| Vehicle distance detection is one of the key technologies in intelligent driving system.Distance detection technology based on monocular vision has the advantages of convenient installation and debugging of equipment,less consumption of computing resources and good real-time performance.It is widely used in intelligent driving research.In recent years,with the development of depth learning,the target distance detection network based on depth learning has begun to appear,but because of its simple structure and easy to lose more spatial information,the prediction accuracy of target depth needs to be improved.Therefore,it is of great theoretical and practical significance to study the monocular visual forward vehicle distance detection model based on depth learning.Based on the analysis of the existing distance detection technology based on vision,this thesis studies the depth learning model of forward vehicle distance detection,the accuracy of detection algorithm and the acceleration of detection model,and realizes the forward vehicle distance detection based on deep learning.The main work and contributions of the thesis are as follows:(1)This thesis introduces the DORN algorithm in the depth estimation algorithm to build the forward vehicle distance detection model D＿FVDD.based on the existing network which is easy to lose space information and affect the detection accuracy by converting the depth estimation problem into the classification problem,the high resolution depth map is obtained while simplifying the network structure,so as to retain the depth information around the target and reduce the parameter fluctuation during the network training.Improve the detection accuracy of forward vehicle distance detection network.(2)In order to improve the accuracy of forward vehicle distance detection network,this thesis improves the accuracy of forward vehicle distance detection by optimizing the key point fitting method and improving the loss function.Aiming at the fitting optimization problem of the key points of the target,the K-means clustering algorithm is introduced to configure the parameters effectively,so as to improve the fitting accuracy.To solve the problem of insufficient pertinence of loss function,by combining the Anchor-free idea,the center point of goal 3 D is taken as the real value corresponding to its depth value,so that the network can obtain the learning information of specific points in training,thus effectively improving the detection accuracy.(3)considering the practical application requirements of the forward vehicle distance detection model,it is necessary to compress and accelerate the model.Based on the analysis of the model optimization method,and the Tensor RT tools are used to accelerate the optimization of the model and greatly improve the detection speed.In order to verify the validity of the proposed distance detection model,a data acquisition platform is built based on the campus scene,and the error law of the prediction is analyzed,then the error correction model is established.The experimental results show that the modified model can significantly improve the accuracy of forward vehicle distance detection.Based on the present situation that there is no open target distance detection data set,this thesis constructs an extended data set for forward vehicle distance detection with the help of KITTI data set,builds a forward vehicle distance detection model,and optimizes the accuracy and speed of the model.Finally,it is verified by multiple comparison experiments.The experimental results show that the optimized forward vehicle distance detection model has high detection performance. |
