A New Framework for People Counting from Coarse to Fine Could be Robust to Viewpoint and Illumination

Abstract

People counting is one of the important tasks in video surveillance. In spite of the significant improvements, this task still has had many challenges such as heavy occlusion in the crowded environment, viewpoint variation, the variety of illumination, etc. People counting process consisted of people detection stage and people tracking stage. This paper focused on boosting the people counting results based on people detection. Our suggested method combines the Deformable Part Models (DPM) and the Deep Convolutional Neural Network (DCNN) to take their advantages and to overcome the shortcomings of each method in people detection. Firstly, to be robust to viewpoint and occlusion, we fuse the people detection results from parts detected by DPM such as head, head-shoulders, upper body, full body. Secondly, to overcome the inefficiency of DPM due to zoom-in view, we use DCNN in detecting head region because the body is often occluded, leaving only head be in full appearance for counting. Finally, we use the late fusion of the detection results from two listed models. PETS 2012 and TUD datasets are selected to experiment and the performance is evaluated by MAE, MRE. The experimental results show that our method could achieve higher performance than the method of Abiol [1], Conte [5] and Subburaman [22] on PETS dataset and especially it could outperform state-of-the-art method as YOLO9000 [10] with parameters fine tuning accordingly to HollywoodHeads dataset. Moreover, it could achieve the high performance in the sparse, medium-density crowd environment and it could be robust to scale, viewpoint, illumination, occlusion, and deformation.