A Research On Target Tracking Based On Depth Learning

With the rapid development of information age,video data presents a state of mass growth.This puts forward new requirements for the ability to extract useful information from video,and also puts forward new demands for the content of video information.One of the needs of an important application background is how to get the trajectory of the target in the video,such as a person in a match or a car in a traffic environment,which is of great help to further analysis and statistics.In recent years,with the powerful ability,deep learning has been outstanding in many areas of computer vision.In order to build a real-time target tracking system using deep learning,the following works are done in this paper.(1)This paper explores the combination of Focal Loss and SSD,modifying the model structure with K-means,and the effect of the large data set pre-training model on the detection results.Tested on the actual data set,the improved model has a 3.4%increase in mAP index compared to the original model.(2)In this paper,a multi-target tracking model based on the integration of deep learning features and traditional artificial features is studied.The influence of target detection on multi-target tracking is analyzed,and the state vector of DeepSort model is improved.On the basis of ensuring the prediction speed of the model,a good result is obtained,which increases 4.4%on MOTA and greatly reduces IDs.(3)In this paper,the path registration problem is defined for the post-processing problem of the output of the multi-target tracking model.A trajectory tracking model based on pedestrian recognition is proposed,and experiments are carried out in real data.Based on the multi-target tracking model,a 0.2%stability improvement is achieved.(4)In this paper,a target tracking system based on target detection,multi-target tracking and path registration is designed and implemented.