Research On Key Technologies Of The 3D Detection Based On LiDAR In Structured Roads

Intelligence is the main trend in the development of future automotive technology,environmental perception is an important part of intelligent driving systems.To achieve safe driving tasks,the vehicle needs to be equipped with various sensors(Li DAR,Camera,RADAR,etc.)to perceive and understand driving scenes,which can provide important reference information for subsequent decision-making and planning.However,the urban driving environment is complex,pedestrians and vehicles are mixed,and there are traffic jams and occlusion.These problems have brought huge challenges to environmental perception.The road boundary is one of the important traffic elements in driving environments.It defines the area where the vehicle can travel and provides guidance information for decisionmaking and planning module.However,the road shape is complex and diverse,and there are occlusions inside and outside roads,these bring great challenges to road boundary detection methods.3D object detection is an important task of environmental perception.The current Li DAR-based 3D object detection methods rely on predefined anchors as prior information.This greatly increases the complexity and computation amount of models and affects the generalization ability and performance of models.Therefore,how to eliminate the dependence of the 3D object detection methods on anchors,reduce the complexity and calculation of models,and improve the generalization ability and performance of models requires further exploration.Besides,the sparseness of Li DAR point cloud and lack of color and texture limits the performance of Li DAR-based object detection methods.The camera can provide images with color and texture information,but the image and point cloud are the expressions of surrounding environments in different perspective views,which brings great challenges to the object detection methods based on the fusion of Li DAR and camera.Therefore,how to realize the adaptive fusion of multi-mode features according to the differences of multi-view features is an urgent problem to be solved in the current 3D object detection methods based on the fusion of Li DAR and camera.To solve the above problems,this paper uses point cloud and image data to carry out indepth research on road boundary detection based on Li DAR,3D object detection based on Li DAR,and 3D object detection based on Li DAR and camera.The specific research contents are as follows:1.Research on road boundary detection method based on Li DAR.To solve the problems of obstacle occlusion and complex road shapes,this paper studies the classification and extraction of road boundary points under the complex road shapes and obstacle occlusions based on road boundary feature points extracted manually.For road boundary point classification,this paper proposes two methods: the method based on improved density clustering and the method based on road segmentation line.Although the improved density clustering method can accurately classify road boundary points,it has a large computational burden.The method based on the road segmentation line can not only ensure accuracy but also greatly reduce the computational burden and improve the efficiency of the algorithm.Besides,the iterative Gaussian Process Regression model is proposed to model road boundaries and extract road boundary points.The non-parametric Gaussian Process Regression model can model a variety of complex road boundary shapes,while effectively eliminating the interference of obstacles inside and outside roads.The experimental results of KITTI dataset show that the proposed road boundary detection method can accurately extract the left and right road boundary points on roads with diverse shapes and occlusions while achieving realtime performance.The proposed method meets the real-time and accuracy requirements for intelligent driving.2.Research on 3D object detection method based on Li DAR.Aiming at the dependence of current Li DAR-based 3D object detection methods on anchors,this paper draws on the idea of key point detection in the image object detector Center Net,and proposes an anchor-free 3D object detection network Center Net3 D for point clouds.In the detection head,this paper designs a detection head based on key points to classify the center points of objects and directly regress 3D bounding boxes.To enable the model to perceive the shape information of objects,this paper proposes an corner point classification module to improve the quality of bounding box regression.Besides,considering that single-stage detectors suffer from the discordance between the predicted bounding boxes and corresponding classification confidences,this paper proposes an efficient keypoint-sensitive warping module by using the center points and the corner points.The experimental results of KITTI dataset show that: compared with the anchor-based models,the proposed Center Net3 D in this paper achieves better performance,significantly improves the performance for difficult-level objects,and eliminates anchor and NMS post-processing.The entire network architecture is more concise,and the inference speed is faster.3.Research on 3D object detection method based on the fusion of Li DAR and camera.Aiming at the problem of feature fusion of image and multi-view point cloud,this proposes a single-stage multi-view adaptive fusion 3D object detection network MVAF-Net.The whole network consists of three parts: Single-View Feature Extraction(SVFE),Multi-View Feature Fusion(MVFF),and Fusion Feature Detection(FFD).In the SVFE part,the data from Bird's Eye View(BEV),Range View(RV),and Camera View(CV)are used as inputs,and a threebranch backbone network is used to extract multi-view features.In the MVFF part,this paper proposes Attention Pointwise Fusion(APF)module and Attention Pointwise Weighting(APW)module to achieve pointwise adaptive learning and fusion of multi-view features.Experiments on the KITTI dataset verify the effectiveness of Attention Pointwise Fusion module and Attention Pointwise Weighting module.The proposed MVAF-Net produces competitive results on the KITTI dataset and surpasses all single-stage fusion methods.At the same time,compared with the proposed Center Net3 D,it significantly improves the performance for longdistance,difficult and small objects,and achieves the best trade-off between speed and accuracy.