I. Introduction
Motion tracking poses many challenges and is vital part of computer vision problems such as motion analysis, event detection and activity recognition. Tracking can be defined as an algorithm to monitor the spatial and temporal changes in a video sequence such as shape, size position, etc. Tracking constraints often include occlusion, noise, moving background etc. which degrade the performance. The proposed algorithm makes use of a simple tracker based on block motion estimation. However, the tracker allows the use of Kalman estimates whenever a mistracking is detected. The proposed method makes the tracker less complex by only relaying on the much intense Kalman when needed. The Kalman uses the predictive estimates to fill in when the block tracker yields erroneous interpretations. Block motion based tracking algorithm have proven to be computationally less intense and faster than other trackers. However, the block motion based tracking often suffers from low accuracy due to more errors. The unconstraint tracker often fails when the object changes direction suddenly or the velocity of the object suddenly changes. During these times the proposed algorithm let Kalman estimate take over thereby improving the accuracy. The proposed work is very useful when there is constraint for hardware upgradation or restrictions and performance is required in real time with minimum delay. Most of the tracking algorithms have to run in real-time and are often limited by computational load that imposes on the algorithm. Kalman filter when used alone for tracking can be sometimes computationally expensive which makes real time application difficult. The proposed method works with a computationally simpler tracker and adapting Kalman filter when needed. Figure 1 presents the schematic of the proposed approach. The input video is subject to motion estimation and extracted motion vectors are used in block matching algorithm where the motion vectors are converted in to moving blocks. The blocks from current frame is matched with similar blocks in the next frame there by reducing computational complexity. The sum absolute difference(SAD) between grey values determines whether the motion exists in successive frames. Also, new motion vectors are searched in successive frames. Whenever the tracker finds deviant from normal behavior with the motion vectors the tracker switches to Kalman tracker which is idle to that point. The Kalman tracker uses its tracking estimates to make sure that the abnormal deviant values don't affect tracking results. If the deviation persists for a longer period then Kalman tracker will return the tracking control block to the block based tracker. The proposed algorithm outperforms existing tracker algorithms as shown through experimental results.