Research On Real-time Detection And Tracking Of Aerial Traffic Flow Based On Deep Learning

With the rapid increase of computing power of the machine,deep l earning has achieved a huge breakthrough in computer vision,and provides solutions to many computer vision related applications,such as traffic flow detection and tracking.Most of the traditional vehicle detection approaches are based on the intersection of different cameras,it is inefficient.With low-altitude aerial vehicles advanced,it is easy to acquire low-altitude aerial images,but still requires a large annotation cost.Currently,traffic detection and tracking methods based on low-altitude aerial data can not balance accuracy with real-time performance.In this paper,we propose a novel detecting and tracking method for real-time traffic flow based on deep learning.Aiming at the problem that the annotated samples of aerial images are very few,this paper proposes weak-supervised learning based with YOLO.After training a vehicle classifier through a small number of annotated images,candidate boxes are selected by a method called selective search,and assigned by vehicle classifier.Coarse-level annotated images are used to initialize the convolutional network layer of YOLO,thereby enhancing the model's detection performance.In the experiments,the modified method proposed in this paper can train the model by using a small number of labeled images combined with a large number of unlabeled images,and achieves the same model performance learnt on a large number of labeled images.For real-time traffic flow detection on aerial images,this paper further revises YOLO by employing the anchor mechanism of Faster R-CNN and the idea of full convolutional network.Utilize the idea of full convolutional network to improve the full connection layer of YOLO and improve the detection speed of YOLO.Through the anchor mechanism to improve YOLO candidate box extraction layer,to learn the prior knowledge of the target box,so as to learn more powerful features to improve the model detection accuracy.To improve the robustness of the detecting model,the multi-resolution and multi-view training methods are adopted to improve the YOLO training network.The experimental results have demonstrated that the proposed method can achieve the real-time detection and 88.8% m AP in complex scenes.For real-time traffic flow tracking issue,this paper adopts matching and tracking method.By designing a vehicle matching model,the detection results of two frames before and after the video are taken as the input of the matching model,and matched according to the size of the detection frame,the color histogram,the position information,and achieve the real-time traffic flow tracking.