SCGait a novel method for person identification applied to legged robots.

Area of Science:

• Robotics
• Computer Vision
• Biometrics

Background:

• Legged robots commonly use speaker recognition or Ultra Wide Band (UWB) positioning for human identification and tracking.
• These methods require active user cooperation, limiting their practical applications.

Purpose of the Study:

• To develop a passive and accurate gait recognition method for legged robots.
• To create an integrated person identification-tracking system for legged robots.

Main Methods:

• Proposed a "Symmetry-encoding and pseudo-Centroid loss optimized Gait recognition method" (SCGait).
• Combined SCGait with the Yolo object detection algorithm.
• Evaluated the system on the CASIA-B dataset and custom test videos.

Main Results:

• SCGait achieved a mean test accuracy of 82.2% on the CASIA-B dataset.
• The integrated system demonstrated 91.8% identification accuracy and 36 FPS in multi-person scenes.
• Ablation studies confirmed the generalization ability of the gait encoding and loss function.

Conclusions:

• The SCGait method offers a robust solution for gait recognition in legged robots.
• The integrated system provides effective person identification and tracking for applications like companion and industrial robots.