Research On Small Target Recogintion And Statistical Technologies For Garden Scenarios

Small target recognition and tracking,as an important research direction,has been widely used in many scenes at present.The garbage problem in gardens often has a great impact on garden management and garden beauty.Based on garden scenes,this thesis studies the related technologies of small target detection and tracking,so as to realize the effective identification,tracking and statistics of garden garbage.The main research contents of this thesis are as follows:(1)The algorithms of small target recognition and tracking are studied.Firstly,the basic image preprocessing algorithm is introduced;Then the mainstream target recognition algorithm and target tracking algorithm are studied.(2)The scene driven yolov5 target recognition algorithm based on anchor box optimal clustering is studied: Firstly,the basic principle and network structure of YOLOv5 model are introduced.The optimal clustering method of anchor frame based on K-means is proposed to obtain the anchor frame size of yolov5 network model quickly and accurately.Subsequently,the scene driven parallel feature fusion pyramid structure is studied,and the scene driving coefficient is proposed.By calculating the scene driving coefficient,the middle feature layer is adaptively fused to solve the problem of feature extraction conflict when large and small targets exist at the same time.(3)The deepsort garbage statistics algorithm based on the motion state of UAV is studied: Firstly,the basic process of deepsort algorithm is studied,and the prediction frame acquisition method based on fuzzy evaluation index is proposed.Through the pre inspection of the picture,the accurate prediction of the prediction frame is realized.Subsequently,the prediction frame matching method based on UAV motion state is studied,and the SIFT feature matching fusion based on height change rate and Hungarian matching compensation based on speed change rate are proposed for the accurate matching of prediction frame and detection frame.Finally,the algorithm proposed in this thesis is tested and verified through experiments.The results show that the algorithm not only ensures the tracking statistical speed,but also improves the accuracy of tracking statistics.