Claims:1. A processor implemented method (300) comprising:
controlling motion of a camera configured to capture a surveillance video such that the motion is substantially parallel to a plane containing one or more open-contour texture-less objects under consideration (302);
estimating incremental motion between a plurality of frames of the surveillance video, wherein a sequence of the plurality of frames within the surveillance video forms tuples of ordered multi-view images, each image of the ordered multi-view images comprising a texture-rich background and the one or more open-contour texture-less objects under consideration in a foreground (304);
performing 2-Dimensional (2D) segmentation of each of the plurality of frames, using an object category specific segmentation technique, to separate out the foreground from the texture-rich background to obtain a Region of Interest (ROI) comprising the one or more open-contour texture-less objects under consideration in each of the plurality of frames (306);
performing grid-based keypoint generation to obtain keypoints, wherein the keypoints are points of intersection of a virtual non-rectilinear grid and contours of the ROI and wherein the keypoints are marked along the contours in each of a left frame within each of the tuples (308);
marking a first set of epipolar lines in each of a right frame within each of the tuples corresponding to the keypoints generated on the foreground in the left frame within each of the tuples based on the estimated incremental motion (310);
performing an ordering constraint based injective stereo matching to obtain stereo matched pairs of keypoints, wherein the injective stereo matching comprises matching the keypoints of the left frame with corresponding keypoints in the right frame using the first set of epipolar lines, and wherein the order of keypoints in the left frame matches the order of the corresponding keypoints in the right frame within each of the tuples of ordered multi-view images based on the controlled motion of the camera (312); and
reconstructing a point cloud corresponding to the one or more open-contour texture-less objects based on triangulation using at least the stereo matched pairs of keypoints and one or more intrinsic parameters of the camera (314).
2. The processor implemented method of claim 1, wherein the step of estimating incremental motion is preceded by a step of calibrating offline, the one or more intrinsic parameters of the camera including an optical center and a focal length.

3. The processor implemented method of claim 1, wherein the step of estimating incremental motion is based on incremental Structure from Motion (SfM) using the texture-rich background.

4. The processor implemented method of claim 1, wherein the step of performing grid-based keypoint generation comprises: imposing the virtual non-rectilinear grid on the foreground of the left frame; and marking the keypoints spaced by a configurable distance along the contour.

5. The processor implemented method of claim 1, wherein the first set of epipolar lines intersect the contours of the one or more open-contour texture-less objects in the right frame within each of the tuples of ordered multi-view images, forming the corresponding keypoints therein.

6. The processor implemented method of claim 1, wherein the step of performing the ordering constraint based injective stereo matching is followed by a step of performing periodic compensation and correction of an accumulative error in pose estimation of the camera (316) by:
identifying uniquely located points on the contours in the left frame and the right frame within each of the tuples of ordered multi-view images to form part of the reconstructed point cloud;
computing a mismatch between the uniquely located points and an estimated location of corresponding projected points in the right frame within each of the tuples of ordered multi-view images using a second set of epipolar lines; and
periodically compensating for the mismatch leading to the accumulative error using a global bundle adjustment technique to minimize error associated with the reconstruction of the point cloud.

7. A system (100) comprising:
one or more data storage devices (102) operatively coupled to one or more hardware processors (104) and configured to store instructions configured for execution by the one or more hardware processors to:
control motion of a camera configured to capture a surveillance video such that the motion is substantially parallel to a plane containing one or more open-contour texture-less objects under consideration;
estimate incremental motion between a plurality of frames of the surveillance video, wherein a sequence of the plurality of frames within the surveillance video form tuples of ordered multi-view images, each image of the ordered multi-view images comprising a texture-rich background and the one or more open-contour texture-less objects under consideration in a foreground;
perform 2-Dimensional (2D) segmentation of each of the plurality of frames, using an object category specific segmentation technique, to separate out the foreground from the texture-rich background to obtain a Region of Interest (ROI) comprising the one or more open-contour texture-less objects under consideration in each of the plurality of frames;
perform grid-based keypoint generation to obtain keypoints, wherein the keypoints are points of intersection of a virtual non-rectilinear grid and contours of the ROI and wherein the keypoints are marked along the contours in each of a left frame within each of the tuples;
mark a first set of epipolar lines in each of a right frame within each of the tuples corresponding to the keypoints generated on the foreground in the left frame within each of the tuples based on the estimated incremental motion, wherein the first set of epipolar lines intersect the contours of the one or more open-contour texture-less objects in the right frame within each of the tuples of ordered multi-view images, forming the corresponding keypoints therein;
perform an ordering constraint based injective stereo matching to obtain stereo matched pairs of keypoints, wherein the injective stereo matching comprises matching the keypoints of the left frame with corresponding keypoints in the right frame using the first set of epipolar lines, and wherein the order of keypoints in the left frame matches the order of the corresponding keypoints in the right frame within each of the tuples of ordered multi-view images based on the controlled motion of the camera; and
reconstruct a point cloud corresponding to the one or more open-contour texture-less objects based on triangulation using at least the stereo matched pairs of keypoints and one or more intrinsic parameters of the camera.

8. The system of claim 7, wherein the one or more hardware processors are further configured to calibrate offline, the one or more intrinsic parameters of the camera including an optical center and a focal length.

9. The system of claim 7, wherein the one or more hardware processors are further configured to estimate incremental motion based on incremental Structure from Motion (SfM) using the texture-rich background.

