FORM 2
THE PATENTS ACT 1970
(39 of 1970)
AND
The Patents Rules, 2003
COMPLETE SPECIFICATION
(See section 10 and rule 13)
1. TITLE OF THE INVENTION:
"METHOD AND SYSTEM FOR PEDESTRIAN DETECTION USING GENERALIZED SPACE FREQUENCY REPRESENTATION"
2. APPLICANT (S):
(a) NAME: KPIT Cummins Infosystems Limited
(b) NATIONALITY: A company registered under the Companies Act of 1956
(c) ADDRESS: 35 & 36 Rajiv Gandhi Infotech Park, Phase 1, MIDC,
Hinjewadi, Pune 411057, India.
3. PREAMBLE TO THE DESCRIPTION:
The following specification describes the invention and the manner in which it is to be performed.

Field of Invention:
The present invention generally relates to image processing for object detection. More specifically, the invention relates to pedestrian detection as part of an automotive safety system and the method thereof.
Background of Invention:
Accidents due to collision between vehicles and pedestrians are of major concern for vehicles on road. Such accidents pose a major concern in heavily populated cities with high vehicular density giving rise to higher chances of vehicle-pedestrian accidents. Such accidents happen due to the driver's failure to detect a pedestrian in time owing to multiple factors. Pedestrian recognition, thus, has a particularly valuable application in road vehicle safety.
Instances of accidents during night time are increased manifold due to low lighting and similar environmental conditions. Night vision systems thus prove extremely useful and critical, as chances of accidents during night time are higher as compared to the day time, due to various reasons like low light, the driver not being alert due to drowsiness, attention diversion and other such conditions. During night time, the driver faces a problem of low contrast vision which leads to unclear interpretation of the apparent scene at the instant time.
There have been various attempts to prevent the occurrence of a collision between a vehicle and pedestrian. The existing systems provide for pedestrian detection by usage of techniques like night image enhancement, super-resolution image and image fusion to combat night vision issues. These techniques, though, result in improper enhancement of object or local level. The methods utilized are complex and require multiple imaging modalities, leading to an increase in the hardware overhead. Due to variations in weather conditions such as snow, fog, rain, etc. images captured by the camera may contain high level of noise. Thus, an ideal image restoration system must consider various levels of noise and must provide a means for eliminating a wide range of noise, for better image processing and restoration. There is a need for a pedestrian detection system which is efficient, simple and economic.

Additionally, image capturing and processing for a moving vehicle for pedestrian detection is an important consideration. A pedestrian detection system for a moving vehicle can provide early warnings of a pedestrian crossing or obstacle in the street and hence help avoid a potential collision. In such a scenario, speed is a major concern in such applications and cannot be compromised.
Summary;
The present invention provides for pedestrian detection by image restoration methodology for use in night vision systems. The restoration technique is based on Wigner distribution, which renders processing of noisy images and generates visibly efficient and better images to enable the driver for pedestrian awareness.
Wigner distribution kernel is operated on a near infra-red camera input image wherein pedestrian objects are detected. Accordingly, for the detected image, the present invention performs selective enhancement of pedestrian pixel values using predefined mapping curve to calculate the resultant pixel value from the input pixel value. The pixel values in an image are pre-multiplied by a scaling factor such that the pedestrian pixels fall in specific layers. The values obtained as a result of enhancement are then quantized into a plurality of layers according to their range. Further, object segmentation is carried out. For this purpose, connected component labeling is used to segment all the probable pedestrian regions which are obtained in the combination image obtained as a result of adding selected layers. The pedestrians are distinguished from other objects in the image using properties like aspect ratio and size.
Brief Description of the drawing:
Figure 1 illustrates the block diagram according to the embodiment of the invention
Figure 2 illustrates the steps for calculating the boundary pixel value
Figure 3 illustrates the X sobel operator
Figure 4 illustrates the Y sobel operator
Figure 5 illustrates the formula for calculating gradient angle
Figure 6 illustrates the Wigner Distribution Kernel
Figure 7 illustrates the modified pear shaped like curve

