Research And Application Of Target Detection And Tracking Technology Based On Deep Learning

Target detection and tracking technology are the two main research directions in the field of computer vision and have a wide range of applications.However,the small size and occlusion of the target seriously affect the performance of the target detection and tracking algorithm.In this paper,aiming at the problem of poor detection and tracking results caused by small target and target occlusion,we propose a target detection algorithm based on multi-granularity feature fusion improved by YOLO,and further propose a target tracking method based on LSTM,which can improve the detection and tracking performance of small target and occlusion.The main contents of this paper are as follows:(1)This paper analyzes the current situation of target detection and tracking at home and abroad,as well as the specific problems of target too small and target occluded in the process of target detection and tracking,and introduces the basic theory of deep learning and the target detection and tracking algorithm based on deep learning,which provides the theoretical support for the later algorithm improvement.(2)A target detection algorithm based on multi-granularity feature fusion improved by YOLO is proposed.Through the improvement of multi-granularity feature fusion in the network structure of YOLO,the feature vector extracted by the model in the final detection takes more underlying features into account,which solves the problem of feature loss caused by the small detected object and the operation of multi-layer convolution and down sampling.(3)A target tracking algorithm based on LSTM is proposed.Aiming at the problem of occlusion and slow tracking speed in target tracking,this paper introduces the interest region determination module of K-Neighborhood search based on the improved multi-granularity feature fusion algorithm of YOLO,quickly determines the interest region of target feature extraction,and uses the expression ability of long-term memory neural network to temporal spatial information and the "gate" mechanism to track the target.The tracking feature of the target is modeled,and the moving feature and semantic feature of the target are retained selectively,and the position of the next frame of the target is predicted,which reduces the dependence of the tracking algorithm on the feature extractor and improves the tracking performance when the target is occluded.(4)The algorithm of target tracking for specific floating objects on the water surface is verified by the unmanned ship.Based on the unmanned vessel,the actual tracking experiment scene of a specific floating object with small target,occlusion and other challenges is designed,and the target tracking algorithm proposed in this paper is verified by experiments.To sum up,this paper proposes the target detection algorithm based on the improved multi-granularity feature fusion of YOLO and the target tracking algorithm based on LSTM,and evaluates the performance of the proposed target detection and tracking algorithm in the public data set.Finally,the tracking experiment is carried out on a specific surface floating object on the unmanned vessel.The experimental results show that the target tracking algorithm proposed in this paper has good performance and stable effect in tracking success rate,accuracy rate and tracking speed,and has good application value.