10. The system of claim 7, wherein the one or more hardware processors are further configured to perform the step of performing grid-based keypoint generation by imposing the virtual non-rectilinear grid on the foreground of the left frame; moving along the contour having two open edges; and marking the keypoints spaced by a configurable distance along the contour.

11. The system of claim 7, wherein the one or more hardware processors are further configured to perform periodic compensation and correction of an accumulative error in pose estimation of the camera by:
identifying uniquely located points on the contours in the left frame and the right frame within each of the tuples of ordered multi-view images to form part of the reconstructed point cloud;
computing a mismatch between the uniquely located points and an estimated location of corresponding projected points in the right frame within each of the tuples of ordered multi-view images using a second set of epipolar lines; and
periodically compensating for the mismatch leading to the accumulative error using a global bundle adjustment technique to minimize error associated with the reconstruction of the point cloud.

12. The system of claim 7, wherein the one or more open-contour texture-less objects is a transmission line and the uniquely located points correspond to lowest points on a catenary curve associated with sag in the transmission line.
, Description:FORM 2

THE PATENTS ACT, 1970
(39 of 1970)
&
THE PATENT RULES, 2003

COMPLETE SPECIFICATION
(See Section 10 and Rule 13)

Title of invention:
STEREO SEGMENTATION BASED SURVEILLANCE OF OPEN-CONTOUR TEXTURE-LESS OBJECTS

Applicant:
Tata Consultancy Services Limited
A company Incorporated in India under the Companies Act, 1956
Having address:
Nirmal Building, 9th Floor,
Nariman Point, Mumbai 400021,
Maharashtra, India

The following specification particularly describes the invention and the manner in which it is to be performed.

TECHNICAL FIELD
The disclosure herein generally relates to a surveillance of open-contour texture-less objects, and more particularly relates to systems and methods for stereo segmentation based surveillance of open-contour texture-less objects.

BACKGROUND
Man-made structures in outdoor environments such as power grid conductors, railway tracks and cable-stayed bridges are open-contour, texture-less objects. Automation of remote surveillance of such objects, for their mandatory periodic health monitoring, is being researched upon in recent times. Given the actual length of such structures, the time taken for subsequent image analysis is an important consideration. Specifically, analysis via stereo segmentation of such challenging structures brings out the depth discontinuities, which can be used in context-driven robust inferencing of localized anomalies and structural faults. However, a practically useful approach that targets fast, large-scale, dense stereo segmentation of texture-less, open-contour objects, especially in aerial multi-view images continues to be a challenge.