Figure 8 illustrates the quantization layers after Wigner distribution kernel has been applied.
Detailed Description:
The pedestrian detection system of the present invention utilizes a single near-infrared (NIR) camera on a vehicle, to capture images of the road that lies ahead of the vehicle. These captured images form the input images which are then processed, according to the method of the invention for pedestrian detection. After detection, the detected pedestrians are highlighted and displayed on a display device for better assistance of a driver. Any display device known in the art, like an LCD, monitor may be used. Additionally, a warning device along with the display device, like an alarm, a blinking light, etc. may be used to alert the driver of the detected pedestrian.
The method and system of pedestrian detection of the present invention performs pedestrian detection by operating the Generalized Space Frequency Representation (GSFR). The GSFR involves application of Wigner distribution kernel. Wigner distribution is a generalized time-frequency representation. To use Wigner distribution function for image processing, it is extended to two-dimensional space. Wigner distribution can be used for achieving good quality of contrast in image and reduction of noise. The technique of the present invention works well with different signal to noise ratios ranging from -1.58dB to 20dB and hence it helps overcome poor image quality due to fog, snow and rain. Such images are stored on a run-time basis and further processed for pedestrian or obstacle detection. The detected pedestrian is then displayed onto a screen for providing assistance to the driver
Fig. 1 depicts the block diagram representation of the method for pedestrian detection according to the embodiment of the invention.
(101) Calculation of edge pixel value of pedestrian:
Pedestrians are vertically oriented objects compared to other object regions in an image i.e. they are associated with vertical edges and hence are unaffected by perspective projection, compared to lane markings, buildings and cars. This property of pedestrians

has been exploited in the method to calculate the pixels in the image which belong to the boundary of the pedestrian in the image, as illustrated in Fig. 2.
1.1 X sobel operator, as illustrated in Fig. 3, is applied on the input image to identify the vertical edges (201). Any other suitable edge detection operator, known in the art may be applied.
1.2 Y sobel operator, as illustrated in Fig. 4, is applied on the same input image to identify horizontal edges (202). Any other suitable edge detection operator, known in the art may be applied.
1.3 Gradient angle of all the pixels using the horizontal and vertical gradient values obtained from the above two steps (201 and 202) are calculated. The equation as illustrated in Fig. 5 is used for calculating this gradient angle (203).
1.4 All the pixels in the gradient image are made zero, except the pixels which have the gradient angle in the range 0-10 degrees(horizontal edges) or in the range70 - 90 degrees (vertical edges) (204).
1.5 The values of pixels in a horizontal edge are made zero in the result image, unless they are associated with a vertical edge. This association is estimated by calculating the aspect ratio and size. Additionally, all vertical edges in the image have non-zero pixel values in the result image.(205)
1.6 The weighted average pixel value (206) calculated from the remaining edges in
the result image obtained from step (205) is obtained as follows.
1.6.1 The result gradient pixel value image is divided by max intensity level i.e. 255 to obtain image matrix corresponding to 0-1 normalized pixel values
1.6.2 Pixel-wise multiplication of the original image with the image matrix obtained in step 1.6.1 is performed.
1.6.3 The average pixel value of the image obtained from multiplication in step 1.6.2 is found.
Knowing the pixel value on the boundary of pedestrian in the image, a scale factor (SF) is determined so that the pedestrian boundary would be visible in the required layer number. The scale factor (SF) is calculated based on the pedestrian boundary pixel value. For example, if the pedestrian boundary pixel value is 120, then the scale factor (SF) is 3.01 which is calculated as illustrated here below:


Wherein,
LNo: Required layer number
PixVal: Pedestrians Boundary Pixel Value as calculated in Step 101
LMax: Maximum layer number
(102) Application of Wigner distribution kernel
Fig. 6 shows the Wigner Distribution Kernel. Wigner distribution kernel is applied to the
input image as follows:
Consider the input image f(x, y) of size 'M x N' and the Wigner kerne! of size 'm x n',
then using the SF as described above, Wigner distribution kernel is applied using the
equation as presented below. The resultant image is improved in contrast and de-noised.

