Research And Implementation Of Improved YOLO_V3+deepsort Muti-target Tacking System

With the development of artificial intelligence and computer vision,the field of target tracking in video has developed rapidly,especially human body multi-target tracking.Human body multi-target tracking has been widely used in traffic management,autonomous driving and robot navigation.Although researchers at home and abroad have achieved good results in the field of target tracking,the application and development of target tracking is restricted when the field of vision is wide,there are many targets of interest,or there are serious occlusions between targets.Real-time target detection and tracking algorithms based on deep learning have gradually become a research focus in human target tracking.This system adopts a combination of detection and tracking algorithms for human multiple tracking.The main research contents are as follows:(1)This thesis first introduce the principle of YOLO_V3 target detection algorithm in detail.YOLO_V3 is a real-time end-to-end deep learning object detection algorithm.YOLO_V3 is improved on the basis of YOLO and YOLO_V2.Firstly use K-means clustering to optimize the prior frame of the detection model,then use the new network structure Darknet-53 as the backbone for feature extraction,secondly use multi-scale features for object detection,and use multiple independent logical classifiers instead Softmax function classification.YOLO_V3 can achieve good detection for both large and small targets,making the detection effect better.(2)In this thesis,the improved YOLO_V3 algorithm is used for target detection,and the Gaussian model is used to model the network output,which can output the reliability of each prediction box.Gaussian YOLO_V3 and loss function reconstruction can improve the detection accuracy by reducing the influence of noise data during training and predicting the positioning uncertainty.In addition,using this predicted positioning uncertainty can further improve the detection accuracy.(3)This thesis uses Deepsort algorithm for target tracking.Deepsort is an improvement based on Sort target.Sort is a practical method focusing on simple andeffective algorithms for multi-object tracking.However,the Sort algorithm cannot handle occlusion problems well,because the association metric used is only association based on motion information.It is accurate only if the state estimates are relatively stable.This thesis integrates appearance information to improve the performance of Sort.By replacing the associated metric with a more comprehensive metric that combines motion and appearance information,and applying a trained Cosine deep network to distinguish large-scale people from re-identifying pedestrians on the dataset.Because of this extension,it can handle target tracking in occlusion situations,thereby effectively reducing the number of identity switches.(4)Finally,this thesis designs and implements a multi-target tracking system based on improved YOLO_V3 and Deepsort according to the actual application requirements.Experiments show that the system is efficient and stable and can meet the actual application requirements.