GDPR-Compliant Person Recognition in Industrial Environments Using MEMS-LiDAR and Hybrid Data

The reliable detection of unauthorized individuals in safety-critical industrial indoor spaces is crucial to avoid plant shutdowns, property damage, and personal hazards. Conventional vision-based methods that use deep-learning approaches for person recognition provide image information but are sensitive to lighting and visibility conditions and often violate privacy regulations, such as the General Data Protection Regulation (GDPR) in the European Union. Typically, detection systems based on deep learning require annotated data for training. Collecting and annotating such data, however, is highly time-consuming and due to manual treatments not necessarily error free. Therefore, this paper presents a privacy-compliant approach based on Micro-Electro-Mechanical Systems LiDAR (MEMS-LiDAR), which exclusively captures anonymized 3D point clouds and avoids personal identification features. To compensate for the large amount of time required to record real LiDAR data and for post-processing and annotation, real recordings are augmented with synthetically generated scenes from the CARLA simulation framework. The results demonstrate that the hybrid data improves the average precision by 44 percentage points compared to a model trained exclusively with real data while reducing the manual annotation effort by 50 %. Thus, the proposed approach provides a scalable, cost-efficient alternative to purely real-data-based methods and systematically shows how synthetic LiDAR data can combine high performance in person detection with GDPR compliance in an industrial environment.

💡 Research Summary

The paper addresses the critical need for reliable, privacy‑preserving person detection in safety‑critical industrial indoor and outdoor spaces. Conventional vision‑based deep‑learning approaches, while effective, capture identifiable visual cues (faces, clothing) that conflict with the European Union’s General Data Protection Regulation (GDPR). To overcome this, the authors propose a system built around a Micro‑Electro‑Mechanical Systems LiDAR (MEMS‑LiDAR) sensor, which records only 3‑D point clouds devoid of any personal identifiers, thereby inherently satisfying GDPR requirements.

A major obstacle for MEMS‑LiDAR‑based solutions is the scarcity of annotated training data. Real‑world data collection is expensive, labor‑intensive, and requires extensive manual labeling. To mitigate this, the authors generate synthetic point clouds using the open‑source CARLA simulator (v9.15, Unreal Engine 4). They model the MEMS‑LiDAR scanning mechanism with sinusoidal mirror motions, extract depth from CARLA’s virtual depth camera, and convert the data into 3‑D coordinates. Virtual pedestrians are spawned and moved along predefined trajectories, allowing automatic extraction of precise 3‑D bounding boxes. This pipeline yields 2,100 synthetic scenes covering 3–10 persons per frame, with full annotation at zero manual cost.

Four training configurations are evaluated: (M1) 100 % real data (2,100 samples), (M2) 100 % synthetic data (2,100 samples), (M3) a balanced 50/50 mix (1,050 real + 1,050 synthetic), and (M4) a 70/30 synthetic‑real mix (1,470 synthetic + 630 real). All models share identical hyper‑parameters: the SECOND (Sparsely Embedded Convolutional Detection) voxel‑based network, Adam optimizer with OneCycleLR (initial LR = 1e‑4), 120 epochs, and data augmentations (rotation, scaling, mirroring). Training is performed on an AMD Ryzen 9 3900X with 64 GB RAM and an RTX 2080 GPU.

Evaluation uses a separate test set of 230 manually annotated point clouds captured in a real industrial environment (Smart Factory OWL, Lemgo). The metric is Average Precision (AP) at IoU = 0.5. Results: M1 (real only) achieves AP = 0.10; M2 (synthetic only) improves to 0.27 (+17 % relative); M3 (50/50) reaches 0.54 (+44 % over M1); M4 (70/30) attains 0.45 (+35 % over M1). Qualitative visualizations confirm that mixed‑data models produce bounding boxes that align closely with ground truth, even under occlusion and multi‑person scenarios, whereas the real‑only model frequently misses detections.

The study demonstrates three key contributions: (1) a GDPR‑compliant perception pipeline that eliminates personally identifiable visual data, (2) a cost‑effective synthetic data generation framework that automatically provides high‑quality annotations, and (3) empirical evidence that hybrid training (real + synthetic) substantially boosts detection performance and generalization to unseen industrial scenes while halving manual labeling effort.

Future work suggested includes extending the sensor model to other MEMS‑LiDAR variants, applying domain‑adaptation techniques to further close the reality gap, and integrating real‑time multi‑object tracking or multi‑sensor fusion for comprehensive safety monitoring. The presented approach offers a scalable, legally sound solution for industrial safety applications.