Person Detection and Tracking from an Overhead Crane LiDAR

summary

Video file (mp4)

The gist

The gist This paper investigates person detection and tracking in an industrial indoor workspace using a LiDAR mounted on an overhead crane.

In short

This research developed a system to detect and track people in industrial indoor spaces using a LiDAR sensor mounted on an overhead crane. Facing challenges from the unique overhead viewpoint, researchers adapted various 3D detection models to achieve high accuracy, reaching an AP of 0.84 within 5 meters. They also integrated lightweight tracking methods that maintain person identities over time, successfully bridging the gap between standard driving datasets and this specific indoor sensing scenario.

Key concepts

Domain Shift
This occurs when a model trained on one type of data (like vehicle-centric LiDAR) performs poorly on a different type of data (like overhead industrial scans). The overhead viewpoint creates a significant domain shift because the sensor sees the environment from above, which is fundamentally different from how standard sensors are usually used in cars. This shift motivates the need for new training and adaptation techniques.
Voxel-based Backbones
These are specific types of deep learning models (like VoxelNeXt and SECOND) that process 3D point cloud data by dividing space into discrete 3D grids called voxels. Instead of looking at every single point, the model analyzes groups of points within these voxels to extract features about objects, making them effective for handling the sparse and perspective-heavy data from overhead LiDAR.
Tracking-by-Detection
This is a method where a separate detection system finds people in each frame, and then a tracking algorithm links those detections across consecutive frames to maintain consistent identities. The paper uses lightweight modules like AB3DMOT and SimpleTrack to associate the newly detected objects with existing tracks based on motion consistency or geometric overlap, ensuring that a person's ID remains stable throughout the monitoring period.

Terminology used across episodes

This episode discusses

The paper

Person Detection and Tracking from an Overhead Crane LiDAR · Read on arXiv

Nilusha Jayawickrama, Henrik Toikka, Risto Ojala

School of Engineering, Aalto University · School of Electrical Engineering, Aalto University

This paper investigates person detection and tracking in an industrial indoor workspace using a LiDAR mounted on an overhead crane. The overhead viewpoint introduces a strong domain shift from common vehicle-centric LiDAR benchmarks, and limited availability of suitable public training data. Henceforth, we curate a site-specific overhead LiDAR dataset with 3D human bounding-box annotations and adapt selected candidate 3D detectors under a unified training and evaluation protocol. We further integrate lightweight tracking-by-detection using AB3DMOT and SimpleTrack to maintain person identities over time. Detection performance is reported with distance-sliced evaluation to quantify the practical operating envelope of the sensing setup. The best adapted detector configurations achieve average precision (AP) up to 0.84 within a 5.0 m horizontal radius, increasing to 0.97 at 1.0 m, with VoxelNeXt and SECOND emerging as the most reliable backbones across this range. The acquired results contribute in bridging the domain gap between standard driving datasets and overhead sensing for person detection and tracking. We also report latency measurements, highlighting practical real-time feasibility. Finally, we release our dataset and implementations in GitHub to support further research

DOI: 10.1109/UR69537.2026.11626794

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Person Detection and Tracking from an Overhead Crane LiDAR".

Jane: The gist This paper investigates person detection and tracking in an industrial indoor workspace using a LiDAR mounted on an overhead crane.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’re starting with the title and the names behind this work, Person Detection and Tracking from an Overhead Crane LiDAR. We need to know who did this because it sets the tone for how serious the research is.

Jane: The authors are Nilusha Jayawickrama, Henrik Toikka, and Risto Ojala from Aalto University in Finland. That gives us a good idea of the academic background we’re dealing with here.

Lu: Their focus seems to be on building a robust system that handles this specific overhead viewpoint by adapting existing three dee detection methods <ref:2603.04938#pg1>.

Meng: Adapting existing methods is fine, but the real question is whether those adapted models actually work when you deploy them on a busy factory floor where things change constantly.

Lalam: The paper specifically mentions they curated a site-specific dataset with three dee human bounding box annotations to handle that domain shift and data scarcity issue <ref:2603.04938#pg1,dataset with 3D human bounding box annotations>.

The paper's summary: Tom: Now, looking at the summary of Person Detection and Tracking from an Overhead Crane LiDAR, the main goal is to process sequential point clouds from that overhead sensor to output human instances frame by frame.

Jane: It’s not just detection; they are integrating a deep learning-based three dee detector with a tracking component to maintain person identities over time <ref:2603.04938#pg1,to maintain person identities over time>.

Lu: They test several candidate architectures, including PointPillars, SECOND, PV-RCNN, VoxelNeXt, and Voxel RCNN. They chose these based on compatibility with BEV representations and low latency needs.

Meng: I see they are comparing voxel-based methods like VoxelNeXt against others to see which one handles the perspective projection better in this specific setup.

Lalam: They also adapted lightweight tracking-by-detection modules, ABthree deeMOT and SimpleTrack, to associate those detections across consecutive frames.

The paper's improvements: Tom: The paper points out a few key improvements they made to the general approach, especially around how the detection works under this challenging viewpoint.

Jane: They suggest that modern voxel-based backbones are actually better at handling the sparsity and perspective you get from a single overhead LiDAR sensor.

Lu: Specifically, they found that VoxelNeXt performs best for ranges less than three meters because of its recall, F1, and AP values in that area.

Meng: That’s interesting because it suggests a specific range optimization rather than just picking the highest overall score across all distances.

Lalam: They also found that SECOND is more robust at larger radial distances greater than three meters where the point density starts to decrease with range.

Conclusion: Tom: So to wrap up, the conclusion of Person Detection and Tracking from an Overhead Crane LiDAR says they can reliably implement person detection and tracking in an indoor workspace using this setup.

Jane: The main finding is that the best adapted detector configurations achieved an average precision up to zero point eight four within a five point zero meter horizontal radius from the sensor.

Lu: That figure, zero point eight four AP, is solid for human targets in their full operating radius of five point zero meters horizontally; it shows a good practical performance ceiling for this setup.

Meng: The limitation they flag is that their dataset was quantity-limited, and they also capped the horizontal radial distance for evaluation at four point five meters because it was a practical constraint of their experimental workspace.

Lalam: That means while the results are strong up to four point five meters, you can't really generalize further without more data or a larger physical space.

More episodes

← Home