Research On Object Tracking Based On Compressive Sensing

This thesis studies single-target tracking technique based on compressived sensing.Based on compressived tracking,the thesis has made detailed analysis and investigation on object appearance description,feature selection and angle adaptation.The thesis descript target by combining superpixel structured information with haar-like features,which improved the tracking accuracy effectively.The tracking robustness is effectively improved by selecting the most discerning features via Boosting algorithm and the target frame is adaptively changed with the movement of target by advanced SURF algorithm.The multi-object tracking system is completed by the introduction of background modeling to detect the target and hierarchical association between the trajectory and detection responses.The main research work is as follows:(1)Compressive tracking based on super-pixel structured information is proposed.Super-pixels can cluster a certain degree of polymerization of adjacent pixels with similar texture,color and brightness,combined with the middle-level cues of the super-pixels and low-level clues of the haarlike features to characterize the target appearance,The Euclidean distance of the super-pixel patchs are extracted to characterize the target scale.Finally,a special measurement matrix with super-pixel structured information is constructed to reduce the dimensionality of haar-like features.The final target postion is determined according to the na?ve Bayesian classifier.(2)Compressive tracking via feature selection based on Boosting algorithm is proposed.According to the Euclidean distance between the extracted positive sample features and negative sample features,we use the main idea of Boosting,weak classifier combination into a strong classifier,to select the most discerning features to represent the target.At the same time,we propose the learning factor adaptive update strategy to provide occlusion judgment processing so that the learning factor can adaptively meet the speed of target.(3)Compressice tracking with adapative angle changes based on advanced SURFF algorithm is proposed.Laplace operator and LoG operator are introduced to SURF descriptor operator,which keep the rotation invariance of SURF and increase the calculation speed of SURF.Calculating the angle variation factor between the target of two consecutive frames,and then generating the rotation matrix.Finally,according to whether the number of successful feature matching is satisfying a certain fixed threshold to determine whether the target position needs to be rotated,which achieve target tracking with adaptive angle changes.(4)A multi-target tracking algorithm based on motion detection and compressed sensing is proposed.Introduce moving object detection in the compressive tracking,adopt the background modeling method based on sample consistency,and use the minimum bounding rectangle corresponding to the difference image of the current frame and the background frame as the detection result.Then use the Hungarian algorithm to solve the minimum problem of the cost matrix in the layer association process.Finally,the compressived tracker is used to rack multiple best matching targets that are detected,and multi-target tracking is completed.The thesis chooses some public video sequence on OTB and MOTB to experiment and verify the effect of proposed algorithms above.The experimental results show that compressive tracking combined with super-pixel can not only adapt to scale conversion but also improve the accuracy of real-time tracking performance.Compressive tracking via feature selection based on Boosting algorithm can effectively improve the robustness and accuracy of tracking.And compressive tracking based on improved SURF algorithm can effectively realize the tracking angle adaptation.The motion detection based on background modeling can detect the target better.The hierarchical association between the detection result and the track to be measured can effectively improve the accuracy of multi-target tracking.The last part of the thesis gives a short summary of the full-text work and looks forward to the follow-up research on the subject.