Computer Science and Electrical Engineering Session:

Title:            Avoiding multiple overlapped detection of single pedestrian
Authors:       N. K. Ragesh, R. Rajesh
Abstract:      Pedestrian detection is one of the major applications addressed in Advanced Driver Assistance Systems (ADAS). There are two approached in detecting pedestrians from a 2D image. One approach is called template matching, where a set of pedestrian templates are matched with various parts of the image in different scales whereas the other method involves extracting some unique features from search windows all over the image and classify the window as a pedestrian or not using some classifier. Both these approaches have their own advantages and disadvantages. The feature based approach is getting more attention because it is computationally less complex and is virtually impossible to find all possible templates of a pedestrian. In both these case, there could be a case where same pedestrian is detected multiple times in overlapped search windows at slightly different position and scales. This paper investigates techniques for avoiding these false repeated detections of same pedestrian. Various
methods are compared against their computational complexity as well as accuracy in repeated detection elimination.

Keywords:    Advanced Driver Assistance Systems (ADAS), Pedestrian Detection, Non Maximum Suppression, Detection Window Clustering.
Pages:            532-536

Download Full Paper