Research On UAV Target Tracking Algorithm Based On Improved TLD

As countries attach importance to UAV(Unmanned Aerial Vehicles)technology,the development of UAVs has entered an unprecedented rapid development track.At present,target tracking is a hot research direction in computer vision,and target tracking algorithms are constantly updated.The combination with UAV technology plays a major role in the promotion of UAV technology.Taking into account that in the process of target tracking,UAVs often encounter tracking targets being occluded,lighting changes,etc.Therefore,this paper is based on the TLD(Tracking-Learning-Detection)tracking algorithm framework and using UAV123 data set is the research object,and the long-term tracking of ground targets from the perspective of UAVs is carried out in-depth research.The main work is as follows:First of all,in view of the problem that the TLD algorithm cannot adapt well to the illumination changes and the in-plane rotation of the target under the UAV’s perspective,the paper proposes an improved TLD algorithm based on local binary features.The local binary features are merged in the nearest neighbor classifier of the detection module,so that the improved TLD algorithm based on local binary features improves the tracking success rate in the scene of illumination changes and target rotation in the plane by an average of 12.1% compared with the TLD algorithm.Secondly,in view of the problem that a large number of meaningless windows to be detected in the TLD algorithm reduces the real-time performance and accuracy of the algorithm,the paper proposes an improved TLD algorithm based on the adaptive detection area.The algorithm uses local search and Kalman filtering to predict the target position and screen the window to be detected.The experimental results show that the average frame rate of the improved TLD algorithm based on the adaptive detection area is increased by 12.67 fps,and the tracking success rate of the improved TLD algorithm based on the adaptive detection area in the occluded scene is compared with the improved TLD algorithm based on local binary features and KCF(Kernel Correlation Filter)algorithm improved 10.5% and 13.2% on average,respectively.Then,the threshold of the variance classifier of the TLD algorithm is a fixed value,which cannot reflect the change of the true variance,thereby reducing the efficiency of the detection module.The paper proposes an improved TLD algorithm based on automatically updating the threshold of the variance classifier.The algorithm uses the variance of the positive samples in the learning module to update the threshold of the variance classifier.The experimental results show that the tracking success rate of the improved TLD algorithm based on automatically updating the threshold of the variance classifier is 4.78% and 13.1% higher than the TLD algorithm and the TLD algorithm based on the adaptive detection area,respectively.Finally,in response to the problem that the disappearance of target feature points causes the output of the TLD algorithm tracking module to fail when the target is occluded from the UAV’s perspective,the paper proposes an improved TLD algorithm based on spatial and channel reliability discrimination related filtering.The algorithm uses discriminative correlation filtering algorithms with spatial and channel reliability to improve the tracking module of the TLD algorithm.The experimental results show that under the UAV123 data set,the algorithm effectively improves the anti-occlusion ability compared with the TLD algorithm.The improved TLD algorithm based on spatial and channel reliability discrimination related filtering has a tracking success rate compared with ASRCF algorithm(Adaptive Spatially-Regularized Correlation Filters)、spatial and channel reliability discrimination related filtering algorithm and the improved TLD algorithm based on automatically updating the threshold of the variance classifier improved by 1.2%、1.6% and 13% on average,respectively.