High-precision Vehicle And Pedestrian Detection Based On Multi-line Lidar

In recent years,benefit from the steady development of deep learning technology and the application of sensors such as lidar,unmanned driving technology has entered the era of rapid development.Among the sensors commonly used in unmanned driving,lidar,as an active sensor,has the advantages of being affected by environmental factors and accurate depth information.In the field of target detection,compared with traditional detection methods,deep learning technology has the advantages of high accuracy and automatic learning features.Therefore,both of them have been widely used in the environment perception of driverless systems.So this thesis mainly studies the deep learning perception algorithm based on multi-line lidar point cloud data.Multi-sensor fusion algorithms in current sensing algorithms have problems such as high time complexity and difficulty in maintaining stability for a long time with multisensor calibration.In addition,considering the requirements for sensors in specific scenarios(such as high ambient light interference),a sensing algorithm based on pure point cloud data is necessary.At present,most of the algorithms used for pure point cloud data are single-step methods.Although its detection speed is fast,it has the disadvantage of low accuracy.Aiming at the above problems,this thesis takes the vehicles and pedestrians in the multi-line Li DAR point cloud data as the main research objects,and on the basis of taking into account the real-time nature,a two-stage object detection network is designed to achieve the purpose of high-precision detection.This thesis uses labeling tools to complete the establishment of point cloud datasets,which is convenient for training detection models and evaluating perception algorithms.In terms of the design of the deep learning perception algorithm,considering the speed requirements,in this thesis,in the data processing,the point cloud data in the three-dimensional space is projected onto the top view,and the statistical features are extracted after rasterization,and the disordered point cloud data is extracted.Converted into a regular two-dimensional feature map,and processed by two-dimensional convolution to ensure the speed of the detection algorithm.Considering the accuracy requirements,this thesis designs a two-stage detection network that is biased towards accuracy,including feature extraction networks,PC-RPN and PCRCNN.In terms of the design of feature extraction network,in order to strengthen the detection of small objects,this thesis designs a encoding-decoding structure.On the basisof the emphasis on shallow feature extraction,the "Squeeze-and-Excitation" structure is used to strengthen the shallow features.In order to further speed up the detection network,this thesis designs a lightweight network structure.In addition,this thesis designs efficient PC-RPN and PC-RCNN networks for generating three-dimensional candidate frames and further classifying candidate regions and accurate regression of bounding boxes,respectively.Finally,in terms of detection speed,the speed of the two detection models designed in this thesis reached 90 ms and 76 ms,respectively,which met the real-time requirements of the sensing system,and the lightweight model design also achieved the purpose of speeding up the network detection speed.In terms of accuracy,the detection accuracy of the two-stage detection network designed in this thesis on the bird’s-eye view evaluation benchmark reached 0.8867.