ScorePillar: A Real-Time Small Object Detection Method Based on Pillar Scoring of Lidar Measurement

Abstract

The small object detection is essential for robot navigation, especially for avoiding vulnerable pedestrians. Usually, the points assigned to small objects in Lidar scans are sparse; detecting them efficiently and accurately is still a challenging problem. This article proposes a real-time and accurate small object detection method (ScorePillar) based on the pillar point scoring mechanism, which focuses on the relationship among points in pillars. Considering that voxel-based object detection methods are not efficient enough for real-time application, compact pillar-based structures are leveraged to represent Lidar scans for improving efficiency. For better extraction of multiscale features on pillar projection of point cloud, an ResNet-based feature extraction module is combined with an attention block and multidilation atrous convolutions to improve efficiency and accuracy further. Extensive experiments on the KITTI and nuScenes datasets show the validity and efficiency of ScorePillar. Note that ScorePillar achieves a 3.5% improvement in mean average precision (mAP) detecting pedestrian objects on the KITTI dataset and first place in the average mAP among Lidar-only methods. The code is publicly available at: https://github.com/Cao-Zonghan/ScorePillar .