Research On Algorithm Of Moving Object Tracking In Complex Scenes

As an important basic branch of computer vision,moving object tracking has been widely used in traffic control,video surveillance,face recognition,human-computer interaction and other fields.Despite the target tracking technology has made great progress,but due to the influence of complex environmental,such as occlusion,background interference,illumination,and target attitude and change the size,angle changes and other factors,makes the target tracking still faces huge challenges.Some difficult problems in single target tracking are studied in this dissertation and the main contents are as follows:First,a color-guided object tracking algorithm is proposed to solve problems such as object drifting and losing in complex environment,which is based on particle filter.Firstly strong object and weak object are separated based on the relationship of color feature between object and background.Secondly,according to the object status,the self-adaptive object model is constructed by tailored feature that includes CNN feature produced by convolution neural network with fixed kernel functions to describe object generalization property,HOG feature to describe object personality property and color feature.Then the search strategy of spatial consistency under the guidance of color feature is applied to get tracking result.The final predicted object is obtained by utilizing the size of bounding box in the mathematical space.The proposed algorithm reduces background noises and improves tracking accuracy with changed object appearance.Eventually experiment results demonstrate the effectiveness of the proposed algorithm.Then,for the object loss problem in the tracking process caused by illumination,occlusion,pose variation,and motion blur,the tracking method based on dual fuzzy low-rank approximation in a particle filter framework is proposed.Firstly,multiple constraint regions are built to filter insignificant samples,and more distinguished candidate samples are selected.Secondly,dual fuzzy observation function of each candidate sample is created based on the designed low-rank approximation representations of object and background.And then the generalized tracking results are obtained by computing membership degrees of dual fuzzy observation functions.Finally,based on the spatial coherency principle,the final tracking result is determined from the generalized results by measuring similarities of consecutive objects.The proposed method shows good performance as compared with several state-of-the-art trackers on challenging benchmark sequences.Next,for a solution of tracking drift and object loss that are resulted from factors such as environmental interference and appearance change of object,a method of object tracking algorithm via object saliency and adaptive background constraint is proposed.Within the tracking framework of particle filter,firstly the pixel characteristics of the object and the extended object are weighted to construct the explicit model of the object according to the principle of Bayesian saliency;secondly the background around object is considered adaptively by exploiting the saliency of the background;and lastly by judging the current appearance state of object,the tracking result is obtained using the correlation between the object and the background.Matching error can be reduced by the object saliency model and tracking accuracy can be improved by the adaptive constraint of background as well,especially when object is occluded and pose is changed.The experimental results suggest that the proposed method has a stronger robustness and a higher precision of object tracking.Finally,sparse representation-based methods have been successfully applied to visual tracking.However,the over complete dictionary mode is single and data is large,the sparse coefficients need to be solved by the complex optimization algorithm,which will limit their tracking performances.In this paper,within the tracking framework of particle filter,we propose a tracking method based on multi-modality dictionary learning.Firstly,a long and short period of object templates are created,which combined with background templates to form a multi-modality dictionary to characterize the current state of sampled object.Secondly,according to the multi-modal coefficient between the sampled objects and the dictioanry,the target is tracked roughly and the candidate tracking results are obtained.Lastly,the observation likelihood functions of the candidate tracking results and the multi-modality dictionary are constructed by using LOMO features,and the candiadate tracking result with the maximum likelihood is taken as the final tracking result.The experimental results demonstrated that the proposed method has strong tracking robustness in the case of occlusion,illumination change and background interference.