SUMMARY
Embodiments of the present disclosure present technological improvements as solutions to one or more of the above-mentioned technical problems recognized by the inventors in conventional systems.
In an aspect, there is provided a processor implemented method comprising: controlling motion of a camera configured to capture a surveillance video such that the motion is substantially parallel to a plane containing one or more open-contour texture-less objects under consideration; estimating incremental motion between a plurality of frames of the surveillance video, wherein a sequence of the plurality of frames within the surveillance video forms tuples of ordered multi-view images, each image of the ordered multi-view images comprising a texture-rich background and the one or more open-contour texture-less objects under consideration in a foreground; performing 2-Dimensional (2D) segmentation of each of the plurality of frames, using an object category specific segmentation technique, to separate out the foreground from the texture-rich background to obtain a Region of Interest (ROI) comprising the one or more open-contour texture-less objects under consideration in each of the plurality of frames; performing grid-based keypoint generation to obtain keypoints, wherein the keypoints are points of intersection of a virtual non-rectilinear grid and contours of the ROI and wherein the keypoints are marked along the contours in each of a left frame within each of the tuples; marking a first set of epipolar lines in each of a right frame within each of the tuples corresponding to the keypoints generated on the foreground in the left frame within each of the tuples based on the estimated incremental motion; performing an ordering constraint based injective stereo matching to obtain stereo matched pairs of keypoints, wherein the injective stereo matching comprises matching the keypoints of the left frame with corresponding keypoints in the right frame using the first set of epipolar lines, and wherein the order of keypoints in the left frame matches the order of the corresponding keypoints in the right frame within each of the tuples of ordered multi-view images based on the controlled motion of the camera; and reconstructing a point cloud corresponding to the one or more open-contour texture-less objects based on triangulation using at least the stereo matched pairs of keypoints and one or more intrinsic parameters of the camera.
In another aspect, there is provided a system comprising: one or more data storage devices operatively coupled to the one or more processors and configured to store instructions configured for execution by the one or more processors to: control motion of a camera configured to capture a surveillance video such that the motion is substantially parallel to a plane containing one or more open-contour texture-less objects under consideration; estimate incremental motion between a plurality of frames of the surveillance video, wherein a sequence of the plurality of frames within the surveillance video form tuples of ordered multi-view images, each image of the ordered multi-view images comprising a texture-rich background and the one or more open-contour texture-less objects under consideration in a foreground; perform 2-Dimensional (2D) segmentation of each of the plurality of frames, using an object category specific segmentation technique, to separate out the foreground from the texture-rich background to obtain a Region of Interest (ROI) comprising the one or more open-contour texture-less objects under consideration in each of the plurality of frames; perform grid-based keypoint generation to obtain keypoints, wherein the keypoints are points of intersection of a virtual non-rectilinear grid and contours of the ROI and wherein the keypoints are marked along the contours in each of a left frame within each of the tuples; mark a first set of epipolar lines in each of a right frame within each of the tuples corresponding to the keypoints generated on the foreground in the left frame within each of the tuples based on the estimated incremental motion, wherein the first set of epipolar lines intersect the contours of the one or more open-contour texture-less objects in the right frame within each of the tuples of ordered multi-view images, forming the corresponding keypoints therein; perform an ordering constraint based injective stereo matching to obtain stereo matched pairs of keypoints, wherein the injective stereo matching comprises matching the keypoints of the left frame with corresponding keypoints in the right frame using the first set of epipolar lines, and wherein the order of keypoints in the left frame matches the order of the corresponding keypoints in the right frame within each of the tuples of ordered multi-view images based on the controlled motion of the camera; and reconstruct a point cloud corresponding to the one or more open-contour texture-less objects based on triangulation using at least the stereo matched pairs of keypoints and one or more intrinsic parameters of the camera.
In yet another aspect, there is provided a computer program product comprising a non-transitory computer readable medium having a computer readable program embodied therein, wherein the computer readable program, when executed on a computing device, causes the computing device to: control motion of a camera configured to capture a surveillance video such that the motion is substantially parallel to a plane containing one or more open-contour texture-less objects under consideration; estimate incremental motion between a plurality of frames of the surveillance video, wherein a sequence of the plurality of frames within the surveillance video form tuples of ordered multi-view images, each image of the ordered multi-view images comprising a texture-rich background and the one or more open-contour texture-less objects under consideration in a foreground; perform 2-Dimensional (2D) segmentation of each of the plurality of frames, using an object category specific segmentation technique, to separate out the foreground from the texture-rich background to obtain a Region of Interest (ROI) comprising the one or more open-contour texture-less objects under consideration in each of the plurality of frames; perform grid-based keypoint generation to obtain keypoints, wherein the keypoints are points of intersection of a virtual non-rectilinear grid and contours of the ROI and wherein the keypoints are marked along the contours in each of a left frame within each of the tuples; mark a first set of epipolar lines in each of a right frame within each of the tuples corresponding to the keypoints generated on the foreground in the left frame within each of the tuples based on the estimated incremental motion, wherein the first set of epipolar lines intersect the contours of the one or more open-contour texture-less objects in the right frame within each of the tuples of ordered multi-view images, forming the corresponding keypoints therein; perform an ordering constraint based injective stereo matching to obtain stereo matched pairs of keypoints, wherein the injective stereo matching comprises matching the keypoints of the left frame with corresponding keypoints in the right frame using the first set of epipolar lines, and wherein the order of keypoints in the left frame matches the order of the corresponding keypoints in the right frame within each of the tuples of ordered multi-view images based on the controlled motion of the camera; and reconstruct a point cloud corresponding to the one or more open-contour texture-less objects based on triangulation using at least the stereo matched pairs of keypoints and one or more intrinsic parameters of the camera.
In an embodiment of the present disclosure, the one or more hardware processors are further configured to calibrate offline, the one or more intrinsic parameters of the camera including an optical center and a focal length.
In an embodiment of the present disclosure, the one or more hardware processors are further configured to estimate incremental motion based on incremental Structure from Motion (SfM) using the texture-rich background.
