Object Tracking Algorithms On Single Camera

Target tracking is a needful technology, no matter in the modern defense, sea and air traffic control system, or in the field of civilian security monitoring, intelligent system of human-computer interaction. Now, target tracking has become a very active research field, and has received much attention of more and more nations. The object of target tracking is to get the location of the targets, build target trajectories(tracklet), and distinguish every target in the whole video sequence.Many problems are remained in target tracking such as the background interference, deformation of target itself, occlusion, and the illumination change. In this paper, two kinds of target tracking algorithm are presented:(1) Category Free Tracking(CFT), in which target's category needn't to be pre-known, and there is also no requirement for pre-trained specific detector in off-line fashion. After manually getting target area in the first frame of the video sequence, CFT performs tracking by utilizing the appearance model of the target or by obtaining foreground target from the background, and adopts filtering algorithm to predict the target.(2) Association based Tracking(ABT), in which target's category needs to be known before tracking, and the specific detector is used to obtain detection responses of the video sequence. The final tracking result is achieved by optimal association among detection responses of video frames.First, three classical target tracking algorithms of CFT are introduced. Then a scale and orientation adaptive tracking algorithm is realized. The adaptive tracking algorithm works under framework of the classical Mean Shift tracking. The algorithm extracts moment features from weight image of candidate target area, estimates scale and orientation adaptively. We compare our method with two classical algorithms(Mean Shift and EM-shift), and the experimental results show that this algorithm not only inherits advantages of the original Mean Shift tracking algorithm, such as simple and efficient, but can effectively track the target with changes in scale and orientation of target.For ABT, in order to precisely compute the similarity between a pair of tracklets and improve the accuracy of target tracking, this paper proposes to establish discriminative appearance models. Firstly, reliable tracklets are obtained from detection responses of each frame. Then training samples are collected from these tracklets. We merge several features to describe the training samples robustly, and use the Adaboost algorithm to train classifier, i.e. discriminative appearance model, in on-line fashion. Finally, the discriminative appearance model is used to link the tracklets into longer ones to form the final complete target trajectories by an iterative process. The experiment results indicate that compared with some state-of-art methods, the proposed multiple target tracking algorithm shows improvements in robustness, and has satisfactory performance in complex scene.