Performance of invariant feature descriptors with adaptive prediction in occlusion handling

Object tracking in computer vision plays an important role for automating the process of video surveillance and robot navigation. The trajectories of every moving object are analyzed to further interpret the events in a scene. Occlusion problem is always the main challenge that interrupts the tracking trajectory and reduces the tracking performance. Thus, this paper aims to investigate an improved performance of invariant feature descriptors in occlusion handling by using adaptive prediction from Kalman filter. The invariant feature descriptors that are extracted from a tracked object are robust against transformation and partial occlusion. These descriptors are combined with Kalman filter prediction to resolve full occlusion in object tracking. Unlike conventional Kalman filter prediction, the error covariance parameters are auto-tuned based on the changing conditions of the feature descriptors in a tracked object. Experiments are conducted to show the response of invariant feature descriptors during partial and full occlusion. The response rate is contributed as the benchmark for parameters tuning in Kalman filter prediction.