Research On Detection And Tracking Of Humans And Vehicles

With the development of society,the impact of targets detection and tracking technology is growing in our production and life.Since humans and vehicles are the most important constituent objects of modern society,detecing and tracking humans and vehicles quickly and accurately is of great significance for building a smart city and a safe society.We conducted the research in the context of humans and vehicles detection and tracking:To solve the problem that one-stage detection models detect humans and vehicles with low accuracy,we propose a new detection model named UYOLOv3 based on YOLOv3.First,we propose a new bounding box regression loss named WIOU-Loss in UYOLOv3.WIOU-Loss is based on IOU-Loss and assigns different weights to the overlapping areas according to their relative positions.Secondly,to solve the problem of unbalanced foreground and background in YOLOv3,we introduce Focal loss which can adjusts the weight of samples during training in UYOLOv3.Finally,we introduce soft-nms in UYOLOv3 to perform better than YOLOv3 on mutual occlusion of the same kind of targets.When dealing with the nonmaximum value,soft-nms adoptes the strategy of suppressing its confidence rather than directly removing it.To detect humans and vehicles quickly and accurately in embedded devices and mobile devices,we propose a lightweight model named Mobile Net V3-YOLOv3.First,we replace Darknet53 with Mobile Net V3 as the backbone to reduce the number of parameters and computation in Mobile Net V3-YOLOv3.Secondly,we introduce the knowledge distillation framework.UYOLOv3 is used as the teacher model and Mobile Net V3-YOLOv3 is used as the student model.In order to improve the accuracy of the student model,the teacher model guides the student model by constructing a specific loss function during training.Finally,we adjust the unreasonable parts in the framework of knowledge distillation to further improve the detection accuracy of Mobile Net V3-YOLOv3.To avoid model drift in correlation filtering algorithms,we propose a new tracking model named A-Staple which is based on Staple.Comparing with the traditional correlation filtering method using fixed learning rate to update model,we propose a new learning rate update strategy based on high confidence criterion in A-Staple.When the target is under occlusion or analog interference,our strategy can slow down the model update rate.