A Study on Occluded Pedestrian Detection Based on Block-based Features and Ensemble Classifier

When dealing with pedestrian detection under actual urban streets environment, pedestrian targets tend to be partially occluded in varying degrees because there are many vehicles and other transportation facilities, which affects the performance of the detection system. This paper presents a method for solving partial occlusions for pedestrian detection. It is based on the block-based feature and random subspace classifier to construct the ensemble classifiers. When the test results of holistic classifier is ambiguous, occlusion inference is conducted; if occlusion does exist, detection window will be classified by the ensemble classifier. Different pedestrian data sets are used: Daimler data set, TUD data set and real on-board pedestrian images. We test three different methods for both partially occluded and non-occluded data. Experimental results show that the method proposed in this paper performed better when dealing with partial occlusion situation and doesn't affect the detection performance for non-occluded targets.