Research On Key Technologies Of Visual Localization And Control For Self-Driving Car

With the progress of society and the continuous development of the automobile industry,driverless cars have emerged at the historic moment.They have become a major force in major automakers for a time,and have become a hot topic in the current society.From a system perspective,unmanned driving systems include environmental awareness systems,localization-navigation systems,and decision control systems.Among them,the localization-navigation system and the decision control system are the two key systems of the unmanned vehicle,and the environment awareness system is to ensure the unmanned vehicle’s understanding of the environment and grasp the service,which is the foundation of the entire process of unmanned driving.At present,the traditional method is still the main technology to support the driverless cars.However,traditional methods have poor robustness to problems such as light changes and scene changes,and it is difficult to meet the application requirements of driverless cars in real scenes.Therefore,this paper focuses on the two key technologies of unmanned vehicle localization and control as research points,using visual sensors to build a deep neural network to achieve end-to-end unmanned vehicle location accurate estimation and vehicle steering control,and from the accuracy and real-time And generalization and other angles to solve the positioning and control problems of unmanned vehicles.This article mainly completed the following three aspects of work:1.Aiming at the traditional feature-based method,the extraction of key points and the calculation of descriptors have the disadvantages of time-consuming,low feature utilization and easy failure when the texture is not obvious,and the direct method is insufficient to be sensitive to changes in illumination.In this paper,CNN is used to extract rich scene feature information to achieve accurate perception of the surrounding environment,and Bi-SLTM is used to extract serialized information as a constraint.An R-CNN network architecture is proposed to ensure the accuracy of unmanned vehicle pose estimation while ensuring In real time,it provides accurate localization information for high-speed unmanned vehicles.The experimental results show that the R-CNN-based vehicle localization algorithm proposed in this paper improves the accuracy of pose translation by 87.6%and the accuracy of pose rotation by 1.7%compared with traditional feature-based algorithms.2.Aiming at the problem that the traditional unmanned vehicle control algorithm cannot meet the long-term prediction in continuous scenes,this paper constructs a new end-to-end deep neural network,which takes a single image as a unit and five continuous images as a sequence to perform car direction corners Prediction.Experimental results show that the network can not only reduce the accumulation of errors,but also mine more effective geometric information from consecutive image frames to improve the prediction accuracy.The control accuracy is 65.4%higher than Nvidia-Net.3.Autonomous driving technology is a very complex system,and no single localization or control module can guarantee the normal operation of unmanned vehicles.Therefore,this paper proposes a method of fusing the localization and image feature extraction modules.Specifically,the localization information and the image information are fused in order to more accurately predict the direction of the unmanned vehicle and control the turning angle,so as to better meet the normal driving needs of the unmanned vehicle in multiple scenarios.The experimental results show that the control accuracy of the control model of fusion localization information and image information proposed in this paper is 74.5%higher than that of the control model without fusion localization information.