Spatial-Temporal Context Aware Object Tracking Method Of Correlation Filter

Object tracking is one of the most important problems in computer vision.It has been widely applied in many fields,such as video surveillance,autonomous driving,human-machine interaction,behavior recognition and so on.Recently,the correlation filter based trackers(CFTs)have made remarkable progress,and they have become a hot research spot.Although a large number of methods have been proposed,the complexity of tracking itself,the variability of scenarios and the diversity of application background make it challenging to achieve robust and efficient object tracking,especially for frequent,or heavy occlusions and the re-detection problems in long-term tasks.To solve the model drift problem,mainly caused by the lack of prior information,occlusions and long-term tracking,this thesis focuses on the correlation filter theory to study robust object tracking methods from the perspective of constructing robust and efficient appearance models,and by reasonably utilizing the spatial-temporal context information of the target.The main works and contributions of this thesis are summarized as follows:(1)A context-aware occlusion detection based object tracking method is proposed.To solve the occlusion problem frequent occurred in tracking,a context-aware occlusion detection method is proposed.Based on the multi-feature kernelized correlation filter(KCF)to estimate the initial position of the target,this method fully uses of the context information and the stability of the continuous frames of target to predict the occlusion state.The correlation filter reflects the intrinsic relationship between the most reliable target and its context,along with the correlation response stability constraint to estimate the current target state.Then,the appearance model is updated adaptively according to the occlusion state and the changes of the appearance between consecutive frames.The experimental results show that the proposed occlusion detection strategy can effectively detect the current state of the target by using the context information.The response reliability constraint between consecutive frames of the target makes the detection result more stable.Both of them work well to improve the tracking accuracy and the success rate.(2)A saliency detection based adaptive object tracking method is proposed.To solve the model drift problem caused by the lack of priori information,visual saliency is introduced into the KCF framework,which is implemented through fast saliency detection,salient object extraction,candidate proposal estimation and optimal location determination,to rectify small inaccuracy in case of low tracking confidence.Then,an occlusion discrimination factor is designed based on the correlation response variations,and the fluctuations of the tracking confidence between consecutive frames are taken into consideration at the same time.Both factors are introduced into the model update process to adaptively adjust the learning rate.Experimental results and analysis show that visual saliency highlights the object of interest in the context,reduces the interference of complex background,and corrects the inaccuracy of initial tracking results.Adaptive updating strategy helps address the occlusion problem as well as maintain the most reliable target appearance.Compared with 12 state-of-the-art trackers,the proposed method yields better tracking performance while operating at a real-time speed.(3)An enhanced occlusion handling and multi-peak re-detection for long-term object tracking is proposed.Long-term tracking tasks are still challenging,especially in areas of occlusion.Herein,an adaptive complementary learners based appearance model is first constructed,which is based on the global color model and the multi-feature discriminative correlation filter(DCF).Each model is trained independently and combined by adaptive merging,and considers the reliability of each representation to provide a preliminary estimation.Secondly,an average occlusion distance metric is proposed.When occlusion is detected,the multi-peak re-detection module is activated,cooperating with the long-term memory preservation mechanism,to recover the target in case of tracking failure.The re-detection module is designed based on the multi-peak property of the merged response,which helps obtain multiple candidates spatially.Long-term memory preservation mechanism maintains the model reliability temporally through adaptive updating,to enable the recovery of the target in case of tracking failure.The experimental results demonstrate that the proposed method handles the heavy occlusions in long-term tracking effectively and it performs favorably against 17 state-of-the-art trackers while running efficiently in real time.(4)A collaborative short-term spatial-temporal correlation model and a long-term re-detector based object tracking method is proposed.Most existing trackers perform well in the presence of short-term occlusions and significant appearance and illumination changes,but struggle with vital challenges in long-term tracking.These include frequent or long-lasting occlusions,and the disappearance of the target from view.Herein,a collaborative long-term tracking method is proposed,which includes three components: a short-term tracker,tracking uncertainty detection,and a long-term re-detector.The spatial-temporal information is explicitly encoded in each component,which operates collaboratively to achieve robust and efficient tracking.Firstly,a spatial-temporal correlation model is proposed as the short-term component for locating a visible target from frame to frame.Both spatial and temporal constraints are introduced to regularize the correlation filter and improve the discrimination power of the target and background.Following this,an uncertainty detection mechanism is presented,which relies on the variations of correlation response measured spatially and appearance similarity estimated temporally,to activate the long-term re-detector when necessary.Finally,a spatial-temporal aware long-term re-detector is proposed.The candidates are sampled by particle filter,and refined by correlation shepherding and nearest neighbor merging.The target is recovered by template matching and confidence re-ranking when it becomes visible.The target template maintains a long-term memory through adaptive updating,which further improves the perceptual recovery ability of the proposed method.The extensive experiments demonstrate that the proposed method performs favorably in comparison with 18 state-of-the-art trackers and effectively deal with frequent or long-lasting occlusions problems in long-term tracking.