Reproduce “Person Detection and Tracking from an Overhead Crane LiDAR” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nilusha Jayawickrama, Henrik Toikka, Risto Ojala
Researchers from Aalto University developed a pipeline for real-time person detection and tracking in industrial environments using overhead-view LiDAR, achieving an Average Precision of up to 0.84 within a 5.0-meter radius and demonstrating real-time operational capability. They addressed the domain shift from autonomous driving by fine-tuning models and creating a dedicated overhead-view dataset.
Key insights
- Transfer learning from pre-trained autonomous driving models is essential for bridging the domain gap, as models without fine-tuning showed significantly degraded performance on overhead LiDAR data.
- Detector performance varies with radial distance: VoxelNeXt excelled at closer ranges (< 3.0m), while SECOND provided more stable and robust performance at larger distances (> 3.0m).
- The quality of the upstream 3D detection module is the primary determinant of overall multi-object tracking performance (MOTA, IDF1), with tracking algorithms showing comparable performance when paired with the same detector.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.