In an embodiment of the present disclosure, the one or more hardware processors are further configured to perform the step of performing grid-based keypoint generation by imposing the virtual non-rectilinear grid on the foreground of the left frame; moving along the contour having two open edges; and marking the keypoints spaced by a configurable distance along the contour.
In an embodiment of the present disclosure, the one or more hardware processors are further configured to perform periodic compensation and correction of an accumulative error in pose estimation of the camera by: identifying uniquely located points on the contours in the left frame and the right frame within each of the tuples of ordered multi-view images to form part of the reconstructed point cloud; computing a mismatch between the uniquely located points and an estimated location of corresponding projected points in the right frame within each of the tuples of ordered multi-view images using a second set of epipolar lines; and periodically compensating for the mismatch leading to the accumulative error using a global bundle adjustment technique to minimize error associated with the reconstruction of the point cloud.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the embodiments of the present disclosure, as claimed.
BRIEF DESCRIPTION OF THE DRAWINGS
The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the disclosed principles.
FIG.1 illustrates an exemplary block diagram of a system for stereo segmentation based surveillance of open-contour texture-less objects, in accordance with an embodiment of the present disclosure.
FIG.2 is a high level pipeline illustrating key stages in a method for stereo segmentation based surveillance of open-contour texture-less objects, in accordance with an embodiment of the present disclosure.
FIG.3A and FIG.3B is an exemplary flow diagram illustrating a computer implemented method for stereo segmentation based surveillance of open-contour texture-less objects, in accordance with an embodiment of the present disclosure.
FIG.4A illustrates background keypoints spread across a frame, in accordance with an embodiment of the present disclosure.
FIG.4B illustrates an output of 2D segmentation of a frame, in accordance with an embodiment of the present disclosure.
FIG.4C illustrates a grid overlay on a Region of Interest (ROI), in accordance with an embodiment of the present disclosure.
FIG.5 illustrates two-view open contour stereo matching in accordance with an embodiment of the present disclosure.
FIG.6 illustrates an injective stereo matching scenario in accordance with an embodiment of the present disclosure.
FIG.7 illustrates overlapping edge formation in adjacent views in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION
Exemplary embodiments are described with reference to the accompanying drawings. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. Wherever convenient, the same reference numbers are used throughout the drawings to refer to the same or like parts. While examples and features of disclosed principles are described herein, modifications, adaptations, and other implementations are possible without departing from the spirit and scope of the disclosed embodiments. It is intended that the following detailed description be considered as exemplary only, with the true scope and spirit being indicated by the following claims.
Conventional algorithmic multi-view reconstruction pipelines such as Structure from Motion (SfM) and Simultaneous Localization and Mapping (SLAM) can be adapted for segmentation of objects using multi-view images. However, stereo correspondence is a known bottleneck in these approaches. Excess false correspondences in cluttered outdoor scenes aggravate this problem. The bottleneck happens due to matching ambiguities, arising from low or repetitive textures, occlusions, wide baseline, and the like. Another dimension of aggravation happens when the object is much larger in some dimension, e.g. open-contour. It becomes worse when it is texture-less as well, since lack of feature points on the foreground implies that the fast feature-based reconstruction approaches cannot be applied to them in a straightforward way. Important man-made structures, especially the category of long-linear infrastructures such as bridges, railway lines, oil and gas supply pipelines, etc. are open-contour texture-less objects. Maintenance of such mechanical structures is important and costly. Periodic remote inspections are hence done to locate anomalies and faults which need to be fixed immediately. Stereo segmentation of such challenging structures brings out the depth discontinuities, which can be used in context-driven robust inferencing of localized anomalies and structural faults.
Drones are popular for surveillance of such long-linear infrastructures that run into hundreds of miles. The remote observation separation mandated by right-of-way requirements, and also payload limit of such platforms (arising from aerodynamics considerations) entails that a monocular camera can only be practically mounted for such missions in general and not depth sensors like Light Detection and Ranging (LiDAR). Unlike terrestrial imaging, aerial imaging leads to capture of significant clutter, in the form of complex and heterogeneous background or terrain. Further, large distance surveillance leads to large-scale image data, which demands a scalable, large-scale reconstruction solution.
Conventionally, in cluttered scenarios, 2-Dimensional (2D) segmentation is performed across a set of images, before doing a multi-view reconstruction of the foreground (object) in focus. A drawback of traditional dense reconstruction methods is that depth values tend to be inaccurate for texture-less object surfaces. To alleviate that, global methods for stereo matching, which use global optimization are used. Such optimization being nondeterministic polynomial time (NP)- hard, such methods are very time and memory consuming. For a large-scale video, especially one arising from long-range surveillance, scalability is a primary issue.
Most work in this domain deal with texture-less planar surfaces only, including plane sweep based algorithms. Some do reconstruct curved surfaces, but it is limited to reconstruction of quadratic surfaces, which are closed-contour second-order surfaces. In contrast, open-contour objects, especially hanging cables, are known to be (infinite-order) catenary curve. Again some work deal with reconstruction of thin objects, but it is limited to objects which are essentially planar, such as a flat beam. The approach in some work is restricted to recovering the location of specific closed contour objects while some approximate any non-planar surface to be piecewise planar, and hence are unable to construct sharp-curving surface of thin open-contour objects.
The present disclosure is directed to providing a fast, large-scale dense segmentation of open-contour texture-less objects in aerial multi-view images. Systems and methods of the present disclosure provide an ordering based fast and dense stereo matching for such objects. Also by leveraging the near-linearity of such objects, periodic compensation is provided for drift or accumulation error in motion estimation using uniquely located points on the objects.
Although further description of the present disclosure is explained with figures illustrating transmission lines, it may be noted that the described application is non-limiting and systems and methods of the present disclosure may be applied to other open-contour texture-less objects such as bridges, railway lines, oil and gas supply pipelines, etc.
Referring now to the drawings, and more particularly to FIGS. 1 through 7, where similar reference characters denote corresponding features consistently throughout the figures, there are shown preferred embodiments and these embodiments are described in the context of the following exemplary system and method.
