Resolving occlusion ambiguity by combining Kalman tracking with feature tracking for image sequences

An improved method for mitigating occlusion in object tracking is proposed in this paper. Using more traditional methods of object tracking and feature detection, a novel scheme is developed based on the use of multiple tracking methods which operate in parallel to minimize the effect of occlusion. A popular feature detection algorithm called Histogram Oriented Gradients (HOG) is used as our baseline tracker. Its ability to detect and track objects during occlusion is then enhanced with a Kalman filter. State variables for Kalman include position and velocity of the object. Additionally, we introduce a new term called a Correlation Constant C(k) which makes use of a HOG trackers noise distribution to minimize the process and measurement variance of the Kalman filter. An online video database is used to experimentally verify our proposed algorithm [1]. Each video frame is provided with ground truth coordinates for the object being tracked. Results were developed in Matlab using online code developed by Henriques [2]. Experimental results show that our proposed algorithm is effective in solving the occlusion problem.