Particle Filter Object Tracking Based On Self-adaptation Mechanism Of Human Vision

Object tracking is a significant subject in research of computer vision; it is the base of understanding target behavior and the important part to ensure that image system works continuously and accurately. Tracking algorithm requires accuracy, robustness, and in many scenarios immediacy. As the motion of object in reality is usually complex, taking into account accuracy, robustness and immediacy is both front edge and hot spot of object tracking all the time, and it is the main objective of this paper as well.Currently most algorithms are based on single feature. Although their computational costs are not great and can meet real-time requirement, the lack of accuracy in depicting object makes it difficult to divide the object from background; because the saliencies of features in different scenes are distinct, at present multi-feature algorithm commonly makes modeling for every feature respectively and merges information of all features together in order to complement each other, but it is required to calculate all features one by one, which will enhance the amount of calculation significantly and impact real-time tracking. In view of the problems in traditional tracking algorithm, this paper has mainly done these jobs as follow:Aiming at inaccuracy of tracking algorithm based on single feature, especially when the object is shaded which is likely to lead to failure of detection and tracking, considering that particle filter algorithm based on posterior probability evaluation can improve robustness, the paper takes measures to merge motion features into particle filter. By recording historical information of object motion features, when the object is shaded, algorithm will adopt Newton’s law of motion to evaluate the probable positions and directions of objects, solving the problem of discontinuous tracking when shading occurs.Referring to the mechanism of human visual tracking, that saliencies stimulated by target object in different scenes are diverse and human eyes capture targets according to their salient intensities, this paper has proposed a self-adaptation object tracking algorithm based on multi-features particle filter. The algorithm sorts saliencies of features by their intensities at first; then sifting particles by sorted features, the higher the saliency the more the discrimination, accordingly the more particles that do not matched with the object will be eliminated, reducing repeating calculation of particles until remaining the last one; ultimately feeding back sifting result, the more particles the feature sifts the higher saliency it will gain, and conversely its saliency will decline. The algorithm will sort those updated saliencies of all features and its result be the reference of the next step. By doing so, as the scenes change unremittingly, the features and saliencies of target object can adapt themselves to the external environment automatically, and consequently realize self-adaptation.For the purpose of verifying the proposed algorithm, we have collected four pieces of video, and contrasted with single and multi-feature tracking algorithm in accuracy, robustness and immediacy respectively. Tracking targets in two of the videos are people who are in athletic contests; the others are automobiles in highway and urban road. The experiment indicates that, in contrast with single feature tracking, the proposed algorithm is better at accuracy and robustness, especially in some complex scenes or sharply changing images; at the same time, in comparison with other multi-feature tracking, the proposed algorithm has higher performance in real-time tracking.