Research On Video Image Processing Algorithms And Applications Based On Deep Learning In Traffic Scene

With the rapid development of road traffic,the importance of video image processing is increasing.Object detection is an extremely significant step in image processing,and the accuracy and speed of it are both of great importance to the whole system.Traditional object detection methods usually rely on experience to extract features manually,having poor performance and robustness.In recent years,object detection based on deep neural network has amazing performance in both accuracy and speed due to booming deep learning.This paper mainly focuses on applying object detection based on deep learning to video image processing of traffic scene.The main work of this paper is as follows:(1)The application of object detection based on deep learning in traffic video monitoring system is studied.The network of YOLOv3 is chosen to be used in this paper,and the method of getting parameters for YOLOv3 network to detect specific objects by self-training is given.Detecting specific objects based on YOLOv3 network and self-training parameters is realized.Also,the network of YOLOv3 is improved.Preprocessing layer is applied to night images for image enhancement,improving the detection performance of night images.(2)The judgement of illegal behaviors in actual monitoring scene is studied.Every object in each frame is detected by object detection algorithm based on YOLOv3.And on this basis,the method of object tracking using maximum object area matching is given.Also,least squares method is used to fit boundaries of no-parking areas,road markings and road isolation facilities in the monitoring area.The criteria for judging typical illegal behaviors in traffic scene is defined by comparing object information and road information.(3)The implement of the intelligent traffic violation system based on deep learning is studied.The hardware framework of the system,including data acquisition module,server computing module and client human-computer interaction module,is established.And the software framework of the system,including data stream acquisition,parameters configuration,object detection,object tracking,violation determination,evidence extraction and evidence upload modules,is also established.The hardware construction and software programming and testing of the integrated and practical system are completed.The test of capturing typical violations and obtaining evidence is conducted and the result is good.