Robust Person Re-Identification for Service Robots Via One-Class Body-Part Transformer and Continual Learning
Enrique Aleman-Gallegos, Sven Wachsmuth
AI summary
Problem
Service robots struggle to maintain accurate person tracking and re-identification in public spaces due to crowds, occlusions, appearance changes, and frequent ID swaps. Existing vision-based methods often fail to adapt online or ignore subtle inter-part appearance cues needed for robust identification.
Approach
The system extends the SORT tracker with depth data to prevent ID swaps and employs a One-Class Body-Part Transformer trained online via continual learning. A memory manager selects diverse positive-negative pairs and pseudo-negative samples to accelerate convergence and maintain a robust target representation.
Key results
- Outperforms state-of-the-art methods on public datasets with high distractor similarity and occlusions
- Successfully re-identifies and follows a target person in real-world robot experiments with multiple distractors and illumination changes
- Achieves reliable real-time performance on embedded NVIDIA GPU platforms
- Memory management with pseudo-negative samples accelerates online learning convergence and improves classifier robustness
Why it matters
Enables service robots to reliably track and assist specific users in complex, dynamic public environments without custom hardware or pre-calibration.
Abstract
This work presents a robust person tracking and re-identification system designed for Human-Robot Interaction applications. The approach introduces the One-Class Body- Part (OCBP) Transformer, trained online to model interactions among body-part features and construct a robust target rep- resentation. To improve data association and reduce identity swaps during the tracking phase, the SORT tracker is extended with depth information in order to provide correct samples for the Online Continual Learning (OCL) setting. The transformer is further enhanced through the use of pseudo-negative samples, which accelerate convergence during the online learning phase. Ablation studies compare the performance of the memory man- agement system using different sample insertion configurations and highlight the benefit of using pseudo-negative samples. The proposed method is evaluated on a public dataset, where it out- performs state-of-the-art approaches in challenging scenarios, and is validated in a real-world person-following experiment with a robotic platform in an environment with multiple distractors, occlusions, out-of-view situations and illumination changes. Despite these complexities, the robot consistently re- identified and followed the target individual. Runtime analysis demonstrates that the system operates reliably on embedded computing platforms with NVIDIA GPUs, making it both robust and resource-efficient for real-world deployment.