Research On Visual Target Tracking In Complex Scenes

Visual target tracking has always been a core question in the field of computer vision.It is widely applied in various fields of national defense military and the people's production and living,such as video surveillance system,intelligent transportation,precision guided system and human retrieval.It is still a challenging task to design a robust and efficient tracking algorithm due to target deformation,fast motion,sudden movement,illumination variation,camera dithering,background clutters,occlusion,scale variation and so on.In this paper,the problem of visual target tracking in the complex scenarios is discussed.The main achievements are as follows:(1)The classifier response of the target model with a single feature could be easily affected by factors such as illumination variation and occlusion.To solve the problem,a target tracking algorithm based on adaptive multi-feature fusion in tracking-by-detection framework is proposed.Features are extracted from the sub-images extracted by dense sampling,and the target appearance models are established respectively.The response of each model is obtained with regularized least squares classifier.The final response is achieved by weighted sums of the responses,in which the weights are updated by solving a regression equation.It helps to obtain accurate and stable detection scores by enhancing local discrimination.Experimental results show that the algorithm outperforms other state-of-the-art tracking algorithms in tracking accuracy and robustness in most complex scenes.(2)Outliers refer to a set of data whose values are quite different with others.It is usually caused by motion blur or partial occlusion during tracking.In order to alleviate the sensitivity problem of outliers,a novel target tracking algorithm based on adaptive observation weight is proposed in this paper.It can effectively alleviate the impact of the outliers on the target model.First of all,a weighted observation model is established by linear visual tracking representation.Then an iterative optimization algorithm is proposed to obtain the parameters of the model,and adaptively update the weight matrix to eliminate negative influences of observation outliers.Finally,an effective likelihood evaluation function is adopted to track the object accurately and robustly.Both qualitative and quantitative analyses on several challenging video sequences demonstrate that the algorithm outperforms other state-of-the-art tracking algorithms in tracking accuracy and robustness.(3)To reduce negative influences of deformation and fast motion in correlation tracking and solve the problem of poor robustness and low effectiveness of visual target tracking in complex scenes,a compensation algorithm for correlation tracking via target color model based on sparse reconstruction is proposed in tracking-by-detection framework.At present,some algorithms based on the correlation filtering framework is largely dependent on the spatial layout of tracked target.Therefore,algorithm is easy to failure due to the large deformation,the camera dithering and fast motion.Considering the statistics of color is very robust to spatial location of pixels,background template is taken to reconstruct the image region via super pixel division in the process of target tracking.Color correlation of target is obtained based on reconstruction error.Discriminative color target model is established to calculate detection score of tracking.The score is then combined with the correlation filter score via probability integration method to alleviate the influence of illumination variation of color target model.Finally,the precise tracking results are obtained.Experimental results show that the algorithm outperforms other state-of-the-art tracking algorithms in tracking accuracy and robustness as for most complex scenes.