Abstract
In this paper, we discuss a gait representation based on the width of silhouette in terms of discriminative power and robustness against the noise in silhouette image for gait recognition. Its sensitivity to the noise in silhouette image are rigorously analyzed using probabilistic noisy silhouette model. In addition, we develop a gait recognition system using width representation and identify subjects using the decision level fusion based on majority voting. Experiments on CASIA gait dataset A and the SOTON gait database demonstrate the recognition performance with respect to the noise level added to the silhouette image.
Introduction
Gait recognition has emerged as a powerful biometric modality due to its non-intrusive nature and ability to operate at significant distances where other biometrics, like facial features or fingerprints, are often obscured. Among various approaches, "model-free" methods are particularly popular because they bypass the computational complexity of 3D human modeling. A common feature used in these methods is the width vector, which represents the horizontal distance between the leftmost and rightmost pixels of a person's silhouette at each row.
While computationally efficient, width-based features are often criticized for their perceived vulnerability to environmental noise. This research provides a rigorous mathematical analysis of this vulnerability, identifying exactly how and why noise in low-quality surveillance footage degrades recognition performance.
Width-Based Gait Representation
The core feature of this system is the width vector extracted from binary silhouettes. The gait feature is defined as the arithmetic mean of these width vectors across a full gait cycle. To identify these cycles, the system detects "valleys" in the norm of the width vector, which correspond to the rest positions in a stride.
Sensitivity Analysis and the "Narrowness Penalty"
To analyze robustness, the researchers employed a probabilistic noise model where background pixels can be flipped to foreground (salt noise) and vice versa (pepper noise) with a probability $\epsilon$.
A critical finding of the mathematical derivation is the "Narrowness Penalty." Because the width vector relies solely on the two extreme boundary pixels in any given row, it lacks internal redundancy. The analysis shows that the expected error from noise increases exponentially as the actual width of the body part decreases. This makes narrow regions, such as the neck and head, significantly more sensitive to noise than the torso.
Experimental Results
The theoretical analysis was validated using the CASIA Gait Dataset A and the SOTON Gait Database. The results revealed a "performance cliff" where accuracy degrades catastrophically even with minimal noise:
- Baseline Performance: In noise-free conditions, the system achieved a Correct Classification Rate (CCR) of 80.53% on the SOTON database and 71.25% on CASIA A.
- Impact of Noise: When the noise probability ($\epsilon$) was increased to just 0.005 (affecting only 0.5% of pixels), the CCR on the SOTON database plummeted to 1.11%.
- Error Characteristics: As noise increased, the False Acceptance Rate (FAR) rose significantly, indicating that noise-distorted width vectors frequently caused the system to incorrectly match subjects to the wrong identities.
Conclusion
This study demonstrates that while width-based gait recognition is efficient, it is highly sensitive to boundary noise, particularly in narrow body regions. The lack of redundancy in boundary-only features leads to a rapid failure in real-world conditions with even minor pixel noise. Future research should focus on projection-based gait representations, which sum pixel densities across rows or columns to provide better noise-averaging and robustness for practical surveillance applications.
Summary updated