Human Body Tracking by Adaptive Background Models and Mean-Shift Analysis.pdfVIP

Human Body Tracking by Adaptive Background Models and Mean-Shift Analysis Fatih Porikli Oncel Tuzel Mitsubishi Electric Research Labs, Murray Hill, NJ 07974, USA Abstract We present an automatic, real-time human tracking and ob- servation system. Robustness and speed are the two major bottlenecks of the existing approaches. We improve upon the robustness and speed of the current state-of-art by in- tegrating a mean-shift based model update technique with an adaptive change detection method. We also provide op- timal solutions for several other stages including illumina- tion compensation, skin color detection, shadow removal, morphological filtering, event analysis of a tracking system. In addition, we introduce a novel background refresh mech- anism. Thus, the proposed framework is capable of han- dling shortcomings of template and correspondence based tracking approaches. The results with the ICVS-PETS data sets show the effectiveness of the algorithm. 1. Introduction Accurate object segmentation and tracking under the con- straint of low computational complexity presents a chal- lenge. A typical detection system is built by finding regions in motion, eliminating shadows and noise, constructing and tracking objects in video. Background Subtraction The most common approach for discriminating a moving object from the rest for sta- tionary camera setup is background subtraction. Basically, background detection approaches can be classified as non- adaptive and adaptive methods. Manual selection, pixel- wise voting, and mean value search algorithms are among the non-adaptive methods. Adaptive methods include aver- aging images over time, alpha-blending [5], Kalman filter- ing, Gaussian mixture models (GMM), etc. Although aver- aging and alpha blending are simple and fast, they are not effective for scenes with many moving objects particularly if they move slowly. Besides, they cannot handle multi- modal backgrounds. They recover slowly when an object occupies the scene