Where:
Lval: 255 * LNo LNo: Required layer number x: Row number of the input pixel y: Column number of the input pixel
(103) Application of pear shaped like curve
The values obtained as a result of Wigner kernel usage are mapped to new values using the predefined curve. This predefined curve is chosen such that the pedestrian detection is facilitated due to boosting of pixel values of interest and suppression of the brighter values and darker values. Pixel values of interest are the values close to the pedestrian pixel value determined from Step (101) of Fig. 2.
The mapping of input pixel values is as described below:
Night-time images captured by an NJR camera generally have a high value of saturation,
due to which they are over-saturated in regions which have higher pixel values. In order

to enhance such an image, the pixel values are scaled according to a predefined standard curve, such as the pear shaped like curve/s shown in Fig. 7. \n Fig.7, the X-axis represents the input pixel values from '0 to 255' and the Y-axis represents a multiplying factor which varies from 0 to 1. The maximum value of the multiplying factor is set to 1, so that the pixel gray values are confined in the range of 0-255. A pear shaped like curve increases gradually from 0 to 1 for smaller values on the X-axis, remains approximately constant at 1 for in-between values and then falls off rapidly for the higher values on the X-axis. The curve equation as illustrated in equationrepresented here below can be changed to get a different 'peak value' if the pixel values of the desired object of interest are known. As such, the peak values are dependent on the object of interest. From the equation of pear shaped curve as mentioned below:
b2 y2= xk (a -x) the curve attains peak value at x = pv,
Wherein, 'pv' is the pedestrian gray value
The power 'k' in the equation (3) depends on 'pv' and is determined using the following formula: k = pv/ (255-pv)
Wherein, 255 represents the maximum range of the input pixel gray values.
(104) Division of image into nine layers
After applying the Wigner kernel and the enhancement method to the input image, the resultant values of the pixels are in the range 0 - 255. In order to quantize the values obtained, the image is then divided into a plurality of layers. According to the embodiment of the invention, the image is divided into nine layers nine layers as illustrated in Fig. 8.
(105) Object segmentation using connected component labeling
The layers in which the pedestrian pixel values lie are combined by simple pixel-wise addition. Next, a connected component labeling is performed on the combination image for object segmentation which segments all the probable regions obtained in the combination image as a result of adding the selected layers.

(106) Pedestrian detection based on aspect ratio
Out of the identified regions, objects other than pedestrians are discarded in order to avoid false positives. This decision is based on the following factors:
• Aspect ratio i.e. the width to height ratio of the segmented region area-Considering the normal average height and width of pedestrian, for the embodiment of the invention the aspect ratio is approximately 2,5 with the tolerance of± 1.
• The smaller objects detected near to the camera i.e. near to the bottom of image and larger objects detected farther from the camera i.e. top of image are removed knowing the fact that the pedestrian near to the camera cannot be smaller and vice versa. If the identified region is located outside the rows corresponding to the specific distance range, based on the distance from the camera and the size of the pedestrian, the object is then classified as random noise. Thus, out of the identified regions, objects other than the pedestrians are discarded.
The present invention is described in scientific terms using the mathematical formulae as stated herein. A person skilled in the art may appreciate that the values of these parameters are relative to application and do not limit the application of the invention.

We Claim,
1. A method for pedestrian detection comprising:
obtaining pedestrian boundary pixel values from the input image using edge
detection operators;
applying Generalized Space Frequency Representation to the image from the
above step to obtain a contrast efficient and de-noised image;
applying a predefined mapping curve to the image from the above step to boost
the identified pixel values of interest;
dividing the output image from the above step into a plurality of layers to extract
pedestrian boundary;
performing connected component labeling on the combined image from the
previous step for object segmentation;
and
performing pedestrian detection on identified regions based on aspect ratio and the
specific distance range.
2. A method for pedestrian detection as claimed in claim 1 wherein, application of the Generalized Space Frequency Representation to the image comprises of utilizing the Wigner distribution kernel.
3. A method for pedestrian detection as claimed in claim 1 wherein, the said predefined mapping curve increases gradually for smaller values on the X-axis, remains approximately constant for interim values and falls off rapidly for higher values on the X-axis thereafter.
4. A method for pedestrian detection as claimed in claim 1 wherein, the input image is divided into a plurality of layers to detect pedestrian like objects in a particular layer or a combination of layers.

5. A method for pedestrian detection as claimed in claim 1 wherein, obvious false positive pedestrians detected are eliminated if the identified region is located outside the row numbers corresponding to the specific distance range.
6. A method for pedestrian detection as claimed in claim 1 wherein, the said pedestrian boundary pixel values are obtained from the gradient angle of pixels calculated using the edge detection operator.
7. A method for pedestrian detection as claimed in claim 1 wherein, the scale factor (SF) is based on the pedestrian boundary pixel value and determined so that the pedestrian boundary would be visible in the required layer number.