Submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Computer Science, School of Engineering and Technology

Research Studies Project Report (M.Eng.) - Asian Institute of Technology, 2009

As public concern about crime and terrorist activity increases, the importance of public security is growing, and video surveillance systems are increasingly widespread tools for monitoring, management, and law enforcement in public areas. The visual surveillance system has become popular research area in computer vision. There are many algorithms exists to detect and track people in video stream. Human detection and tracking in high density crowds where object occlusion is very high is still an unsolved problem. Many preprocessing techniques such as background subtraction are fail in such situations. I present a fully automatic approach to multiple human detection and tracking in high density crowds in the presence of extreme occlusion. We integrate human detection and tracking into a single framework, and introduce a confirmation-by-classification method to estimate confidence in a tracked trajectory, track humans through occlusions, and eliminate false positive detections. I search for new humans in those parts of each frame where humans have not been detected previously. I increase and decrease the weight of tracks through confirmation-by-classification process. This is helpful to remove those tracks (false positives) which not confirmed for long time. I keep the tracks which have not been confirmed only for a short period of time, to give chance to those tracks where human are occluded fully or partially for a short period of time to rejoin their tracks. We use a Viola and Jones AdaBoost cascade classifier for detection, a particle filer for tracking, and color histograms for appearance modeling. An experimental evaluation shows that our approach is capable of tracking humans in high density crowds despite occlusions. On a test set with 35.35 humans per image achieves 76.8% hit rate with 2.05 false positives per image and 8.2 missed humans per image. The results form a strong basis for farthar research.