Research On Multi-object Tracking Based On Visual Detection

With the development of computer and artificial intelligence technology,camera multi-object tracking technology and track modeling based on tracking results play an important role in intelligent monitoring analysis,which has received more and more attention in the field of computer vision.In recent years,the number of camera monitoring terminals has increased rapidly,resulting in massive monitoring video data.However,the current monitoring system is still based on artificial monitoring and retrieval analysis,and there are many problems such as missing alarm,false alarm and low monitoring efficiency.Therefore,the intelligent monitoring system with independent monitoring and analysis ability has become a research hotspot.In view of the application status and requirements above,a multi-object tracking system based on camera visual detection is proposed in this thesis,which can achieve stable and reliable tracking of multi-pedestrian target in the monitoring scene.Based on the multi-object tracking framework of vision detection,this thesis first improves the object detection technology.The main network of the current object detection algorithm is migrated directly from the traditional image classification network.The problem of single scale and spatial resolution loss of feature map exists.In view of these problems,based on the difference analysis of image classification task and object detection task,this thesis improves and optimizes the backbone network structure of object detection task.The sampling optimization is carried out in the highlevel feature map,the sampling scale remains unchanged,the generalization of the feature map is enhanced,the feature results are enriched,and more robust detection results are obtained.Experimental results show that the proposed algorithm can obtain higher detection accuracy.In the data association stage of multi-object tracking,a multi-object tracking method based on fusion feature measurement is proposed in this thesis,aiming at the common problem of object intersection and object occlusion in the actual scene.Firstly,the apparent features of the target are introduced in the distance measurement,and a robust lightweight feature extraction network is constructed to extract the features of the target and obtain the more discriminative feature representation and distance measurement results.On this basis,Kalman filter is used for state prediction and Hungarian algorithm for matching to achieve multi-object tracking.Experimental results verify the effectiveness of the proposed algorithm.On the basis of the above work,a multi-pedestrian tracking system for monitoring scenarios is designed in this thesis.In this system,referring to the structure frame of the denoising autoencoder,a convolutional autoencoder network is proposed to preprocess the scene image,effectively remove the background noise and enhance foreground target,and improve the performance of the multi-pedestrian tracking system.Finally,the effectiveness of the system is verified by experiments under actual monitoring scenarios.