Head-Body Correlation for Robust Crowd Human Detection

A simple yet effective proposal-based human head and body detection framework is developed in crowded scenes. Human heads are too small for detectors to locate and human bodies are frequently occluded in the crowds, which requires more robust location capability of detectors. To tackle the issues above, we propose a head-body correlation module to utilize the location prior knowledge of human body and human head. Compared with Faster R-CNN, our approach can improve the Average Precision (AP) gains for human body and head detection by 2.15% and 2.52% on the challenging CrowdHuman dataset.