DPM Score Regressor for Detecting Occluded Humans from Depth Images
- Tsuyoshi Usami, Hiroshi Fukui, Yuji Yamauchi, Takayoshi Yamashita, and Hironobu Fujiyoshi
- Korea-Japan Joint Workshop on Frontiers of Computer Vision, 2016
Download: PDF (English)
A part-based object detection method called deformable part models (DPMs) is known as a robust method for detecting objects that have detection method for posture variation.In the detection stage, the DPMs assume that all parts are visible.If some parts of people are partially occluded by an object such as a table or wall, detecting them becomes difficult.This paper proposes a robust method for detecting occluded humans to regress scores with a reduced influence of occlusion.We apply 3D raster scanning to depth images for finding occluded regions.We compute occlusion rates in each part of humans as the occlusion rate from regions.We regress from the DPM detection scores, DPM root scores, DPM part scores, and occlusion rates as explanatory variables to enable detecting the scores of humans with no occlusion.Using these detection scores makes detecting humans easier. Experimental results show that the precision of the proposed method was improved by almost 20% compared with that of conventional DPMs.