+**********************************************************************

*

*

*                          Einladung

*

*

*

*                     Informatik-Oberseminar

*

*

*

+**********************************************************************

Zeit:  Freitag, 11. Juli 2025, 15.30 Uhr

Ort:   Raum 025, Mies-van-der-Rohe Str. 15 (UMIC Gebäude)

Referent: Dan Jia, M.Sc.

                Lehrstuhl Informatik 13

Thema: Efficient Person Detection with LiDAR Sensors for Mobile Robots

Abstract:

Person detection is a fundamental capability for mobile robots operating in unstructured and dynamic environments, including public spaces, healthcare facilities, and households. While RGB(-D) cameras combined with deep learning-based object detectors have become standard solutions, they face notable limitations such as restricted field of view, reduced depth accuracy at long range, and increased computational complexity in multi-camera setups. These limitations are particularly critical in real-time robotic applications. In contrast, 2D LiDAR sensors—widely deployed for mapping and navigation—offer accurate range measurements, wide field of view, and high frame rates, while also being less intrusive in terms of privacy. 

This talk introduces DR-SPAAM, a novel and efficient deep neural network designed specifically for detecting persons using sparse 2D LiDAR data. At the core of DR-SPAAM is a Spatial Attention and Auto-Regressive Module that aggregates temporal information across consecutive scans. This architecture addresses the inherent sparsity of 2D LiDAR data while maintaining minimal computational overhead. DR-SPAAM achieves state-of-the-art performance on the DROW benchmark dataset, reaching 70.3% average precision, while operating in real-time on both high-end and embedded GPUs. In addition, DR-SPAAM demonstrates robustness to changes in sensor specifications such as resolution and scan rate, making it adaptable to a wide range of robotic platforms without the need for re-training. 

To address the scarcity of annotated 2D LiDAR training data, an approach is introduced that automatically generates pseudo-labels using image-based detectors from calibrated cameras. These pseudo-labels enable training or fine-tuning of LiDAR-based detectors during deployment without the need for manual annotation, and performance can be further enhanced through techniques such as mixup regularization and robust loss functions.

Finally, the talk provides a comparative analysis of 2D and 3D LiDAR sensors for person detection in terms of accuracy, inference speed, localization performance, and occlusion handling. The results reveal that 2D LiDAR can match 3D LiDAR in detecting nearby, visible persons at a fraction of the cost and computational demand. However, for environments with frequent occlusions or the need for long-range detection, 3D LiDAR remains the preferred choice. Taken together, these findings offer practical guidance for selecting sensor modalities and designing robust, efficient person detection systems for a wide spectrum of robotic applications.

Es laden ein: die Dozentinnen und Dozenten der Informatik