FlowGait: Enabling Robust Long-Term Gait Recognition Across Real-World Covariates with mmWave Radar

Paper Title

FlowGait: Enabling Robust Long-Term Gait Recognition Across Real-World Covariates with mmWave Radar

Publication Info

• Topic area: Long-term gait recognition using mmWave radar for smart home applications.
• Keywords: Gait recognition, mmWave radar, continual learning, self-training, transformer networks, pseudo-labeling, biometric systems, smart home, long-term adaptation, real-world covariates.

Background and Problem

• Problem / challenge: Existing gait recognition systems struggle with long-term reliability due to domain shifts caused by covariates like clothing, carrying items, and walking routes. These systems often rely on static models validated in controlled environments, which fail to adapt to real-world variability.
• Significance: Robust gait recognition is crucial for enabling personalized, proactive, and unobtrusive smart home interactions, such as safety monitoring and ambient intelligence. Addressing these challenges can enhance usability and security in domestic environments.
• Motivation and related work: Prior approaches using cameras, Wi-Fi, and ultrasound have limitations in privacy, resolution, and environmental sensitivity. mmWave radar offers advantages like privacy preservation and robustness but has not been effectively adapted for long-term deployment. Existing methods lack mechanisms for continual learning and adaptation to evolving gait patterns, leaving a gap in practical applications.

Solution

• Proposed approach: FlowGait, a mmWave-based framework integrating self-training and continual learning to adapt to evolving gait patterns using unlabeled data.
• Novelty:
  1. Introduction of a Transformer-based feature extraction network tailored for elongated mmWave heatmaps.
  2. Development of a two-stage step-traversal labeling algorithm for accurate pseudo-labeling of unlabeled data.
  3. Implementation of a core-set mechanism to manage data accumulation and prevent catastrophic forgetting during continual learning.
  4. Validation across diverse datasets, including cross-covariates, cross-routes, and long-term scenarios, demonstrating superior robustness and accuracy.
• Procedure and key techniques:
  1. Preprocess raw radar signals into Range-Doppler heatmaps and segment them into discrete steps.
  2. Use a hierarchical Transformer model to extract spatial and temporal gait features.
  3. Apply the step-traversal labeling algorithm to propagate high-confidence pseudo-labels across traversals.
  4. Employ a core-set mechanism for efficient daily model updates, balancing historical and new data.

Results

• Concrete findings:
  • Recognition accuracies: 94.8% (cross-covariate), 98.6% (cross-route), and 96.6% (cross-day).
  • Long-term performance decay reduced from 13.6% to 1.4%.
  • Exceptional accuracy for elderly (97.7%) and children (97.5%) groups.
  • Real-time inference latency: 7.16 ms per sample; daily model updates completed in ~6 minutes.
• Advantage over baselines:
  • Outperformed state-of-the-art methods (e.g., RDGait) by 3.2% in Top-1 accuracy and 0.06 in mAP.
  • Achieved 6–21% higher accuracy in cross-covariate and cross-route scenarios compared to supervised and semi-supervised baselines.
• Experiments / evaluation:
  • Datasets: Cross-covariate (12 conditions), cross-route (11 routes), cross-day (two weeks), and demographic datasets (elderly and children).
  • Metrics: Top-1 accuracy, mAP, false acceptance/rejection rates.
  • Hardware: IWR6843boost mmWave radar, consumer-grade laptop for real-time testing.
• Limitations and future work:
  • Challenges in multi-user scenarios, such as close-proximity walking and signal occlusion.
  • Need for a large-scale pre-training dataset to enhance robustness.
  • Exploration of environmental variables (e.g., furniture layout) and multi-user identification.

Summary

FlowGait introduces a self-learning framework for robust, long-term gait recognition using mmWave radar, addressing real-world covariates and temporal variability. By leveraging a Transformer-based architecture and a novel step-traversal labeling algorithm, it achieves state-of-the-art accuracy across diverse datasets while mitigating performance decay during extended deployment. The system demonstrates practicality for smart home applications, with low latency and high usability confirmed through user studies. Future work will focus on multi-user scenarios, dynamic environments, and large-scale dataset development to further enhance adaptability and scalability.