Research On Robust Single Object Tracking Algorithm In Complex Scenes

Object tracking is an important research direction in the computer vision and widely used in video surveillance,human-computer interaction,driverless and military fields.It establishes a model based on video information of image sequence,then determines the position and posture of the interested object in each frame according to the spatial-temporal correlation to obtain the trajectory of target,which lays a foundation for higher-level processing in video surveillance system.With the development of over 20 years,object tracking has made many progress and breakthroughs.However,it is still a challenging task to build a stable and efficient target appearance model,which can effectively cope with the appearance variations such as illumination,occlusion,and deformation in complex and changeable scenes.This is also a key element in object tracking.Therefore,the thesis combines compressive sensing theory,correlation filter and deep learning technology to investigate how to design and learn robust target appearance model based on discriminative tracking methods.Firstly,according to the efficiency of the compression tracking method,there are defects of insufficient performance,we propose a multi-model real-time compressive tracking(MMCT)algorithm,which adopts the compressive sensing to decrease the high dimensional features for the tracking process and satisfy the real-time performance.Moreover,The MMCT algorithm selects the most suitable classifier by judging the maximum classification score difference of classifiers in the previous two frames,and enhances the accuracy of location.The MMCT algorithm also presents a new model update strategy,which employs the fixed or dynamic learning rates according to the differences of decision classifiers and improves the precision of classification.The experimental results show that the MMCT algorithm can well adapt to illumination,occlusion,background clutter and plane-rotation.Meanwhile,the introduced multi-model does not increase the computational burden and still shows excellent real-time performance.Secondly,in order to further improve the overall performance of the tracking,we exploit the deep learning and correlation filtering techniques,and propose spatial and semantic convolutional features for robust visual object tracking.We exploit spatial and semantic convolutional features extracted from convolutional neural networks in continuous object tracking.Besides,we construct the multi-scale pyramid correlation filter by extracting spatial features of target,which effectively determine the scale level.Furthermore,we present a novel model updating strategy,which exploits peak sidelobe ratio(PSR)and skewness to comprehensively measure the fluctuation of response map.The strategy guarantees the freshness of the model and avoids unnecessary updates.The experimental comparisons on OTB-2013 show that our algorithm performs favorably against 12 state-of-the-art trackers.Finally,due to the deep learning and the correlation filtering method,there is a limitation that the model is too single to adapt too many challenges,we introduce the classifier re-detection model,and propose dual model learning combined with multiple feature selection for accurate visual tracking.We fuse the hand-crafted features with the multi-layer features extracted from convolutional neural network to construct correlation filter models based on multi-level feature fusion,which can precisely localize the target.We also propose an index named Hierarchical Peak to Sidelobe Ratio(HPSR).The fluctuation of HPSR determines the activation of an online classifier learning model to redetect the target.The target locations predicted by the dual learning models mentioned above are combined to obtain the final target position.With the help of dual learning models,the accuracy and performance of tracking have been greatly improved.The results on the OTB-2013 and OTB-2015 datasets show that the proposed algorithm achieves the highest success rate and precision compared with the 12 the state-of-art tracking algorithms and is better adaptive to various challenges in visual object tracking.