Research Of Object Detection Method Based On 3D Roadside LiDAR

Environmental perception plays a vital role in ensuring the safety of autonomous driving vehicles.Roadside perception units can enlarge the perception field of view,increase the perception distance,and achieve vehicle-road collaborative perception with wireless communication technology,endowing autonomous driving vehicles with beyond-visual-range perception capabilities.3D Li DAR has advantages such as high detection accuracy,wide perception range,strong stability,and all-weather operation.By deploying Li DAR on roadside perception unit,a“God’s eye view”can be obtained to acquire high-precision traffic information.However,roadside Li DAR point clouds have characteristics such as sparse distribution and uneven numbers foreground and background points,which can impact the performance of object detection methods.Additionally,existing roadside Li DAR point cloud datasets suffer from limitations such as small data size and significant background homogeneity,leading to suboptimal results when used for model training.In order to improve the accuracy of traffic object detection based on roadside Li DAR,this thesis relies on the Zhejiang Provincial Key Research and Development Program project"Research on Holographic Digital Sensing Technology for Intelligent Expressways"（211424200512）,focuses on traditional and deep learning-based object detection methods,and obtains the following innovative results:（1）To address the problem of background filtering difficulty caused by roadside Li DAR swing and background dynamic change,this thesis proposes a background filtering and updating method considering roadside Li DAR swing.A point cloud height map is constructed based on the laser beam azimuth angle and point height,and a background model is constructed by performing temporal median filtering on the point cloud height map;a filtering threshold selection strategy considering the swing angle of the Li DAR and the vertical angle of the laser beam is proposed to achieve accurate segmentation of foreground and background points;a background update method based on point cloud height map frame difference is proposed,which reduces the impact of background dynamic change on background filtering accuracy.Real point cloud sequences are collected by deploying roadside Li DAR for experiments.The experimental results show that the proposed method’s rate of background point filtering(R_BPF)remains above 98%,and the rate of foreground point extracting(R_FPE)remains above 93%.Compared with two typical methods,the proposed method improves R_BPF by more than 1.47%and R_FPE by more than 5.12%.Combining DBSCAN and SVM to achieve object classification,the classification accuracy of pedestrians is improved by 14.81%,and that of background is improved by 13.33%.（2）To address the problem of low-quality roadside Li DAR point cloud dataset affecting the accuracy of object detection network,this thesis proposes a rapid roadside Li DAR point cloud construction method based on CARLA.A variety of typical traffic scenarios are constructed in a simulation environment,and simulation Li DAR parameters are set according to the real Li DAR,and a large number of simulation point clouds are generated quickly;a point cloud simulation method is proposed to simulate the characteristics of real point clouds such as motion distortion and noise with high fidelity,reducing the difference between simulation point clouds and real point clouds.The simulation point clouds and BAAI-VANJEE real point clouds are used to construct a hybrid dataset for training various networks.The experimental results show that compared with the baseline dataset BAAI-VANJEE,using the constructed hybrid dataset to train networks such as Point Pillars can improve the detection accuracy by more than4.27%.（3）To address the problem of domain shift between simulation point clouds and real point clouds leading to lower object detection accuracy,this thesis proposes a domain adaptive pillar-based 3D point cloud object detection method.Adversarial training is introduced to encourage the backbone network to extract common deep features of simulation point clouds and real point clouds,improving the training effect of mixed dataset on the network;farthest point sampling is introduced to improve the uniformity of sampling process and extract local point cloud spatial features more comprehensively;sparse convolution is introduced to improve model computation efficiency;based on the loss value,difficult negative samples are selected for training,alleviating the problem of imbalance between positive and negative samples.The proposed method is evaluated on the constructed hybrid dataset.The experimental results show that the proposed method can significantly alleviate the overfitting problem caused by domain shift,compared with existing object detection networks,the proposed method improves car detection accuracy by more than 4.35%,and the inference speed reaches 89Hz.This thesis conducts in-depth research of traffic object detection tasks based on the characteristics of roadside Li DAR point cloud,effectively improving object detection accuracy and efficiency,providing technical support for promoting application of roadside perception units with roadside LiDAR as the core.