The saliency selection reliable particles are used to estimate the target state. Finally, experimental results demonstrate the efficiency and effectiveness of the proposed method in presence of occlusion and large illumination variation, even non-target with similar features. | Journal of Automation and Control Engineering Vol. 3, No. 5, October 2015 Adaptive Probabilistic Tracking with Visual Saliency Selection Reliable Particles Linshan Liu, Suiwu Zheng, and Hong Qiao State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China Email: {, , }@ However, color distribution becomes ineffective in the presence of illumination changes. Edges are sensitive to clutter and computationally expensive. Mahadevan and Vasconcelos [6] show that saliency feature played an important role in object tracking. For the better performances, one can combine saliency and other features as in [7]-[9]. In [7], an improved Itti model [5] is employed to measure the similarity between tracked object and candidate one, which was more accurate. In addition, target model was used as prior knowledge to calculate the weight set which was utilized to construct their saliency features adaptively [8]. D. Sidibé [9] introduces frequency tuned saliency features into particle filter. However, all of the approaches either suffer very large computational complexity in real applications or do not use the background information which is likely to improve tracking stability. Therefore, these methods are apt to be distracted by background with similar features which is likely to cause drift. Target model as prior knowledge may be ineffective when it changes drastically. In this paper, we provide an effective and efficient adaptive probabilistic tracking approach with saliency selection reliable particles. We introduce a binary, simple and holistic image descriptor called the “image signature” [10] to extract saliency features of the object. The proposed method estimates the weight of each particle through the saliency measurement implemented by Hamming distance between target model and each hypothetical observation, but also through background model and each .
