Research On 3D Object Detection Algorithm Based On Radar Point Cloud And Image

With the improvement of the performance of hardware equipment and the rapid development of artificial intelligence technology,more and more fields begin to use point clouds collected by lidar to understand the three-dimensional scene,such as autonomous driving and augmented reality.The RGB image collected by the camera has rich texture information,while the radar point cloud is not easily affected by external factors and has strong anti-interference ability.This article uses radar point cloud and images as the basis to detect objects in three-dimensional space.The main work is summarized as follows:(1)A 3D object detection algorithm(Semantic-based Frustum Pointnet,SFP)based on 2D semantic segmentation is proposed.The algorithm makes full use of the texture information of the RGB image to semantically segment the image and project it into the point cloud space.The purpose is to remove the influence of the noisy point cloud in the point cloud space on the 3D object detection task.The Deeplabv3+ network is introduced to segment the region of interest and non-interest in the automatic driving scene.After the segmentation result being obtained,it is projected into the point cloud space to form a frustum.Finally,3D instance segmentation and 3D frame estimation are performed within the scope of the frustum.The experimental results prove that the proposed algorithm can well separate the detection object point cloud from the non-interest point cloud,effectively reduce the impact of noisy point clouds in the scene on the object detection task.In the KITTI dataset,the m AP of object detection in three levels of easy,moderate and hard has been increased by 4.44%,2.54% and 2.26% respectively.(2)A 3D object detection algorithm with sparse point cloud completion is proposed.The algorithm adopts the encoder and decoder mechanisms to address the sparse and incomplete characteristics of the point cloud collected in the actual three-dimensional scene.Through the point cloud completion network,part of the input sparse point cloud becomes dense.This algorithm adopts the cascade decoding method of fully-connected decoder and folding-based decoder in the decoding network.According to the characteristics of the cascading decoder method,a new composite loss function is defined.In addition to the original folding-based decoder stage,the composite loss function also adds the loss in the fully-connected decoder stage to ensure that the total error of the decoding network is minimized,so that the point cloud completion network generates a dense point cloud with more complete information,and applies the completed point cloud to the 3D object detection task.The experimental results show that the proposed algorithm can well complete the sparse car point cloud in the KITTI dataset,and makes the average accuracy of 3D object detection significantly improved,especially for the data of moderate and hard levels.The improvement reached 6.81% and 9.29% respectively.