Improved occlusion handling for human detection from mobile robot

This paper presents an approach for detecting multiple persons from a mobile robot using a combination of depth and thermal information. Previous works were mainly based on RGB and thermal data, but in this work we present a system where we combine depth and thermal data for occlusion handling. First, a classifier is trained using thermal images of the human upper-body. This classifier is used to obtain the bounding box coordinates of human. Initial detection process in thermal image is used as an estimate of the position of a person. The depth image is later fused with the region of interest obtained from the thermal image. Using the initial bounding box, occlusion handling is performed to determine the final position of human in the image. The proposed method significantly improves human detection even under occlusion.