Most solutions to similar problems such as the automatic traffic light for pedestrians requires object detection and tracking algorithms. In this work detection is made on a gray level images.
Most solutions to similar problems such as the automatic traffic light for pedestrians requires object detection and tracking algorithms. In this work detection is made on a gray level images. The SUSAN algorithm give a good answer for the distinction of the pedestrians from their surroundings. The SUSAN algorithm is method for finding corners in a digital images and combining it with the fact that key points such as corners are most likely to appear on human beings gives a reliable human detection method,Further more working on difference image of two images in adjacent time removes unwanted inanimate objects and thus enhances reliability of the detection process. This project is a matlab implementation of these methods intragrated with a tracking method which gives a robust solution for automatic traffic light.
A camera located nearby a trrafic light has to detect pedestrians trying to cross the zebra crossing. In order to perfom such a task a way of discerning human pedestrains from the backround is needed. Other objects such as various vehicles other pedestrians who are not trying cross increases the diffculty one might find in the detection process of the relevant pedstrains and thus good method for “”surgical”” detection is required.
A digital image of Human being is likely to have corners most likely in the intersection between the human the image backround. The SUSAN algorithm gives an optimal way to find these keypoints in an image. A simple explanation of the alogrithm can be found in figure 1. Let us define the USAN-Univalue Segment Assimilating Nucleus, as the circular masked a can be seen in figure 1. A simple glance can tell that in an area where a corner lies a quarter of the pixels inside the USAN are similar in their gray level scale and thus finding a corner is almost equivalent to finding a USUAN which a quater of it contains similar pixels.
The SUSAN algorithm relies on the assumption but of course correction are needed to adjust the algorithm to gray level scale by applying a simlilarity fuction as can be seen in figure 2.
in figure 2. three different kind of similarity function are suggested for the use of corner detection. ( A more comprehensive explanation can be found in the site http://www.fmrib.ox.ac.uk/~steve/susan/ ) In order to enhance corner detection the SUSAN algorithm is applied on a diffrence image of to adjacent frames in order to remove all inanimated object that can be found in the image.
Figure 3 – Original image and the frame difference image after applying the SUSAN algorithm
As it seems the algorithm depicted above discerns well the moving objects such as human and vehicels and thus the alogrithm should be updated to run on a block from the original image and also to defince a noiseless area inside that block in order to prevernt other moving object from interfering with the detection process.
Finally one must define a threshold line and by checking whether the center of mass of the key points found by the SUSAN algorithm can be used in order to find whether the pedestrian wish to cross or not.
The tools used in this project were digital camera supplied by the VISL lab. The source code is written as a matlab script.
Detection of object in an image is more then possible. The use of keypoints such as found by the SUSAN algorithm is a reliable way of object detection and even more when combining it with the frame difference method. A major disadvantage of the SUSAN algorithm is that not always a keypoint found in the a certain framer will be found in the next ,this problem was solved by the calculation of the center of mass which gives the location of the object with good precision. The major advantage derived from the use of the SUSAN algorithm is the fact the one can replace an even large objects with small amount of key points which make the tracking of several object more easy and also saves computation time.
We are grateful to our project supervisor Johanan Erez for his help and guidance throughout this work, and for his kind and professional approach. We also would like to thank the Ollendorff Minerva Center for supporting the project.