FIG.1 illustrates an exemplary block diagram of a system 100 for stereo segmentation based surveillance of open-contour texture-less objects, in accordance with an embodiment of the present disclosure. In an embodiment, the system 100 includes one or more processors 104, communication interface device(s) or input/output (I/O) interface(s) 106, and one or more data storage devices or memory 102 operatively coupled to the one or more processors 104. The one or more processors 104 that are hardware processors can be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, graphics controllers, logic circuitries, and/or any devices that manipulate signals based on operational instructions. Among other capabilities, the processor(s) are configured to fetch and execute computer-readable instructions stored in the memory. In an embodiment, the system 100 can be implemented in a variety of computing systems, such as laptop computers, notebooks, hand-held devices, workstations, mainframe computers, servers, a network cloud and the like.
The I/O interface device(s) 106 can include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like and can facilitate multiple communications within a wide variety of networks N/W and protocol types, including wired networks, for example, LAN, cable, etc., and wireless networks, such as WLAN, cellular, or satellite. In an embodiment, the I/O interface device(s) can include one or more ports for connecting a number of devices to one another or to another server.
The memory 102 may include any computer-readable medium known in the art including, for example, volatile memory, such as static random access memory (SRAM) and dynamic random access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable ROM, flash memories, hard disks, optical disks, and magnetic tapes. In an embodiment, one or more modules (not shown) of the system 100 can be stored in the memory 102.
In an embodiment, the system 100 includes one or more data storage devices or memory 102 operatively coupled to the one or more processors 104 and is configured to store instructions configured for execution of steps of the method 200 by the one or more processors 104.
FIG.2 is a high level pipeline 200 illustrating key stages in a method for stereo segmentation based surveillance of open-contour texture-less objects and FIG.3A and FIG.3B is an exemplary flow diagram 300 illustrating a computer implemented method for stereo segmentation based surveillance of open-contour texture-less objects, in accordance with an embodiment of the present disclosure. The steps of the method 300 will now be explained in detail with reference to the components of the system 100 of FIG.1 and the key stages depicted in FIG.2. Although process steps, method steps, techniques or the like may be described in a sequential order, such processes, methods and techniques may be configured to work in alternate orders. In other words, any sequence or order of steps that may be described does not necessarily indicate a requirement that the steps be performed in that order. The steps of processes described herein may be performed in any order practical. Further, some steps may be performed simultaneously.
In accordance with an embodiment of the present disclosure, the one or more processors 104 are configured to control motion of a camera, at step 302, to capture a surveillance video such that the motion is substantially parallel to a plane containing one or more open-contour texture-less objects under consideration. The camera may be any optical device configured to capture images in the form of still photographs or sequence of images constituting videos.
In an embodiment, at step 304, incremental motion between a plurality of frames of the surveillance video is estimated. In accordance with the present disclosure, a sequence of the plurality of frames within the surveillance video forms tuples of ordered multi-view images or stereo images. Each image comprises a texture-rich background and the one or more open-contour texture-less objects under consideration in a foreground. In an embodiment, the step of estimating incremental motion is based on incremental Structure from Motion (SfM) using the texture-rich background.
In incremental SfM using the surveillance video, a new frame is added to a current estimation of scene geometry and camera motion in each iteration of Bundle Adjustment (BA). In each such BA iteration, motion parameters between a current frame and a previous frame (1 frame delay) are calculated before the estimation of a scene geometry to estimate the latest position of (monocular) camera’s point-of-view. It may be noted that in accordance with the present disclosure, only a segmented region of the image is updated. The overall pipeline is as depicted in FIG.2 with detailed steps described herein under.
It is known that accuracy of stereo reconstruction process has a critical and nonlinear dependency on the accuracy of the camera calibration parameters. Further, it is known that if one or more intrinsic parameters of the camera are known a-priori to reconstruction, then the inherent ambiguity in SfM based reconstruction gets limited to similarity transformation only. Also, unlike self-calibration, offline calibration using specific calibration-only methods also help by not disregarding any significant nonlinear camera distortions, which arise especially in wide-angle cameras used in aerial imaging. Accordingly, in the present disclosure, the step of estimating incremental motion is preceded by a one-time offline step of calibrating the one or more intrinsic parameters of the camera. The scene geometry (foreground model/structure) is estimated using the texture-rich background by estimating a fundamental matrix on a per-frame basis (rather than on a minimal subset offline, as known in the art) that can be later used for reconstruction in each iteration. The fundamental matrix is configured to map associated 3-Dimentionsal (3D) world coordinates to coordinates of the ordered multi-view stereo images. In an embodiment, the intrinsic parameters may include an optical center and a focal length of the camera.
For robust estimation of a fundamental matrix F, it is always desired that the corresponding keypoints be concentrated in a 2D / projected region. Hence, in accordance with the present disclosure, first an Oriented FAST and Rotated BRIEF (ORB) -based keypoint detection is performed in a current and previous frame. The image descriptors (shape, color, texture, etc.) are putatively matched and a standard RANSAC-based outlier removal technique is used for robustness. The current and previous frames are then divided into 32 x 32 blocks, and tuples of block in the two frames which are bound by exactly one (inlier) keypoint match are picked up. Alternately, the tuples are chosen randomly in case of multiple available matches. If there is no matching detected in a block pair, then that block pair is disregarded. FIG.4A illustrates background keypoints spread across a frame, in accordance with an embodiment of the present disclosure.
In accordance with an embodiment of the present disclosure, the one or more processors 104 are configured to perform 2-Dimensional (2D) segmentation of each of the plurality of frames, at step 306, to separate out the foreground from the texture-rich background to obtain a Region of Interest (ROI) comprising the one or more open-contour texture-less objects under consideration in each of the plurality of frames. FIG.4B illustrates an output of 2D segmentation of a frame, in accordance with an embodiment of the present disclosure. Object-category agnostic methods do not work well for segmentation of open-contour objects. Hence, in accordance with the present disclosure, an object category specific segmentation technique is used at step 306. Improved accuracy of 2D projection eventually helps in reconstructing as much surface of the object as possible. In another embodiment, a lightweight approach may be used that involves tracking via, say, optical flow, the object, in next few successive frames, rather than segmenting in each frame. But for long open-contour objects across perspectively transformed scenes, the flow field is not uniform and may be difficult to track.
