Human Tracking within a Crowd

Abstract

The labour cost required to supervise surveillance cameras is by far outgrowing the ever decreasing cost required to purchase, install and maintain such equipment. Consequently, the development of intelligent video surveillance systems is gradually replacing the traditional cameras and the manpower required to monitor the video footage. Suspicious motion or activities are detected automatically as they occur and the appropriate security personnel are alerted as soon as salient events happen so that they can take immediate action. As a result, security in public places is further enhanced, alleviating human errors like fatigue and negligence which can occur whilst monitoring the traditional cameras. In this dissertation, a tracking system which is able to track people in medium to high density crowd scenes is implemented. To uniquely identify the target person, the proposed tracking system first generates its respective feature distributions, where the colour and the texture features are unified homogeneously for increased accuracy. The particle filter method is then used to randomly launch particles in order to identify the search regions within the next frame. The respective feature distributions are extracted for each propagated particle and a distance metric is utilised to determine the best match by means of a correspondence based matching. To overcome the problem of occlusions, which is very common in high density crowded environments, an adaptive threshold method was used. The human tracking system presented in this dissertation has achieved very promising results, where an average Track Detection Rate (TDR) of 97.83% and 89.98% were registered for non-occluded and occluded scenarios respectively. Furthermore, the average Occlusion Success Rate (OSR) for occluded scenarios is 84.62%, indicating that the proposed system can be used to detect and track humans within a crowd.