Research On Local Feature-based Tracking Methods For Infrared Target

Infrared target tracking is to find out the position, velocity, size and motion trail of a certain target in infrared video sequence, which implements the function of an understanding and description of target behavior and then carries out advanced tasks,it is widely used in many military visual guidance, robot visual navigation, safety monitoring, traffic control and medical diagnosis field. In practical application, being affected by complex background, target variation(e.g. scale variation) and external change(e.g. occlusion and illumination change), infrared target tracking algorithm with robust and accuracy is still facing a wide range of problems.To solve the problems above, this paper introduces an improved infrared target tracking algorithm, which is based on KF(Kalman Filter) and SURF(Speeded Up Robust Feature). The innovative contributions of this are as follows:1.A tracking algorithm combining visual attention mechanism with weighted SURF is presented. The algorithm first matches SURF features in current frame with the counterparts in target template, then assigns a decreasing sequence of weight values from the close-by examples to those far off successively, which is based on the distance of target center from SURF in current, and finally achieves target location and tracking box with size self-adaption according to location and scale information of weighted SURF.2.A scale self-adaptive algorithm based on weighed SURF is proposed for infrared target tracking. The method first obtains a region of interest as the target template, and employs Kalman filter to predict the target location in current frame, on which a to-be detected region is located. Then it extracts SURF features in the region and match them with counterparts in target template, and finally obtains an accurate target location and scale information by the algorithm mentioned in 1. The improved algorithm efficiently sieves out some unreliable features, which would affect location even frustrate tracking process. Abundant experiment has proved that the algorithm performs well in scale-varied target tracking with accuracy.3.An anti-occlusion algorithm based on five-rectangle of gray scale model is proposed for infrared target tracking. Firstly, a tracking box which is the same size as in previous frame is built in the location determined by Kalman prediction, and another four tracking boxes of the same size are built, with their centers at the midpoints of four sides of above box, which determine five temporary candidate regions in all. Secondly, similarity coefficients of gray histogram between target template and these regions are calculated respectively, and the region of maximum value is chosen to be target candidate region. Finally, target area and location is obtained by the algorithm mentioned in 1. Experiment results show that the algorithm has an accurate performance and robust to pre-occlusion, partial occlusion, serious occlusion and reappearance.