Driving Scene Perception And Automobile Body Automatic Modeling Based On Deep Learning

With the rapid development of artificial intelligence technologies such as machine learning in the context of big data,intelligence has become one of the important characteristics and trends of automobiles in the new era.Environmental perception based on visual sensors and methods which relate to deep learning have been extensively and deeply researched and applied in the automotive field.On the one hand,because binocular vision has both the recognition ability of monocular vision and the precise ranging ability similar to lidar point clouds,it has been extensively studied.This article will study the perception of complex road information based on vision to improve driving safety and ride comfort with deep learning method.On the other hand,research on the modeling method of car body model based on monocular vision can realize the automatic modeling of complex curve and surface based on image with deep learning,in order to shorten the automobile development cycle.In this paper,we have the following contributions:For the information perception problems of road bumps and non-standard obstacles,a three-branch road information semantic segmentation deep network based on binocular vision is proposed.The network introduces a spatial attention mechanism and a channel attention mechanism,which can simultaneously receive image information and binocular visual depth information,and complete the perception and segmentation of road information by distinguishing the differences between the two and fusing effective features.This paper collects,organizes,and constructs a large-scale binocular vision data set in real driving scenarios.On this basis,numerical experiments such as performance verification and algorithm comparison are carried out.The results show the effectiveness and effectiveness of the perception and segmentation algorithm proposed in this paper.For the problem of real-time segmentation and tracking of non-standard small obstacles on the road,a semantic segmentation and tracking method based on monocular vision and deep learning network is proposed.This paper proposes a lightweight VGG+FPN semantic segmentation network to complete the real-time and accurate segmentation of small obstacles on the road,and uses the Center Track algorithm to track the small obstacles obtained by the segmentation.We created a dataset of small obstacles on the road in real driving scenarios.Numerical comparison experiments show that the method in this paper can realize the integration of sensing,segmentation and tracking of small road obstacles,and the frame rate can reach 50 FPS,which makes the detection and the segmentation of small road obstacles reach the application level.An automatic modeling method of single-view car body 3D curve based on deep regression network is proposed.The cubic Bézier curve is used to define the three-dimensional curve feature template of the car body,and the multi-view image of the car and the corresponding feature wire frame data set are created based on the three-dimensional mesh model library.The cubic parameter curve in the characteristic wireframe data set is uniformly sampled with equal arc length,and the car curve characteristic template is converted into a low-dimensional shape coefficient representation by the PCA method,and used as the corresponding multi-view image annotation information.Using Res Net101 as the regression network,using the mean square error to supervise the regression of the shape coefficients,the end-to-end automatic reconstruction from a single view to a three-dimensional curve model of the car body is realized.Numerical experiments show that the method in this paper can effectively realize the automatic modeling of auto body with arbitrary single-view input.