Hybrid Skeleton-Based Motion Templates for Cross-View and Appearance-Robust Gait Recognition

Area of Science:

• Computer Vision
• Biometrics
• Human Motion Analysis

Background:

• Silhouette-based gait recognition (e.g., Gait Energy Image) excels in controlled settings but falters with appearance variations.
• Skeleton-based methods offer interpretability but are susceptible to pose-estimation inaccuracies.

Purpose of the Study:

• To introduce novel compact 2D skeletal descriptors, Gait Skeleton Images (GSIs), for enhanced gait recognition.
• To develop a unified pipeline for processing 3D joint trajectories into static templates compatible with 2D CNNs.
• To evaluate the robustness and performance of GSIs against appearance and viewpoint changes.

Main Methods:

• Encoding 3D joint trajectories into line-based and joint-based 2D skeletal templates (GSIs).
• Implementing a processing pipeline involving topology normalization, view alignment, projection, and rendering.
• Analyzing core design factors on the GRIDDS dataset and evaluating performance on the CASIA-B dataset.

Main Results:

• GSI variants demonstrate reduced performance degradation under varying appearance conditions (viewpoint, clothing, carried objects) compared to Gait Energy Images.
• GSIs exhibit greater stability under severe cross-view gait recognition scenarios.
• While Gait Energy Images achieve higher same-view accuracy, GSIs offer superior robustness.

Conclusions:

• Compact skeletal templates like GSIs can effectively complement appearance-based gait recognition methods.
• The proposed Gait Skeleton Images show promise for improving the robustness of human identification systems.
• Future advancements in 3D human pose estimation could further enhance the utility of skeletal-template-based gait recognition.