Abstract
Traditional gait-based age and gender estimation methods typically require a full gait cycle, introducing latency and sensitivity to occlusions. This paper presents a deep learning framework that achieves state-of-the-art accuracy from a single silhouette image. By employing a Phase-Aware Gait Cycle Reconstruction (PA-GCR) network to "hallucinate" a full walking cycle from a single frame, the system enables real-time estimation without waiting for a complete stride. Evaluated on the OU-MVLP dataset, the method achieves a Mean Absolute Error of 8.39 years for age and 94.27% accuracy for gender classification, processing frames in approximately 10-12ms.
Introduction
Gait analysis—identifying characteristics based on how someone walks—is a powerful biometric because it works at a distance and doesn't require the subject's cooperation. However, traditional methods rely on a full gait cycle (usually about one second of video). This creates two major hurdles for real-world deployment: latency, as systems must wait for a full stride, and occlusion, as crowded areas rarely offer unobstructed views for extended periods.
A single image lacks explicit motion information like stride length or arm swing frequency. This research addresses the challenge: Can we reconstruct the missing motion from a single snapshot well enough to make accurate predictions?
Methods
The authors propose a unified framework consisting of three main stages, designed to work across arbitrary viewing angles using a single CNN model.
1. Gait Cycle Reconstruction (PA-GCR)
The core innovation is the Phase-Aware Gait Cycle Reconstruction (PA-GCR) network. It takes a single silhouette and its estimated "phase" (the person's position in their stride) to generate a sequence of silhouettes representing a full gait cycle. Essentially, the model learns the dynamics of a full walk based on a single pose.
2. Feature Extraction (GaitSet)
The reconstructed cycle is fed into GaitSet, a state-of-the-art gait recognition network. GaitSet treats the sequence as a set of independent frames and learns a discriminative representation that is robust to different walking speeds and styles.
3. Multi-task Prediction
The final features are passed through fully connected layers to produce two outputs:
- Gender: A binary classification (Male/Female).
- Age: The model predicts a probability distribution across 100 possible age labels (0–99). This "Label Distribution Learning" helps the model handle the inherent uncertainty and biological noise in aging.
Datasets and Evaluation
The model was trained and evaluated on OU-MVLP (Osaka University - ISIR Multi-View Large Population Dataset), the world’s largest multi-view gait dataset, featuring 10,307 subjects with ages from 2 to 87 years across 14 viewing angles.
Key Insights
- Reconstruction as a Feature Booster: Even a synthetic gait cycle is more informative than a single frame. Forcing the network to reconstruct motion encourages it to extract deeper structural cues from static images.
- The Power of Symmetry: The authors used Jensen-Shannon (JS) divergence as a loss function for age estimation. Unlike KL divergence, JS divergence is symmetric, which proved more suitable for their dual-input training strategy used to ensure consistency.
- Uncertainty Modeling: By predicting an age distribution rather than a point estimate, the model expresses its confidence level, which is crucial for real-world applications like social robotics.
Results
The proposed method was tested in real-time online systems, demonstrating high efficiency and accuracy:
- Age Estimation: Achieved a Mean Absolute Error (MAE) of 8.39 years, outperforming baselines like GEINet (9.11 years).
- Gender Classification: Achieved a Correct Classification Rate (CCR) of 94.27%.
- Speed: On an NVIDIA RTX 2080 Ti, the system processes a single person in 10–12ms, maintaining a real-time 30 FPS even with multiple people in the frame.
- View Robustness: Performance is most accurate at side views (75°–90°) but remains stable across all 14 angles.
Conclusion
This work demonstrates that gait-based biometrics can be effectively extracted from a single image through cycle reconstruction. Future work will explore continuous age distributions and temporal fusion of multiple frames to further enhance stability and accuracy in "in-the-wild" environments.
Summary updated