Research On Pose Estimation Of Front Vehicle Based On Computer Vision

The research of vehicle pose estimation in intelligent driving is helpful to the information acquisition and understanding of the surrounding scene,and it can also promote the development of the field of automatic driving perception.However,the vehicle pose information cannot be obtained directly from the image.Based on the analysis of this problem,this paper proposes a method of vehicle pose estimation based on computer vision.Combined with the method of deep learning,three-dimensional detection and tracking are carried out to determine the attitude and position information.The main contents of this paper are as follows:(1)The research builds a model for estimating the speed and distance of the vehicle ahead.The error source and magnitude of the relative distance of the vehicle are analyzed,the error compensation method is studied,and an inverse perspective transform ranging method based on dynamic compensation of the pitch angle is established to reduce the ranging error.The experimental results show that the model can accurately obtain the speed and distance of the vehicle in front in the moving scene in real time.(2)The 3D detection algorithm for forward vehicles based on computer vision is proposed.Based on Res Net-34 model for basic feature extraction,a 3D vehicle detection algorithm model for a multi-tasking sub-network is studied.The target coding is represented by 3D parameters,a candidate region generation network module is added to the model,and 2D and 3D information coding is integrated into the feature pyramid network,which can better detect the pose of small target vehicles.The experimental results show that the proposed monocular 3D vehicle detection algorithm finally predicts the 3D parameters through an end-to-end method,and returns the pose information of the vehicle,including the attitude angle,size,and world coordinates of the center point.The camera 3D vehicle detection algorithm has the same effect.(3)The 3D tracking algorithm for forward vehicles based on computer vision is proposed in this study,and images and spatial information are integrated into Markov decision algorithms.The 3D Kalman filter is used to predict the pose and state of the vehicle to generate similarity measures between all detection results and tracking results.The occlusion problem encountered during vehicle driving is solved with the help of data association method and depth ranking matching.Obtaining the newly matched trajectory can re-estimate the actual three-dimensional position of the vehicle in space.Experimental results show that the center point tracking error drift of this algorithm is small,and the 3D pose of the vehicle can be continuously obtained in real time.