In accordance with an embodiment of the present disclosure, the one or more processors 104 are configured to perform grid-based keypoint generation, at step 308, to obtain keypoints, wherein the keypoints are points of intersection of a virtual non-rectilinear grid and contours of the ROI and wherein the keypoints are marked along the contours in each of a left frame within each of the tuples. Firstly, the contours of the object are matched, to estimate the depth. The virtual non-rectilinear grid is imposed on the foreground of the left frame and then the keypoints are marked along the contour spaced by a configurable distance. FIG.4C illustrates a grid overlay on the ROI, in accordance with an embodiment of the present disclosure. The grid can be arbitrarily dense because the grid is imposed only on the foreground, which typically occupies a small percentage (for instance < 5%) of total frame region.
Image matching is one of the most time-consuming steps of SfM as well as a bottleneck in the quality of reconstruction. In accordance with the present disclosure, a fast injective, geometry-based mapping of keypoints marked on the foreground contours in adjacent frames is implemented. The method of the present disclosure is independent of a specific image descriptor since the keypoint location information or geometry is used to perform stereo matching and not the keypoint appearance. Also, geometric description is just a few bytes, and hence speeds up the matching procedure. Choosing a geometric matching approach over an appearance-based approach also helps to bypass the problem of poor appearance-based feature correspondences in typical wide baseline setups of aerial imagery.
In accordance with an embodiment of the present disclosure, the one or more processors 104 are configured to mark a first set of epipolar lines in each of a right frame within each of the tuples corresponding to the keypoints generated on the foreground in the left frame within each of the tuples, at step 310, based on the estimated incremental motion.
Given a highly common motion of camera relative to the open-contour object, each marked keypoint on the contour of the left frame assuming two-view geometry, unambiguously geometrically matches to another point on the contour of the right frame. This property naturally generalizes to multi-view geometry as well, given consistent camera motion. Further, an ordering constraint based injective stereo matching is performed at step 312, wherein the injective stereo matching, in accordance with an embodiment, comprises matching the keypoints of the left frame with corresponding keypoints in the right frame using the first set of epipolar lines, and wherein the order of keypoints in the left frame matches the order of the corresponding keypoints in the right frame within each of the tuples of ordered multi-view images based on the controlled motion of the camera. It may be noted that the first set of epipolar lines intersect the contours of the one or more open-contour texture-less objects in the right frame within each of the tuples of ordered multi-view images, forming the corresponding keypoints therein. In the absence of ordering, the epipolar lines would have criss-crossed and matching of keypoints is not possible. The controlled motion of the camera in step 302 helps to achieve the ordering constraint based injective stereo matching. FIG.5 illustrates two-view open contour stereo matching in accordance with an embodiment of the present disclosure.
At step 302, the motion of the camera is controlled such that it is substantially parallel (maintaining similar distance from the object in successive two to three views) to the plane containing the object under consideration. Based on this controlled motion, the matched right frame contour keypoint will not be in the forbidden zone virtually marked by the left frame contour keypoint. Alternatively, any point on the contour of a left frame, which is to the right of another point in the same frame, will not change relative ordering in the right frame, which can give rise to potential matching ambiguity. Specifically, parallel motion indicates that the 3D direction formed by two points on contour of the left frame is near-parallel (in 3D) to the instantaneous direction of the camera movement, especially catenary curved objects such as hanging cables which fit a single plane. FIG.6 illustrates the injective stereo matching scenario in accordance with an embodiment of the present disclosure. The forbidden zone is essentially the plane spanned by three points viz., the two locations of the (monocular) camera’s optical center in adjacent frames, and a 3D scene point (corresponding to one marked keypoint, say Pt1 (?Pt1?^' in a subsequent frame).
Given the typical distances (> 20 meters from the object), and distance covered between two frames at typical frame rate of 30 fps and nominal speed of fast-moving camera platform (up to 60 kmph), the shorter angle subtended in the isosceles triangle of the forbidden zone is < 0.8º as represented below.
tan^(-1)??(1/2·(60×1000)/(60×60)·1/30)/20˜0.8°?
At such highly acute angles, it is obvious from FIG.5 that in the near-parallel direction case as described above, the next point Pt2 (?Pt2?^' in a subsequent frame) marked on the contour, especially that part of contour which is closest to the camera optical axis, does not fall in the forbidden zone. For the keypoints which are sufficiently far away from the camera optical axis, this property will hold true when Pt1 and Pt2 are not arbitrarily close to each other. In practice, the grid dimensions ensure that such closeness never happens.
Oftentimes, open-contour objects are also highly smooth objects. Hence the contours of their 2D projections can potentially differ across different images, observed from different camera positions. Matching points on different contours actually mean putative matching of different 3D points, which does not make photogrammetric sense. However, open-contour objects are relatively thin objects. Furthermore, in remote surveillance, the distance of their observation is also much more than their mean cross-sectional diameter. In such a case, hypothetical tangent lines to any specific infinitesimally thin locally normal cross-section of the surface of such open-contour objects from two successive views impinge the cross-section at points, whose 3D depths are extremely close. FIG.7 illustrates overlapping edge formation in adjacent views in accordance with an embodiment of the present disclosure. Extending this to multiple adjacent cross-sections, it may be noted that the projected contours in adjacent frames can be triangulated to a common, mean 3D contour. The estimated depth of such mean contour is sensibly within the limits of minimal reconstruction error.
The main differentiators of the matching technique of the present disclosure from conventional techniques include: Firstly the keypoints are not detected but imposed via a grid. Secondly, the keypoints are located on the epipolar lines itself and not in the vicinity of the epipolar line of each keypoint. Thirdly, keypoints are considered on the desired foreground of interest and not on the entire scene, given the end application of remote surveillance.
In accordance with an embodiment of the present disclosure, the one or more processors 104 are configured to reconstruct, at step 314, a point cloud corresponding to the one or more open-contour texture-less objects based on triangulation using at least the stereo matched pairs of keypoints and one or more intrinsic parameters of the camera. The reconstruction so far only generates a wire-frame model, by reconstructing the top and bottom 3D open contours respectively. If the cross-section is regular (e.g. a flat beam having rectangular cross-section and a cable having circular cross-section), then interpolation methods can be used to assign depth to virtual points on the object’s visible surface. For flat open-contour objects, a simple discrete interpolation may be performed. Four 3D end-points of the two contours maybe treated as corner-points of a 3D plane, and any known plane interpolation technique may be employed. For other open-contour objects such as wires, the cross-section is a regular algebraic curve (e.g. a circle). Hence from the estimation of the two depths for any specific point on the upper and lower contour of an open-contour object, the parameters of the cross-section (e.g. circle radius) may be estimated. The parameters may be used to synthetically interpolate and reconstruct the entire surface of such a cross-section. Concatenation of such cross-section leads to reconstruction of the surface of the entire curved-surface open-contour object. Such an interpolative reconstruction works even when the open-contour object is curved along its leading length, e.g., a cylindrical parabola for a hanging wire.
