LiDAR-Based 3D Object Detection using Self-ensembling Feature Preservation PointPillars in Urban Environments

초록

Object detection, which is essential for configuring autonomous driving, generally includes 2D, bird’s eye view (BEV), and 3D detection. Among these detection methods, 3D object detection is becoming more active to be compatible with real environments. However, most of the currently released deep learning networks for 3D object detection have been developed using the KITTI dataset, which is labeled only for objects in the front camera angle. To solve this problem, this paper presents a method of 3D object detection using a point cloud dataset collected by a 128-channel LiDAR sensor installed on a vehicle for driving data collection and labeled on objects of 360 degrees. The proposed method based on PointPillars, which currently boasts the highest speed, is designed to be trained on groups of related features and uses self-ensembling in SE-SSD, which shows the best accuracy on the KITTI dataset. This deep learning model demonstrates better accuracy for R40 evaluation indicators than other deep learning networks, and the practicality of the model is validated by verifying the performance of an actual autonomous vehicle. © ICROS 2022.

키워드