Research On Key Technologies Of Computer Vision For Autonomous Driving

With the rapid development of artificial intelligence,autonomous driving technology has received a high degree of attention from academia and industry.Road element detection based on computer vision is a key technology for unmanned vehicles to perceive the road environment and make control decisions.However,complex road scenes and various environmental interference factors will seriously reduce the performance of computer vision related algorithm s.At the same time,the existing deep learning-based vision technology is difficult to balance accuracy and real-time.In view of the above problems,this paper takes traffic signs,pedestrians,vehicles,and lane lines in road elements as research objects,and proposes target detection,classification,and tracking methods based on deep learning ad related filtering,which improves the robustness and practicality of various algorithms Sex.The main research contents of this article are as follows:(1)In order to classify road traffic signs with high precision,a road traffic sign classification network model based on Goog Le Net is proposed.The network is composed of sub-modules with a multi-branch asymmetric structure design,which ensures the robustness of the network to extract target features;the output characteristics of each sub-module are normalized to solve the disappearance of gradient and node deaths during network training;before the fully connected layer,using continuous pooling layer reduces the amount of network parameters and improves the classification and training speed of the network.Experiments prove that the network model can achieve high-precision classification of road traffic signs.(2)In order to achieve pixel-level positioning in lane line detection,a multi-model fusion lane line network model is designed and built.The network model introduced Deeplabv3+ network,and designed Unet1 network and Unet2 network based on Unet network.The Unet1 network increases the depth and width of the network and obtains the deep features of the lane line.The Unet2 network combines spatial pyramid and hollow convolution to get multi-scale features of the target and strengthen the connection between adjacent pixels of the input features.After implementing each sub-network,a pixel voting algorithm is designed to fuse the output of each network.The experiment proves the effectiveness of the voting algorithm and the excellent segmentation effect of the lane detection network.(3)In order to meet the demand for long-term robust tracking of specific targets in automatic driving,a single target long-range tracking algorithm based on re-detection is proposed.The algorithm based on the relevant filtering algorithm with excellent short-range tracking performance.This re-detection algorithm uses histogram of oriented gradient(HOG)to describe the target to be tracked,and uses support vector machine to obtain the target position,and proposes to use the correlation between the target and the filter template as the index to start the re-detection module.The comparison experiment with other high-level algorithms shows that Robustness and effectiveness in tracking targets.(4)In order to meet the demand for multi-target real-time detection on domestic roads and improve the ability of the detection network to perform real-time detection on Chinese roads when there are few labeled data,a network training scheme based on transfer learning is proposed.The program uses the BDD data set as the source domain for transfer learning and the self-labeled domestic data set as the target domain.Meanwhile,use YOLOv3 as the migration learning model,and utilize the model-based transfer learning method to migrate large amounts of foreign marked data to domestic roads.Experimental results show that the proposed scheme could be used for real-time multi-target detection effectively on domestic roads even when dataset amount is limited.