Research On Unsupervised 3D Object Detection Algorithm Based On Prior Distribution

With the rapid development of artificial intelligence technology,autonomous driving is gradually moving from theory to reality.3D object detection provides basic environmental perception and 3D scene understanding for unmanned driving,and lays a foundation for subsequent machine behavior decision and path planning.As a derivative task of 2D object detection,3D object detection needs to speculate for the smallest circumscribed cube of each object in the camera coordinate system.Its properties include center localization,cube size and yaw angle.Although great progress has been made in vision-based 3D object detection,the demand for expensive 3D annotations has greatly limited its scalability in practical applications.Facing this problem,inspired by the principle of disparity,we fully explores the reverse transformation from 2D projection to 3D bounding box,and proposes the first 3D object detection model without real 3D annotation.The main idea is to sample a large number of virtual 3D bounding boxes and project them to 2D plane to construct an artificial dataset without image data.These virtual 2D-3D pairs can replace real 3D annotations,so as to learn the conversion model from 2D projection to 3D bounding box.This paper proposes a lightweight neural network model 3DNP-Net based on the principle of Gaussian process,which only needs stereo 2D bounding box to realize the regression of 3D bounding box.In order to train the model,the sampling and generation method of the virtual data set is introduced in detail.In addition,this paper further proposes a non-maximum suppression method for 3D structures,combined with a strategy of 3D refinement using parallax alignment and photometric errors,which greatly improves the average accuracy of the results from 3DNP-Net.The results on the most influential public data set KITTI surpassed some fully supervised 3D object detection algorithms that used 3D real labels,which have proved the possibility of unsupervised 3D object detection.