Small object 3D detection algorithm based on lidar point cloud

Abstract

Aiming at the problems that point cloud has few available features and low positioning accuracy of small object in 3D object detection process. A new 3D object detection algorithm CFPointPillars based on improved PointPillars is proposed. First, before the feature extraction of the pseudo-image, Coordinate attention(CA) is introduced to obtain the position perception and direction perception information of the small object. Secondly, Feature Pyramid Network (FPN) structure is introduced to integrate the extracted features to obtain the precise semantic information of the small object. Finally test on KITTI public data set; The experimental results show that in terms of BEV detection benchmark, 3D detection benchmark and Average orientation Similarity(AOS), the mAP of CFPointPillars detection algorithm for car, pedestrians and cyclists reaches 70.79%, 64.44% and 71.75% respectively, which is 1.93%, 0.87% and 2.29% improvement compared with original network PointPillars respectively.