Robust 3D Object Detection for Moving Objects Based on PointPillars

Ryota Nakamura, Shuichi Enokida; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2022, pp. 611-617

Deep learning techniques have been applied success-fully to detecting objects from video images, and the use of three-dimensional (3D) point clouds obtained from light detection and ranging (LiDAR) with VoxelNet and other techniques have previously been proposed for use in highly accurate object detection methods that are robust against lighting changes. However, while object detection from video images with deep learning has been observed to be continuous and stable, there are times when a few continuous frames suddenly go undetected, thereby resulting in a phenomenon known as a momentary missed detection.Extending the methodology discussed in a previous paper that examined the cause of these momentary missed detection in object detection in 3D point clouds with VoxelNet, this study proposes a robust network for detecting moving objects while considering the cause of similar momentary missed detection in PointPillars, which is an encoder developed based on VoxelNet.