Improving Pedestrian Detection in Crowds With Synthetic Occlusion Images

Chong Shang, Haizhou Ai, Zijie Zhuang, Long Chen Rui Chen

Abstract: Pedestrian detection has been boosted significantly with the rise of convolutional neural networks (CNNs). However, detecting pedestrians in crowded scenes still remains challenging. Numerous efforts have been invested to handle occlusion with domain knowledge, such as occlusion reasoning, non-maximum suppression, etc. Unlike existing works, in this paper we propose a simple yet effective method to handle crowded scenes, i.e., synthesizing more occlusion images, which also enables an empirical analysis on the impact of the amount of occlusion data. We emphasize that a regular CNN detector has not reach its full potential for handling occlusion due to lack of occlusion examples. With synthetic occlusion images, we achieve a state-of-the-art performance on heavy occlusion scenes.