In practice, there are multiple open-contour objects supporting a mechanical system. It is also known that for mechanical design aspects such as load balancing, such open-contour objects are laid out symmetrically, and do not cross each other, especially in frontal view. The present disclosure leverages this property to uniquely mark out different instances of the object in each frame (for example, mark each power line top-to-bottom in each frame). Using unique markings, it is possible to match the contours of the correct common instance.
Given that offline calibration is performed as stated above, the ambiguity in SfM-driven segmentation is restricted to only similarity transformations. Many of the corresponding equivalences, especially that of scale, may be disambiguated by a well-known technique of using few ground truth depth measures (of few ground control points), and then imposing the popular smoothness constraint in disparity space (which also maps directly to the depth space).
The problem of scale can be of critical importance, when it comes to using stereo segmentation for fault detection purposes. Surveillance videos for open-contour objects are generally quite long. So the possibility of accumulation of pose estimation errors is expected to be present. The motion of camera being restricted to near-linear due to nature/shape of such objects, it is not possible to perform loop closure or use symmetries for compensating such error. In many cases such as aerial imaging over non-urban regions, it may not be possible to reliably employ external references such as GPS, GIS data or 3D models to periodically compensate for such error.
In accordance with an embodiment of the present disclosure, the one or more processors 104 are configured to perform periodic compensation and correction of an accumulative error in pose estimation of the camera, at step 316. In almost all cases of open-contour objects, there are expected to be uniquely located points on the contour of the object. For example, each catenary curve representing a transmission line has a lowest or minimal point associated with sag in the transmission line, which is captured by the camera periodically (say every 20th frame). In accordance with the present disclosure, such uniquely located points on the contours in each of the left frames and the right frames form part of the point cloud being reconstructed at step 314 along with the stereo matched pairs of keypoints. A mismatch between the uniquely located keypoints and an estimated location of corresponding projected points in each of the right frames is computed using a second set of epipolar lines. The computed mismatch leading to the accumulative error is then periodically compensated using a global bundle adjustment technique to minimize error associated with the reconstruction of the point cloud.
EXPERIMENTS AND DATA COLLECTION
Aerial surveillance data related to a power transmission grid was collected. It was captured using a 11 MP f/2.8 wide-lens RGB camera, GoPro Hero3, mounted on a mini-drone. The image size captured was 3000 x 2250. Two test sites were used for imaging the grids. Both surveillance missions were carried in different times of day: afternoon and evening. The drone was flown so as to have a sideways view of the power grids. Enough length of power grids was imaged, providing around 70 different views between each single span of the powerline (an open-contour object) between two successive transmission towers. The background of the powerlines in both cases was different, but was varying all through, typically consisting of vegetation, sky, unpaved roads and houses. Given the minimum required distance of observation, the typical width of projection of powerline was mere 7 pixels, against a heterogeneous background as wide as 2250 pixels. This made the segmentation problem very challenging. Given the vastness and operational costs of laser scanners, it was not possible to comprehensively label the ground-truth for the challenging dataset collected. Rather, the limited ground truth was manually recorded for the core novelty and the fast stereo correspondence. The accuracy of estimated correspondence was verified against that. The reconstruction/correspondence quality was compared against two other popular state-of-the-art SfM approaches: VisualSfM and PhotoScan. The former software (closed source) includes PMVS2 in its pipeline, which being patch-based direct, dense method, may potentially reconstruct texture-less objects.
RESULTS AND PERFORMANCE ANALYSIS
A typical SfM pipeline is evaluated against many measures, such as: cardinality of subset of images/keyframes used for reconstruction (represents speed), number of reconstructed scene points, overall time taken, average reprojection error or error rate (represent reconstruction accuracy), to name a few. However, since the present disclosure focusses on foreground stereo segmentation, particularly solving the problem of keypoint matching and. the overlay grid is made arbitrarily dense, to have a dense reconstruction, the choice of evaluation narrows down to judge the trade-off between speed and quality of reconstruction. Especially in a video sequence with mostly linear motion of camera due to open-contour nature, the choice of keyframe subset generally gets limited to a trivial choice of, for example, every-5th-frame. Hence the analyses is further restricted to evaluating reconstruction quality only.
In general, quality of correspondence-driven stereo segmentation can be judged via stereo correspondences themselves, as is the protocol for evaluation of most stereo correspondence algorithms. However, since there are too many (dense) matches in the dataset under consideration, manual counting of those can lead to errors. So the quality is evaluated by:
1) Observing the overall shape of reconstructed object and visually judging the reconstruction. It may be understood that change in relative scale along the jittery camera motion may subdue certain localized reconstruction errors.
2) Quantitatively estimating and measuring using average reprojection error.
Since ground truth for reconstructed scenes is absent, the standard protocol of computing pixel wise validation error rate cannot be followed. Hence mean reprojection error has been used as the unsupervised quality measure.
For VisualSfM and PhotoScan, 100 frames were used to reconstruct the scene. For the method of the present disclosure, currently with few limitations in pose estimation, 16 frames were used. It was observed that both VisualSfM and PhotoScan are not able to reconstruct texture-less objects, at all, in outdoor scenario. On the other hand, the method of the present disclosure is able to reconstruct it for both datasets. Further, it was seen that the 3D shape of the open-contour object (thin cylindrical parabola) is retained. The reconstruction is complete to the point of density of choice in virtual grid made earlier. The accuracy of reconstruction is estimated as shown in Table 1 herein below.
Table 1: Segmentation results for state-of-the-art approaches and the method of the present disclosure.
Dataset #Reconstructed Foreground Points Average reprojection error Normalized runtime
VisualSfM PhotoScan Present disclosure VisualSfM PhotoScan Present disclosure VisualSfM PhotoScan Present disclosure
Dataset 1 9.6K 332K 915K 3.2 1.2 2.8 30 mins 104 mins 72 mins
Dataset 1 9.6K 164K 527K 2.6 1.1 3.4 17 mins 50 mins 41 mins

