Improved TLD Human Target Tracking Method Fusing Direction Prediction And Corner Point Detection Algorithm

With the promotion of intelligent applications,intelligent video surveillance has become a new focus in the computer field.By applying image processing related methods to analyze the video captured by the monitoring equipments,it is possible to identify,locate,and track the target until the target is specified.The core is the target tracking technology.This paper focuses on human target tracking.The TLD(tracking-learningdetection)human target tracking algorithm is used as a framework to deeply study and analyze the long-term tracking algorithm of single target.The TLD algorithm only needs less a priori information such as the initial position of the target to achieve long-term online tracking of the target.The target tracking system based on the TLD algorithm consists of four parts: tracking module,detection module,integration module and learning module.Integrating the calculation results of the tracking module and the detection module,determining whether there is a target,a current state of the target in the current frame,and determining whether to train the current frame results.if necessary,learning new samples obtained by the detection module and the tracking module,updates the series parameters of the detection module and tracking module,and improves the overall target tracking ability of the system.The TLD algorithm effectively solves the problem of tracking effect caused by target deformation and occlusion during long-term tracking.However,this method has the problem of large calculation speed and low system running speed.In this regard,based on the TLD framework,this paper implements the improvement of the target tracking algorithm from two aspects.Improve the greedy detection method adopted by the detection module of the TLD algorithm,When the tracking module is running normally,The method of predicting the moving direction of the target by the Markov model direction predictor dynamically adjusts the area to be detected,and performs target detection.Reduce the scope of detection,fundamentally reduce the time cost of testing.The uniform sampling method adopted by the TLD algorithm tracking module is improved,and the SUSAN corner point of the block detection is used as the sampling point of the optical flow method for target tracking.Reduce the number of sampling points of the tracking module and improve operational efficiency.Contrast experiments with the video library provided by the original author of the TLD and the otcbvs infrared image library.The results show that the improved algorithm has improved the tracking accuracy rate by 16.58%compared with the original algorithm,and the overall system operating speed has increased by 84.41%.