Research On Key Techniques Of Human Target Detection And Tracking

Target Detection and Tracking is a hot and key point for computer vision.The research results of this direction are used in many fields such as video surveillance,virtual reality,human-computer interaction,behavior understanding and automatic driving.However,since the actual scene is complex and variable in practical application,detecting and tracking the target needs to meet the challenges such as occlusion,light source change,attitude change and scale change brought by the camera.In recent years,with the excellent performance of the deep learning,it has brought a new solution to these challenges in the field of computer vision.This paper proposes a method combining deep learning to solve the problem of target tracking.By referring to the idea of TID algorithm,this algorithm uses DetectionRecognition-Tracking to achieve the task of target tracking.The neural network has good detection recognition rate and robustness in detecting partial occlusion,illumination change,camera pose change and target scale change.In this paper,convolution neural network can be used to detect targets better,which provides good preliminary conditions for tracking targets.However,the neural network recognition results are not 100% correct,so it is necessary to remove the false results during the tracking process.In the detectionbased tracking framework,this paper extracts the HOG feature of the target in two consecutive frames in the single target tracking algorithm.Then,the multi-target tracking algorithm is studied,and the better tracking effect is obtained by combining the different features of the image.The main research contents and innovations of this paper are as follows:1)Based on the target-tracking framework of detection-recognition-tracking,a single target tracking algorithm based on YOLO is proposed.Instead of relying on the previous linear motion model or quadratic motion model,the target is detected and identified by neural network,and then a tracking algorithm is designed to track the target.2)Based on the single target tracking algorithm,this paper attempts to fuse the different features of the detected image region: target spatial location,color features and texture information.Experiments show that the robustness of multi-target tracking algorithm is further improved by different features.At the same time,the target box regression module based on convolution neural network can realize the adaptive adjustment of the tracking frame.Compared with other methods,the experimental parameters in this paper are simple,and good results are obtained on different data sets.