Research On UAV Object Tracking Based On Spatial-temporal Regularized Correlation Filters

Unmanned Aerial Vehicle(UAV)is widely used in military reconnaissance,disaster monitoring,intelligent transportation,and other fields.Visual object tracking is the basis for the widespread application of UAV,however,it is extremely challenging to achieve accurate and robust UAV object tracking due to the problems of motion blur,similar objects around,occlusion,and multi-angle and multi-directional scale changes.In addition,the computing resources of UAV are limited,which puts forward higher demands on the real-time performance of tracking algorithm.Therefore,the study of accurate and robust UAV object tracking algorithms under the condition of ensuring real-time performance has certain theoretical and application value.Correlation filters have been widely studied and applied in the field of object tracking,but they suffer from boundary effects and model degradation.The spatial-temporal regularized correlation filters have improved the robustness of tracking method by addressing these problems.Auto Track is an improved spatial-temporal regularized correlation filter that performs better by using adaptive regularization terms and learning rates.However,deep research and analysis have revealed three main problems with these algorithms.Firstly,they ignore the high uncertainty of response maps when the object is occluded or similar object around.Secondly,inappropriate update strategies cause model degradation even tracking failure.Thirdly,the lack of scale estimation module cannot adapt to multi-angle and multi-directional scale changes in UAV scenarios.Therefore,this thesis will solve above three problems.The main goal of this thesis is to propose and design an algorithm based on Auto Track that can achieve tracking accuracy and robustness in complex UAV scenarios while ensuring real-time performance.This thesis mainly conducts research work in the following two aspects:The probabilistic adaptive spatial-temporal regularized correlation filters(PASTRCF)is proposed to solve the problem of high uncertainty in response maps and inappropriate model update strategies.Firstly,the classification training module in the object tracking is improved by using a probabilistic model to construct the reliability information of the response map.This probabilistic model is used to design a new objective function,and the highly certain response map is utilized to adaptively adjust the regularization coefficients.Secondly,the model update module is improved by introducing an adaptive threshold mechanism to assist model updating,which effectively alleviates the problem of model degradation.The adaptive threshold mechanism was analyzed by visualizing the tracking results and the changes of threshold function.Then,the experimental evaluation and comparative analysis of PASTRCF algorithm were carried out on the UAVDT and DTB70 datasets.The results show that PASTRCF has good precision and success rates,and can achieve stable tracking in various UAV scenarios.Scale adaptive spatial-temporal regularized correlation filters(SACF)is proposed to solve the problem of multi-angle and multi-directional scale changes in UAV scenarios.Firstly,a Kalman position estimation module is introduced to predict the object trajectory and obtain the optimal position estimation.Secondly,the scale filter is trained using multi-scale information and rotation information to obtain a rough scale estimation.Thirdly,the scale-adaptive module is used to extract candidate regions and locate the local object region,and scale factors are constructed according to the localization results,so as to adapt the unique scale changes in the UAV scene.Then,the experimental evaluation and comparative analysis of SACF were carried out on UAVDT and DTB70 datasets.The results show that SACF has achieved advanced effects in terms of accuracy and success rate,and can adapt to multi-scale and multi-directional scale changes in various UAV scenarios,meeting the requirements of UAV scenarios.This thesis’ s work has achieved the established goals.However,the experimental results show that the algorithm our work proposed is sensitive to camera shake and has limited ability to express manual features.Therefore,the next step will focus on these two problems to further improve the robustness.