600 points were chosen on the two open-contours (upper and lower), of the 2D segmented object, per image pair. Since the method of the present disclosure focuses on foreground reconstruction, all the results are scaled with respect to % of scene points constructed by various methods. It can be seen that on-an average, the reprojection error is less on both datasets for the method of the present disclosure. This happens due to the fact that the wide baseline setup for aerial view leads to reduction in depth error (and consequently, reprojection error). By doing a next-best-view selection and dropping intermediate frames, it is expected that the implicit baseline increase will lead to further reduction in this error. The running time was measured on a single machine running 8-core 3.4 GHz Intel® core i7 processors, having 64 GB RAM, and with NVidia® Quadro K620 as GPU card. Both VisualSfM and PhotoScan run relatively faster but lead to sparse reconstruction. Both of them use PMVS2 to post process and densify the point cloud. The method of the present disclosure performs direct densification by using a dense virtual grid.
Since matching is performed over the length of open contours, the expected complexity of the matching step of the present disclosure is ~ O(n). However, in the overall stereo segmentation pipeline, it is the 2D segmentation stage, which has the (highest) complexity of O(n2). Hence the overall complexity of the method of the present disclosure is O(n2). It is obviously higher than the fastest SfM reported O(n). But that is the price that has to be paid, as a trade-off, to perform dense segmentation of a very challenging object category, using the O(n) method (visualSfM).
Systems and methods of the present disclosure provide a stereo matching technique for open-contour texture-less objects, and a supporting stereo reconstruction pipeline. Many man-made structures of vast sizes, especially linear infrastructures, have such objects. Reconstruction of such objects is an important step towards the eventual goal of detection and segmentation of structural damages to surfaces of such objects. The method of the present disclosure is both fast and scalable, and makes use of matching of marked keypoints on the outer contours of such objects. The matching step can fit in any SfM pipeline till date, though an optimized reconstruction pipeline is provided as well. The methods of the present disclosure can also help in solving related geometric problems such as clearance from obstructions, encroachments etc. It has been proven through experiments that the method of the present disclosure is able to stereo-segment such objects, that conventional benchmark algorithms are unable to do.
The written description describes the subject matter herein to enable any person skilled in the art to make and use the embodiments. The scope of the subject matter embodiments is defined by the claims and may include other modifications that occur to those skilled in the art. Such other modifications are intended to be within the scope of the claims if they have similar elements that do not differ from the literal language of the claims or if they include equivalent elements with insubstantial differences from the literal language of the claims.
It is to be understood that the scope of the protection is extended to such a program and in addition to a computer-readable means having a message therein; such computer-readable storage means contain program-code means for implementation of one or more steps of the method, when the program runs on a server or mobile device or any suitable programmable device. The hardware device can be any kind of device which can be programmed including e.g. any kind of computer like a server or a personal computer, or the like, or any combination thereof. The device may also include means which could be e.g. hardware means like e.g. an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a combination of hardware and software means, e.g. an ASIC and an FPGA, or at least one microprocessor and at least one memory with software modules located therein. Thus, the means can include both hardware means and software means. The method embodiments described herein could be implemented in hardware and software. The device may also include software means. Alternatively, the embodiments may be implemented on different hardware devices, e.g. using a plurality of CPUs.
The embodiments herein can comprise hardware and software elements. The embodiments that are implemented in software include but are not limited to, firmware, resident software, microcode, etc. The functions performed by various modules described herein may be implemented in other modules or combinations of other modules. For the purposes of this description, a computer-usable or computer readable medium can be any apparatus that can comprise, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
The illustrated steps are set out to explain the exemplary embodiments shown, and it should be anticipated that ongoing technological development will change the manner in which particular functions are performed. These examples are presented herein for purposes of illustration, and not limitation. Further, the boundaries of the functional building blocks have been arbitrarily defined herein for the convenience of the description. Alternative boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. Alternatives (including equivalents, extensions, variations, deviations, etc., of those described herein) will be apparent to persons skilled in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope and spirit of the disclosed embodiments. Also, the words “comprising,” “having,” “containing,” and “including,” and other similar forms are intended to be equivalent in meaning and be open ended in that an item or items following any one of these words is not meant to be an exhaustive listing of such item or items, or meant to be limited to only the listed item or items. It must also be noted that as used herein and in the appended claims, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.
Furthermore, one or more computer-readable storage media may be utilized in implementing embodiments consistent with the present disclosure. A computer-readable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processor(s) to perform steps or stages consistent with the embodiments described herein. The term “computer-readable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., be non-transitory. Examples include random access memory (RAM), read-only memory (ROM), volatile memory, nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and any other known physical storage media.
It is intended that the disclosure and examples be considered as exemplary only, with a true scope and spirit of disclosed embodiments being indicated by the following claims.