Object Detection And Tracking For Traffic Statistics In Ferry Scenarios

Target counting has broad application prospects in the field of intelligent security and is a typical application in the field of computer vision.The target detection and target tracking technology it contains is the most basic problem that computer vision needs to solve,and it is also a hot spot in current research.However,in actual monitoring scenes,changes in lighting and complex backgrounds,especially pedestrian occlusion and posture changes,will affect the quality of the final result.In recent years,with the wide application of deep learning,it can effectively improve the algorithm quality in terms of robustness and accuracy.Therefore,this paper is based on deep learning methods to study target detection and tracking.This paper takes pedestrian vehicles based on the uplink and downlink in the data set as the research goal,comprehensively considers the real-time and accuracy,and combines with the actual scenario,proposes an improved YOLOv3(You Only Look Once v3)target detection method,based on IOU(Intersection Over Union)Vehicle tracking method and improved Deep SORT(Simple Online And Realtime Tracking With A Deep Association Metric)pedestrian tracking algorithm,and finally get the number of pedestrians,assisting security surveillance video.The main work and innovations of this article are as follows:(1)Establish data set for pedestrian and vehicle detection and trackingThe data collection in this paper comes from two parts,one is to collect data on real passenger ferry scenes,and the other is to collect data on real ferry scenes.The data set contains two categories of vehicles and pedestrians.The data set made by marking the head and shoulders of pedestrians includes 8874 images,of which 7986 training sets and 888 test sets.(2)Target detection algorithm based on improved YOLOv3In this paper,an improved YOLOv3 detection algorithm is proposed.In view of the relatively low accuracy of the original YOLOv3 algorithm in the detection,in the original YOLOv3 loss function,CIOU Loss is used instead of the mean square error loss function as the boundary box regression function,and the Focal Loss function is incorporated into the boundary box confidence loss function.To solve the problem of slow detection of YOLOv3 algorithm,Mobile Net network is used instead of Dark Net-53 network.The experimental results show that this algorithm has a gooddetection effect in this data set,and has real-time performance.(3)Multi-Object tracking and counting based on IOU tracker and improved Deep SORTThe size of pedestrians and vehicles differs greatly in the image,and the area of pedestrians' borders is much smaller than that of vehicles,so two different counting methods are used.For vehicle counting,IOU-tracker is used for tracking and counting.For pedestrian targets,this paper proposes an improved Deep SORT multi-target tracking algorithm for counting.The improved Deep SORT tracking algorithm uses Res Net101 network and Hard Triplet Loss loss function to replace the original network.Experiments show that the improved algorithm has better effect on target tracking.The experimental results show that the detection algorithm proposed in this paper can achieve real-time effect in the self-made test set of this paper,and achieve 93.27%accuracy.The counting system that combines detection and tracking can achieve good results in both vehicle and pedestrian categories.