Research On Multi-Target Tracking Algorithm Based On Deep Learning

As a basic and important task in the field of computer vision,multi-target tracking is widely used in video surveillance systems,autonomous driving and intelligent interaction.In the process of multi-target tracking,there will be problems such as deformation of the targets,occlusion of the targets,appearance of new targets and disappearance of old targets.So it is extremely important to improve the accuracy of the algorithm in practical applications and the real-time performance of rapid landing.With the advent of the era of deep learning,the performance of target detection algorithms based on deep learning has been rapidly improved.Therefore,this paper adopts the detection-based multi-target tracking algorithm framework for research.The main contents are as follows:(1)Aiming at the problem of poor detection effect and poor real-time performance when small target detection and pedestrian occlusion occur in the target detection task,this paper proposes a real-time pedestrian detection algorithm based on the fusion attention mechanism.Firstly,the lightweight network Mobilenet V3-large is used to replace the CSPDarknet53 of the original YOLOv4 algorithm as the backbone feature extraction network to reduce model complexity and number of parameters.Then,the SE channel attention module in the original Mobilenet V3-large network inverted residual structure is replaced by a fusion attention module based on channel dimension and spatial dimension,which is used to enhance the network’s ability to represent local key information in dual dimensions and to use inhibition of information.The improved YOLOv4-Mobilenet V3-FA algorithm greatly improves the real-time performance of pedestrian detection tasks while ensuring that the accuracy is not much different.Finally,it is tested on the Crowd Human dataset.The accuracy of the improved algorithm is increased by 1.6%,and the detection speed is increased by 9.5 frames per second,which meets the real-time requirements for target detection.(2)Aiming at the inaccuracy of using the intersection ratio as an association measure and the frequent exchange of identities in the process of multi-target tracking and matching,this paper proposes an improved Deep-SORT pedestrian tracking algorithm.Firstly,the KM algorithm that can solve the optimal matching problem of the weighted bipartite graph is used to replace the Hungarian algorithm used in the original Deep-SORT algorithm for the data association process,so that the matching target with the largest weight is preferentially selected for matching instead of each the matching probability of matching targets is considered the same.Secondly,the complete intersection ratio is used as a measure of the distance correlation between the target detection frame and the trajectory tracking frame.The complete intersection ratio not only considers the overlap area between the target detection frame and the trajectory tracking frame,but also considers the bounding box.The aspect ratio of the scale information and the center point distance information between the two bounding boxes can better reflect the similarity between the two,and also have a better judgment on the trajectory matching process.Finally,it is tested on the MOT16 dataset,the improved Deep-SORT algorithm multi-target tracking accuracy is increased by 5.54%,and the number of identity exchanges is reduced by 36.88%,which further verifies the effectiveness of the algorithm in this paper.