| Abstract |
This paper addresses the need for reliable person detection systems in public spaces by developing a novel dataset tailored for solid-state 3D-LiDAR sensors and evaluating various neural network architectures. The dataset was created using a Blickfeld solid-state 3D-LiDAR, capturing 265 point clouds in a controlled test environment modeled on a three-lane pedestrian crossing. The neural network architectures evaluated include VoxelNeXt, PillarNet, SECOND, PointPillar, CenterPoint, Voxel-R-CNN, PointRCNN, PartA2, and PV-RCNN. The evaluation methodology follows the KITTI benchmark metric for performance analysis. Key results indicate that voxel-based approaches like SECOND and VoxelNeXt achieve inference speeds of 10.3 FPS and 9.8 FPS on an NVIDIA Jetson AGX platform, respectively, with mean Average Precision (mAP) scores of 95% and 90%. In contrast, the hybrid approach PV-RCNN, which combines voxel-based and point-based methods, achieves a mAP of 92% but a slower inference speed of 2.5 FPS. These results underscore the trade-offs between speed and accuracy in person detection using solid-state 3D-LiDAR, highlighting the potential of voxel-based methods for real-time applications. The results contribute to the advancement of person detection technologies in public security and smart city initiatives. © 2024, KIT Scientific Publishing. All rights reserved. |
