An Approach for Handling Partially Visible Human Shapes in People Detection Systems

In this paper, we propose a novel approach for handling occlusion in people detection systems using deep learning technique. The occlusion is handled by applying a part-object detection classifier on top of a whole-object classifier for detections with low confidence in our proposed architecture. In contrast to the other approaches, the proposed framework uses only one part-object detection classifier that is capable of performing a multi-class inference on the proposed regions. Experimental results show that our approach is able to detect partially visible human shapes better than other state-of-the-art models like ResNet50, ResNet101 and MobileNet and maintain the same level of performance as Inception. Although the paper presents results for people detection, the model can be applied to other domains as well.