Research On Optimization Of Precision For Visual Tracking

With the rapid advancement with computer performance,the application of computer vision system has been widely found in more and more aspects of lives such as vehicle navigation,video surveillance,human-computer interaction,robotics,so on and so forth.Target tracking algorithm has been in a center for both theoretical research and practical applications.However,in the realworld,due to the large variety of influencing factors such as the complexity of the backgroundand-foreground scene and the variation of target motion,the great challenges remain to achieving the robust and accurate real-time tracking of the targets.In this paper,the author focuses on overcoming the shortcomings of the original STC algorithm applied in the process of tracking,and building the model to fuse saliency detection with spatio-temporal context,which can better take image features into account for accurate tracking.Furthermore,by combining the scale transform and motion prediction,the performance is improved and the satisfactory accuracy of target tracking is achieved.The research content can mainly be divided into the following three parts:1.Due to the fact that the generic STC fails in stably tracking the target when the target rotates out of the predefined shape or is occluded,or when the target is motioning with high mobility as manifested by varying speed and orientation,and by the changing scale of target,a robust visual tracking approach via joint the covariance moment features with spatio-temporal context learning algorithm is proposed.The STC algorithm are first used on the given video sequences to track the target of interest,and then the results are utilized as the selection criteria,which determines whether or not to turn to using the covariance moment features.The target size is adaptively determined,and then the Mean-shift algorithm is utilized to determine target center location for positioning purposes.The experimental results from testing on video clips from public databases indicate that the presented algorithm outperforms to a certain extent some existing algorithms.2.In order to better take advantage of image feature information and improve the tracking accuracy,as well as robustness of target tracking,an improved super pixel tracking approach via fusing salient region detection into spatio-temporal context is proposed.In this work,the approach is achieved by firstly conducting super pixel segmentation in the context region of target,then using motion relevance of target context and the regional covariance information to calculate the correlation saliency of the image super pixels.Based on Bayesian framework,the model of fusing saliency detection into spatio-temporal context is built in the frequency domain;then the color and texture histograms of current frame and reference template is employed to calculate the Bhattacharyya coefficient,which is intended to update the spatial and temporal context model.Next scale pyramid model is induced to estimate the target scale.3.In order to solve the problem of the target with high mobility,the super pixel tracking algorithm via fusing saliency detection into spatio-temporal context is proposed by incorporating the adaptive low-pass filtering motion prediction module.By updating online dynamic model sample set and using ridge regression method,the parameters of a low pass filter is determined,and the adaptive motion prediction is implemented in a bid to improve the accuracy of the predicted target location.The results from testing the presented approach on public database indicate that the proposed method can achieve better performance than some existing algorithms to a varying extent.The developed approach is expected to be further applied to visual target tracking in more sophisticated conditions,such as illumination change,complex background,object rotation,high mobility,low resolution and so on.In addition,the estimation of target scale is also improved to a certain extent,as presented in this thesis.