Multipart person detection for video surveillanceapplications

Abstract: CEA LIST is a research institute aiming at developing new and innovative technologies in several domains such as interactive systems, embedded systems and signal processing,it works in close collaboration with industrial partners to provide them with solutions matching their technological needs. In this thesis we describe how we integrated a persondetection module in the institute’s C++ library. We present several existing approaches for person detection and we describe the one we chose to implement. We present the results of our detector on the Pascal VOC 2007 and Caltech Pedestrian Dataset. In a second time we present severalal ternatives to improve the speed of the detector andshow how implementing a cascade resulted in a significant improvement of detection type. Finally we discuss several ways to tackle the issue of detecting partially occluded persons