Research On Vehicle Forward Collision Warning System Based On Monocular Vision

In recent years,due to the rapid increase in car ownership,car accidents have occurred more frequently,among which collision accidents have a strong correlation with human factors.At the same time,advanced driver assistance systems are constantly developing,and the collision warning system is one of the important branches of the system.Therefore,in order to remind the driver to avoid the occurrence of collision accidents,this paper proposes a vehicle forward collision warning system based on monocular vision.First,the object detection algorithm is utilized to detect the forward object of the video stream obtained by the camera and output the position information,and then the object tracking algorithm combines the detection result with the continuous image data to match the object to obtain the continuous information record of the object.Finally,the distance calculated by monocular vision ranging method are used to be analyzed so as to realize the early warning of forward vehicle collision,pedestrian and non-motor vehicle collision.The main research work of this article is divided into the following three parts:Firstly,a object detection model for traffic scenes suitable for data processing platforms with low hardware levels is constructed.The importance of the object detection model is to determine the accuracy and real-time performance of the entire warning system.This paper will use simplified existing convolutional neural network model to extract features of the traffic scene object,set an appropriate number and size of anchor frames according to the object situation of the training set,and detect traffic objects on two scales.Experiments show that this method makes the model lighter and speeds up training and detection under the working conditions of this paper.Then,a multi-object tracking algorithm based on Kalman filter is realized,and use the improved Hungarian algorithm to associate and match prediction frame of the Kalman filter with the detection frame.the Hungarian algorithm takes the intersection over union the location information as the cost function.Experiments show that the algorithm is fast and has good adaptability to object scale changes.Finally,a vehicle collision warning strategy,pedestrian and non-motor vehicle collision warning strategy based on monocular vision ranging are proposed.The camera calibration method and the monocular ranging method are utilized to get the distance information of the traffic object from the image.The object distance information matching achieved by the tracking algorithm can calculate the relative speed and time to collision of the traffic object.According to whether the movement state of the traffic object is stable,the warning strategy is divided into the forward vehicle collision warning strategy based on time to collision and the pedestrian and non-motor vehicle collision warning strategy based on the dangerous distance.Experimental shows that this warning algorithm has better accuracy on the urban road.