ROBUST PEDESTRIAN TRACKING USING IMPROVED TLD ALGORITHM

Abstract

Object tracking is defined as the estimation of location of an object of interest in image sequence, whose initial position is defined in the first frame. It has applications in the field of surveillance, traffic management, sports event monitoring and most recently in Driverless assistance systems .Most existing digital video surveillance systems rely on human observers for detecting specific activities in a real-time visual scene. However, there are limitations in the human capability to monitor simultaneous events in surveillance displays. Hence, human motion analysis in automated visual surveillance has become one of the most active and attractive research topics in the area of computer vision and pattern recognition. Earlier methods of object tracking used either tracking or detection, neither of which were independently sufficient for tracking under complex situation .TLD suggested by kalal et.al. [1] is an award winning technique in which ,tracking and detection are integrated along with a new learning component called P-N learning .These 3 components forms a strong feedback loop. Comparison of TLD with earlier tracking methods shows that TLD has better performance in many aspects of difficulty such as - long lengths of video, occlusion, zoom and background clutter. In this thesis, implementation of the TLD algorithm is described in detail. It is evaluated specifically from pedestrian tracking prospective. We extended TLD to track multiple objects. Harris features [2] are added to benefit the algorithm by providing robustness to out-of plane rotation of the object in image sequence. Automatic initialization is also accomplished using Histogram of oriented gradients [3]. Experiments are conducted on complex datasets and results are compared. Implementation results shows that the final result is robust to occlusion and worked on frame rates comparable to real life scenarios. The final implementation also provides the trajectory of each pedestrian which can be used for crowd flux analysis.