Research On Environmental Perception Technology For Intelligent Vehicles Based On Deep Learning

Traffic accident?congestion?environmental pollution and energy depletion are becoming increasingly serious with the continuous growth of car ownership in China,in recent years.Therefore,it has attracted great attention of research institutions at home and abroad that researching and developing more energy-saving?environmentally?friendly and safe intelligent vehicle technology by combining of artificial intelligence technology,especially deep learning algorithm and advanced control technology.As the basic procedure for decision-making and planning task of intelligent vehicle,environmental perception plays an important role in further realizing the autopilot function by sensing the external environment state of the vehicle through camera or radar sensors,detecting and recognizing the obstacle.It has become a research hotspot in the field of intelligent vehicles in recent years due to the fact that the performance of environmental awareness technology has also become one of the main reasons for the overall performance of higher-level intelligent vehicles.This paper focuses on the application of deep learning in intelligent vehicle environment perception and the related research contents are as follows:(1)Complete the installation and debugging of hardware systems such as vision sensor?millimeter wave radar?multi-layer LADAR and IPC(Industrial Personal Computer)on the basis of investigating and researching the software and hardware framework of Intelligent Vehicle for the requirements of Intelligent Vehicle Environment Perception algorithms.Realizing data communication between multi-source sensors and IPC based on Socket/UDP communication protocol.(2)The joint calibration of camera and radar coordinate system is completed aiming at the demand of multi-source sensor fusion.In addition,complete the camera internal parameter calibration task based on Zhang's calibration method.And for the external parameter calibration of LADAR,a calibration method of optimizing the level function by genetic algorithm is designed,which completes the external parameter calibration of lidar without any data from manual measurements.(3)Researching Intelligent Vehicle obstacle detection algorithm based on deep convolution neural network,and two kinds of algorithms based on region extraction and regression are emphatically analyzed.Finally,the advantages and disadvantages of each are expounded through qualitative comparison and actual road scene test.(4)Aiming at the speed of object detection algorithm based on deep learning,the lightweight feature extraction network is reconstructed to replace the feature extraction network of SSD algorithm;aiming at the accuracy problem of SSD algorithm,the anchor scale and ratio automatically learned from the dataset that contain pedestrians and vehicles object,and the ultimate goal of this improvement is to improve the accuracy of the algorithm.(5)Aiming at the disadvantage of poor illumination robustness of vision sensor and the shortcomings of radar point cloud such as lack of texture information and too many noise points,vehicle detection algorithms based on deep learning and multi-source sensor fusion is established.The algorithm consists of two stages: first of all,target hypothesis be generated by traditional point cloud segmentation algorithms,and project the information to pixel coordinate system in order to generate fusion image containing strength and distance information by nearest neighbor interpolation based on Delaunay triangulation;the second step is hypothesis verification,training a deep residual network for target classification and finally achieving the goals that accurate orientation of target in near-end scene.