Dynamic scene analysis using Kalman filter and mean shift tracking algorithms

Video surveillance in a dynamic environment is one of the current challenging research topics in computer vision. In video surveillance, detection of moving objects from a video is important for object detection, target tracking, and behavior understanding. The present work is about locating a moving object (or multiple objects) over a time using a stationary camera and associating it in consecutive video frames. In this perspective, a video captured by digital camera is used for motion analysis. In the first stage of experiment background subtraction and frame differencing algorithms are used for object detection and its motion is estimated by associating the centroid of the moving object in each differenced frame. Tracking of non-stationary foreground regions is one of the most critical requirements for surveillance systems. In the second stage of experiment same algorithm is chosen for object detection but motion of each track is estimated by Kalman filter. However the best estimate is made by combining the knowledge of prediction and correction mechanisms that were incorporated as part of Kalman filter design. Subsequently kernel based tracking using mean shift theory is implemented for tracking single object under partial occlusion. Histogram based target representations are regularized by spatial masking with an isotropic kernel. The masking induces spatially-smooth similarity functions that are suitable for gradient-based optimization. In this regard a metric derived from the Bhattacharyya Coefficient is used as similarity measure, and subsequently mean shift theory is used to perform the optimization. In order to improve the track efficiency, an object tracking algorithm using Kalman filter (KF) combined with mean shift (MS) is also proposed. Firstly, the system model of KF is constructed, and the center of the object predicted by KF is used in MS algorithm for finding the target in the frame. The result obtained from the mean shift is given to KF as a measurement and is correctly updated using correction technique. The corrected value is taken as a reference position by mean shift for finding the object location in the successive frame. Again the obtained position from the mean shift is sent to KF for correction. The idea of combining Kalman filter theory and mean shift theory has given a direction in bringing out the efficient and reliable tracking results in case of partial occlusion.