Vehicle Detection And Ranging Based On Convolutional Neural Networks

With the development of artificial intelligence technology,smart cars have become an important direction of a new round of technological change.As an important part of smart cars,vehicle detection and ranging systems are of great significance for improving the safety and reliability of smart cars.Vehicle detection is the premise and basis of vehicle ranging.This paper proposes a vehicle detection algorithm based on enhanced YOLO v3 for complex road scenes.Vehicle ranging is the upper application of vehicle detection.This paper proposes a vehicle ranging algorithm based on edge width regression and Vehicle ranging algorithm based on image depth estimation.In this paper,the traditional vehicle detection method has high false detection rate and poor robustness in complex road scenes.The powerful feature learning and expression ability of Convolutional Neural Networks(CNN)improves the detection accuracy.Considering the accuracy and real-time,this paper proposes an enhanced YOLO v3 algorithm based on the general object detection algorithm YOLO v3 combined with the scene of vehicle detection.The number of optimal anchor frames is calculated by clustering,and the pre-training is performed by using the vehicle classification data set,and finally training the whole model with multi-scale.We achieved 89%AP(Average Precision)accuracy in complex road scenarios and 6%AP accuracy over YOLO v3 algorithm.In view of the shortcomings of traditional camera ranging,such as manual camera calibration,road modeling and pitch angle measurement,this paper proposes a vehicle ranging algorithm based on edge width regression.Based on the regression analysis theory,a regression equation for vehicle edge width and distance is established.The distance is calculated according to the regression equation during distance measurement,and no additional manual measurement is needed.Furthermore,in order to realize a more robust and universal ranging system,this paper innovatively proposes a vehicle ranging algorithm based on image depth estimation,using binocular images to train image depth estimation network unsupervised,and the network outputs a dense depth estimation image.The K-means clustering algorithm is used to calculate the vehicle distance in the vehicle detection boxes of the image.The whole process does not require any manual participation,and can be extended to the distance measurement of pedestrians,bicycles and the like.Based on the above algorithm,this paper implements a vehicle ranging system based on edge width regression.The experimental measurement has an error of less than 7%within 100 meters and can achieve real-time ranging of 30FPS(Frames Per Second).In addition,this paper proposes a new image depth estimation network,which achieves the accuracy of the latest algorithm on the KITTI dataset,with a relative error of 10%within 50 meters,which is better than the latest algorithm;and based on the proposed depth estimation network.The vehicle ranging system has a ranging error of less than 15%within 40 meters.