Integrating Multiscale Deformable Part Models and Convolutional Networks for Pedestrian Detection

Abstract

Pedestrian detection has many real-world applications, such as advanced driver assistance systems, security surveillance, and traffic control, etc. One of the pedestrian detection challenges is the presence of occlusion. In this study, a jointly learned approach using multiscale deformable part models (DPM) and convolutional neural networks (CNN) is presented to improve the detection accuracy of partially occluded pedestrians. Deep convolutional networks provide a framework that allows hierarchical feature extraction. The DPM is used to characterize non-rigid objects on the histogram of oriented gradients (HoG) feature maps. Scores of the root and parts filters derived from the DPM are used as deformable information to help improve the detection performance. Experimental results show that the proposed jointly learned model can effectively reduce the miss rate of CNN-based object detection models tested on the Caltech pedestrian dataset.