Research On Visual Object Tracking Technology Based On Correlation Filtering

With the development of UAV technology,the phenomenon of "heifei" has been prohibited everywhere,so the research of control technology for UAV has a broad application prospect,and among the various control technologies,visual tracking technology plays a key role.This paper investigates the visual tracking based on correlation filtering proposes a background-detection tracking algorithm with multi-feature adaptive fusion and a UAV small objects tracking algorithm based on YOLO,then develops a UAV visual tracking platform.The main research of this paper is as follows:1.A background-detection tracking algorithm with multi-feature adaptive fusion.Firstly,aiming at the problem that a single feature cannot fully express the target feature,an adaptive fusion method of color feature,HOG feature and edge detection feature is proposed.Then,in view of the problem of filtering the background in the traditional method,a background detection framework is introduced,and the background information around the target is added to the filtering.Finally,an improved ADMM(Alternating Direction Method of Multipliers)optimization algorithm is proposed to participate in the operation.The proposed algorithm enhances the feature representation of targets in dynamic environments,improves the robustness to background changes as well as occlusions,and reduces the number of optimization iterations.Simulation on the OTB2013 dataset verifies that the accuracy of the proposed algorithm is improved by 4.1% and the success rate is improved by 3.9% compared to the Siam FC algorithm.2.A UAV small target tracking algorithm based on YOLO is proposed.First,for the target prior problem,the algorithm updates the initial prediction frame through the K-means clustering analysis method to output more robust prediction results in this chapter.Secondly,in view of the small number of small target samples,a small target sample enhancement strategy is adopted to increase the quantity and quality of training samples and increase the number of anchor points that can be matched by small targets.Thirdly,for the problem of unbalanced sample discrimination,the EIOU＿Loss(Efficient IOU Loss,efficient IOU loss)loss function and focusing function are combined to balance the sample classification and reduce the loss value shock caused by low-quality samples.Finally,for the problem that overlapping targets depend on the threshold setting,an improved weighted NMS method(Non-Maximum Suppression,non-maximum suppression)is used to optimize the confidence weighting and improve the detection accuracy of overlapping targets.Using the UAV dataset for testing and comparing with the YOLO model,the method proposed improves the average detection accuracy of UAV small targets by 23% in this chapter.3.Based on the research in visual tracking algorithms,a vision tracking platform for small UAV objects is developed.The platform is divided into three modules: model training,video transmission and processing and gimbal control,which can realize the function of automatic tracking of UAVs that appear in the coverage area.In the model training module,the training and tracking data set is built with more than 4300 images collected and labeled to train the tracking model.In the transmission and processing module,the main body is the small target tracking algorithm proposed in this paper,while the design of dual-threaded and mutually exclusive thread locks for simultaneous acquisition and analysis,and the design of chain queue buffer to reduce resource overflow and speed waste caused by thread waiting.In the control module,the PTZ is controlled by an improved PID algorithm.Test and analysis in practical application scenarios have proven the tracking performance and reliability of the platform.