Person Detection and Tracking from an Overhead Crane LiDAR

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

Problem Setting and Motivation

Ensuring safety for human workers near automated systems requires reliable person detection to enable downstream monitoring or collision avoidance The overhead viewpoint introduces a strong domain shift from common vehicle-centric LiDAR benchmarks and limited availability of suitable public training data This work addresses the problem of person detection using an overhead-mounted static LiDAR sensor within a prediction framework where a deep learning–based 3D detector processes sequential point clouds to output per-frame human instances The overhead viewpoint of this indoor work environment constitutes a domain shift from conventional frontal type (e.g. vehicle-centric) LiDAR perception pipelines, motivating the research on transfer learning and comparative evaluation of adapted detectors.

Detection Methodology

The detection module adapts and evaluates multiple LiDAR-based 3D detection modules under a unified training and evaluation protocol The primary criteria for selecting the candidate detection architectures were compatibility with BEV/voxel representations to capture top-down geometry, support for low-latency deployment, documented open-source implementation, and reported evidence on standard 3D detection benchmarks. The selected architectures include PointPillars (PP), SECOND (SEC), PV-RCNN, VoxelNeXt, and Voxel RCNN. For training, a collected dataset was manually annotated with 3D bounding boxes for human targets using labelCloud Each detector was trained for 250 epochs with a batch size of 4 and using the AdamW optimizer. Input point clouds were shuffled during training to improve robustness, and backbones were initialized from pretrained weights on open-source autonomous driving datasets like KITTI [28] and nuScenes [29].

Tracking Methodology

The tracking module associates detections across consecutive frames to maintain person trajectories and IDs Two lightweight tracking-by-detection modules, AB3DMOT [20] and SimpleTrack [21], were adapted for this task. Both trackers take per-frame 3D detections as input and maintain tracks over time The AB3DMOT-style tracker performs data association based on motion consistency using the Mahalanobis distance between predicted track positions and detected object centers in 3D space, with BEV-based overlap used as a secondary cue. In contrast, SimpleTrack relies directly on geometric overlap between predicted and detected bounding boxes in BEV for association.

Results and Discussion

The distance-sliced detection results suggest that modern voxelbased backbones (VoxelNeXt and SECOND) generalize better to the sparsity and perspective induced by a single overhead LIDAR VoxelNeXt is best for < 3 m range as observed by its respective recall, F1 and AP values SECOND demonstrates more robustness at larger radial distances (> 3 m) where point density decreases with range. The best-performing detector configurations achieved average precision (AP) up to 0.84 within a 5.0 m horizontal radius, increasing to 0.97 at 1.0 m Tracking quality is broadly similar between the two implemented modules, with the dominant factor being the upstream detector quality. The end-toend tracking results further showed that tracking-by-detection methods can maintain the track IDs of persons reasonably well over time.

Conclusion and Limitations

Overall, this study indicates that reliable overhead-view LiDAR person detection and tracking can be accurately implemented in an indoor workspace The best-performing detector configurations achieved AP values up to 0.84 when evaluating human targets within our full operating radius of 5.0m from the LiDAR (horizontally) A limitation with the current study is that our dataset was quantity-limited The horizontal radial distance for evaluation was also capped at 4.5 m as it was a practical constraint of our experimental workspace. Future work could push the distance boundaries to identify the point at which LiDAR point sparsity becomes too low to sufficiently represent human targets for detection. The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.5 m as it was a practical constraint of our experimental workspace (thus, human targets observed only up to this range). The horizontal radial distance for evaluation (Table I) was also capped at 4.

1. Bold backbone selection based on distance: The system should dynamically select between VoxelNeXt for near-field (3 m), as "VoxelNeXt is best for < 3 m range as observed by its respective recall, F1 and AP values while SECOND demonstrates more robustness at larger radial distances (> 3 m)."

2. Integration of distance-sliced evaluation into deployment: The system can perform real-time risk assessment by quantifying the practical operating envelope using distance-based slicing to depict how performance changes across the working radius, allowing operators to understand performance degradation as they move away from the crane's sensor.

3. Enhanced tracking robustness through IoU threshold tuning: To improve ID consistency in close interactions, the tracking module can be tuned by lowering the association threshold; this increased MOTA and IDF1 for all detector-tracker combinations which indicates that a portion of errors at IoU = 0.3 are attributable to modest BEV misalignment rather than complete failures in association.

4. Real-time feasibility benchmarking: The system's deployment strategy can be informed by comparing inference speeds, as the reported latency measurements provide a practical indicator to estimate which architectures remain feasible for real-time applicability based on percentile data.

5. Domain gap mitigation through transfer learning protocol: To deploy the system in new industrial settings, it should leverage transfer learning and evaluation of relevant LiDAR models for person detection based on an overhead-view LiDAR viewpoint within an industrial crane workspace, bridging the gap between standard driving datasets and overhead sensing.

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