Regularization Methods For Non-Specific Visual Object Tracking

Video object tracking is a fundamental problem of computer vision and provide strong support for intelligent applications,aiming to accurately and e ciently obtain the status of targets such as location and size from a real-time video.However,since the target is non-specific and constantly changing and the background are complex,this problem is significantly challenging.How to construct an accurate and robust appearance model e ciently via few-shot examples is the most important problem of visual object tracking.This dissertation studies how to design effective regularization methods for video object tracking by considering prior information,such as the target structure,spatial distribution and spatial-temporal changing of the target and context,to construct accurate,robust and real-time trackers.In summary,our main contributions are as follows:1.We propose a superpixel-guided structure regularization method for accurate compressive tracking to include target structure information and obtain more discriminative features.We find that superpixel-guided compressive projection produce powerful features by capturing rich local structural information of targets.Then,we propose fast directional integration to realize object rotation detection.Finally,we present online data-driven sampling strategies to produce less yet more e?ective candidates and training samples.Extensive experiments on real-world benchmark datasets validate the superior performance of the proposed approach over state-of-the-art trackers.2.We propose a selective spatial regularization(SSR)method via the Markov Decision Process(MDP)for robust correlation filter based tracking to take advantages of the spatial distribution of the target and its context.First,we extend the objective function of correlation filter based tracking scheme to learn the target-contextregularized filters using target-context-driven weight maps.We then formulate the online selection of these weight maps as a decision-making problem by a Markov Decision Process(MDP),where the learning of weight map selection is equivalent to policy learning of the MDP that is solved by a reinforcement learning strategy.Finally,the proposed SSR is used to equip two popular trackers to significantly boost their tracking accuracy,while achieving an average 5 times speeding-up.3.We propose a truncatedly iterative meta-learning based fast spatial regularization(FSR)algorithm to speed up the spatially-regularized correlation trackers.We first present a low-cost spatial regularization function for the correlation filter based trackers to learn spatially-regularized filters.We then present a truncatedly iterative meta-learning(TIML)method to e?ectively and e ciently optimize the objective function through a dual LSTM meta-learner.Extensive experiments validate that the FSR method significantly speeds up 5 existing SR-based CF-based trackers and helps achieve better accuracy with near real-time performance.4.We propose a Siamese filtering based dynamic regularization method for beyond real-time deep trackers by considering the spatial-temporal changing of the target and its context.We first add two transformations to the Siamese network,i.e.target appearance variation and background suppression transformations.Then,we e ciently learn these two transformations by solving an objective function in the frequency domain.In addition,through multi-layer fusion and joint training,the tracking accuracy is further improved.Moreover,we analyze the relationship between Siamese filtering and correlation filter as well as Siamese network and find that Siamese filtering unifies the two frameworks and overcomes their main demerits while maintaining the merits.Experiments on benchmarks show that the Siamese filtering based trackers achieve similar precision to classification based deep trackers but run over 20 times faster.