Research Of Visual Tracking Based On Video Image Sequence

Vision is an important way for human to acquire the information and perceive the world.Visual tracking based on video image sequences is the key technology of image processing,largely required and widely used in biomedical science.To improve tracking performance in real-timeliness,precision and robustness,numerous state-of-the-art tracking algorithms have been proposed in recent years.However,this technique remains a challenging task because of various interferences,such as scale variations,illumination variations,occlusion,and fast motion.This paper focuses on the problem of efficient multi-scale tracking using Haar-like features.To address this issue,we propose a novel Coarse-to-Fine tracking(CFT)approach with three highlights.First,visual representation based on Haar-like feature is investigated in this paper.The relationship between Haar-like feature and illumination is analyzed and therefore,a simplified illumination observation model is built.Haar-like features are improved in illumination invariance on the basis of the observation model.A two-stage feature selection method is introduced for informative and discriminative Haar-like features.Second,the mechanism of visual tracking algorithm is explored by comparing generative appearance model with discriminative appearance model.CFT is designed by combining “coarse” tracking based on particle filtering with “fine” tracking based on a discriminative classifier,which simulates human vision mechanism efficiently from “coarse” scale to “fine” scale.CFT improves the efficiency of multiscale visual tracking efficiently by optimizing the assignment of computing resource.Third,the training and updating methods of discriminative appearance model are researched.A weighted bootstrap sampling strategy is adopted;this strategy improves the precision and separability of the discriminative appearance model.On the other hand,a simple yet effective adaptive threshold method is developed;this method can robustly detect large appearance variations and preserve the appearance models from noisy appearance variations.Experimental results on various benchmark challenging sequences verify the superior performance of CFT over other state-of-the-art methods.