Pedestrian Detection Using Effective Feature Extraction and Disparity Map

Abstract

In this paper, we propose the pedestrian detection using effective feature extraction and disparity map. The pedestrian detection plays an important role in the various industries. Unlike other sensors, the image sensor provides a lot of information similar to the human vision. Thus, the utilization of the image sensor has increased. However, the vision-based pedestrian detection has a weak point in the processing speed due to large amounts of data. Furthermore, the use of a large number of features in the classification is another factor which slows the speed of pedestrian detection. Thus, it is very important to extract and select only the features which help to enhance the detection performance. Our pedestrian detection method consists of several steps, namely disparity map-based pedestrian candidate detection, obstacle segmentation, obstacle classification and pedestrian verification. The first stage is to detect all of the pedestrian candidates in the whole image using the disparity map. The disparity map helps to distinguish the object and the background by offering 3D information. We use road feature information from v-disparity map and perform obstacle detection with pedestrian height to detect the candidates. It is followed by obstacle segmentation using disparity map and stereo vision system parameters to further improve detection performance. The third stage is to classify pedestrians from all candidates using multiple regression analysis-based effective feature extraction. The use of effective feature extraction helps to enhance the processing speed, while maintaining the detection performance significantly. The last stage is to verify the results of classification. Our pedestrian method is verified by applying it to ETH database.