Real time human tracking system

Abstract

The use of surveillance cameras in various locations has rapidly increased in its popularity. Since the purchase, installation and maintenance of these cameras has become relatively cheap, their use has increased rapidly. A problem which was encountered by various organizations investing in these cameras was the lack of manpower to surveil them. It is now no longer feasible to employ people to monitor these cameras and so the captured video has often been used as a forensic tool to analyse what happened after an event has already occurred. This problem of human surveillance can be solved by the implementation of intelligent cameras which can detect events at the time they happen. The implementation of such a system would enable people to be alerted immediately if an event occurs without the need of human surveillance. In this dissertation, a real-time human tracking system is implemented. The development of this system was segmented into three main parts. A motion detected module was used to obtain the location and shape of moving objects. Noise removal and shadow suppression techniques were used to eliminate background noise while at the same time retaining the detail in the detected objects. Once the moving objects are clearly identified, the object classification module was implemented to classify all objects as human or non-human. When testing this module using hundreds of images, an accuracy of 96.84% was obtained. The human objects are then tracked using the final module. The centres of the boxes enclosing the objects are used to keep track of the object location in consecutive frames. When no occlusion is present in the scene, a tracking detection rate of 96.45% was successfully achieved. In situations where occlusion between multiple object occurs, a tracking rate of 98.49% was obtained. This system was successfully designed to operate in a real-time environment with an average processing rate of 15 frames per second.