Research On Machine Learning Based Object Detection And Augmented Reality Technology For Intelligent Vehicle

The augmented reality head-up display system(AR-HUD)integrates driving assistance information with the actual traffic scene in the driver's normal driving vision,which can not only expand environmental perception information,but also avoid the driver from checking the instrument panel too much,which effectively improves the driving safety.The AR-HUD technology has been highly valued by the automotive electronics industry and has become a hot spot of automotive intelligent technology research.However,if the real-scale real-vehicle application is truly realized,there are still many related technical issues to be solved urgently.In this thesis,the lightening problem,system combination calibration and virtual image distortion processing of object detection deep network in the AR-HUD engineering application are studied.The main contents are as follows:1.To slove the real-time problem of object detection algorithm based on deep learning in an embedded environment,this thesis uses lightweight neural network Mobile Net to reduce the amount of network parameters and model size,and uses long and short-term memory networks to design a road object detection algorithm based on Mobile Net-SSD with time dimension features.Experiments verify that the algorithm can effectively improve the real-time performance of object detection while ensuring the accuracy of object detection in the embedded environment.2.Aiming at the problem of system combination calibration and virtual image distortion processing in AR-HUD engineering application,this paper proposes the AR-HUD system calibration method and various pre-distortion algorithms of virtual image.The AR-HUD system calibration method includes multi-camera combination calibration and the measurement of the position and size of the AR-HUD projection virtual image;the virtual image pre-distortion algorithm includes the linear interpolation algorithm of the pre-distortion of AR-HUD image under the condition of static eye position,the multi-linear interpolation algorithm of the pre-distortion of AR-HUD image under the condition of dynamic eye position,and the neural network learning method of the AR-HUD virtual image pre-distortion under the condition of dynamic eye position.Experiments have proved that the calibration method of AR-HUD system proposed in this paper can effectively improve the calibration accuracy of AR-HUDsystem,and the virtual image distortion correction algorithm can effectively improve the distortion of AR-HUD virtual image.3.According to the actual needs of AR-HUD system engineering applications,this paper comprehensively applies the above methods and researches and designs the virtual and real registration,virtual image pre-distortion,safety warning and other functional software modules in the AR controller of vehicle-mounted AR-HUD system based on multi-eye position;in addition,a systematic road experiment was conducted on a real vehicle to verify the effectiveness of related algorithms in the vehicle-mounted AR-HUD system,and the experiment verified that the AR method plays an important role in enhancing the driving environment information,improving the safety of the driving system and the user experience.