Research On Uncertain Motion Tracking Method Based On Deep Learning And MCMC Sampling

Visual object tracking is one of the important research subjects in the field of computer vision at present.It has been widely used in the fields of medical,military,and industrial and etc.Despite the remarkable research achievements in object tracking technology in recent decades,there are still many unpredictable factors that may lead to tracking failure in real-world scenarios,such as fast motion,background occlusion,and motion blur.The main reason for these factors is the drastic position change of the target between two frames due to its own motion or camera condition change.Therefore,the problem of adapting to target motion uncertainty becomes a key issue for achieving long-term tracking.However,the motion models built by conventional algorithms based on the assumption of smooth motion of the target are usually insufficient to achieve stable tracking.In order to solve the uncertain motion problem of the target,this thesis uses Markov chains Monte Carlo(MCMC)method with Deep Learning technique to improve the tracking performance in complex scenes.The main research work and innovations are described as follows.(1)An uncertain motion object tracking method based on Extend Wang-Landau Monte Carlo and Deep Correlation Filters(EWLCF-DP)is proposed.First,the Multi-Scope Marginal Likelihood(MSML)strategy is introduced into the Wang-Landau Monte Carlo(WLMC)algorithm to improve the sample acceptance rate in the promising region.Second,a more reliable density-of-states(DOS)distribution is used to mark promising regions,where the iterative search process is simplified by the correlation filtering operation to improve the tracker's efficiency.Target localization is achieved by the maximum response in the promising region.Finally,a unified tracking framework is designed to enable correlation filters and WLMC with MSML strategy to exploit and complement each other to cope with uncertain motion tracking.(2)A Target-objectness Proposals(TOP)based uncertain motion object tracking method is proposed.Firstly,the MCMC approach is derived as a regional proposal mechanism to obtain target-level proposals from a new perspective.It integrates learning features into the feature space sampling strategy for global object candidate boxes generator.Secondly,an objectness label is adopted to constrain the feature space sampling and improve the probability that the candidate frames contain targets.In addition,a TOP candidate evaluation method based on object score and likelihood score is also used to filter the target-level suggestions to obtain a small number of high-quality TOP candidates.In the end,a unified tracking framework is designed to enable sampling and regression strategies to exploit and complement each other to cope with uncertain motion tracking.And a target verification module is included to determine whether the target uncertainty movement occurs.(3)For the target uncertain motion problem,the proposed algorithm is extensively experimented on uncertain motion sequences containing ten videos and four target tracking test datasets,OTB-2013,OTB-2015(Object Tracking Benchmark),TC-128(Temple Color),and UAV-123(Unmanned Aerial Vehicles),and the results demonstrate that the proposed EWLCFDP with TOP tracker is highly competitive with advanced tracking methods.