What Is Gait Recognition?

Gait recognition is a biometric identification method that analyzes the distinctive patterns in a person's walking motion to verify or identify individuals. As a behavioral biometric, it operates at a distance without requiring cooperation from the subject and does not depend on contact with a sensor, distinguishing it from fingerprint, iris, and facial recognition systems. The characteristic patterns of a person's gait arise from the interaction of anatomy, neuromuscular coordination, and habitual movement, and are sufficiently individual that they can serve as a stable identifier across time. Gait recognition has roots in biomechanics, computer vision, and pattern recognition, and has grown into an active research area within IEEE's biometrics and signal processing communities since the 1990s.

Input signals for gait recognition fall into two broad categories: video or depth imagery captured by cameras, and inertial measurements from accelerometers and gyroscopes embedded in wearable devices or smartphones. Each modality carries different trade-offs between coverage area, data quality, and deployment constraints, which has driven the development of parallel methodological branches.

Appearance-Based and Model-Based Approaches

Traditional gait recognition methods divide into appearance-based and model-based categories. Appearance-based methods represent the walking figure as a sequence of silhouettes or energy images extracted from video frames, compressing the temporal sequence into a single Gait Energy Image (GEI) that can be fed to a classifier. These holistic representations are computationally efficient and work on low-resolution footage, making them suited to wide-area surveillance. Model-based methods instead fit a parameterized body model, typically a stick figure or articulated skeleton, to each frame and track the joint angles and limb trajectories over time. Model-based representations are more interpretable and scale-invariant but require higher image resolution and more computation. A comprehensive survey of deep gait recognition methods identifies five persistent challenges shared by both families: speed variation, clothing and carried objects that alter the body silhouette, partial occlusion, camera viewpoint changes, and adversarial spoofing attempts.

Deep Learning Methods

Since 2015, convolutional neural networks and recurrent architectures have displaced hand-crafted feature pipelines as the dominant approach to gait recognition. CNNs extract spatial features from silhouette or skeleton sequences, while long short-term memory (LSTM) networks and 3D convolutions capture the temporal dynamics of the walking cycle. Graph neural networks that model the human skeleton as a graph of joints have shown strong performance on skeleton-based gait data. A hybrid CNN-RNN architecture for inertial-sensor gait recognition achieved above 93% accuracy in both identification and authentication tasks on smartphone accelerometer data from 118 subjects walking in unconstrained conditions. Generative adversarial networks have also been applied to synthesize gait sequences under unseen viewpoints and covariate conditions, augmenting training datasets and improving cross-condition generalization. IEEE research on biometric identification through deep gait analysis documents continued progress on feature learning from both silhouette and inertial modalities.

Covariates and Robustness

Real-world deployment of gait recognition requires robustness to covariates, factors external to the individual's inherent gait that alter the observed motion signature. Wearing a coat, carrying a bag, changing footwear, or walking on an uneven surface all shift the gait features in ways that can degrade recognition accuracy. Covariate-aware training strategies, domain adaptation, and multi-modal fusion that combines video with inertial data have been shown to reduce covariate sensitivity. Aging is a slower covariate that degrades long-term recognition performance and requires periodic re-enrollment.

Applications

Gait recognition has applications in a wide range of disciplines, including:

• Access control and identity verification in security installations without requiring subject cooperation
• Forensic analysis using surveillance video for post-event identification
• Clinical gait assessment for monitoring neurological conditions such as Parkinson's disease and stroke rehabilitation
• Continuous authentication on mobile devices using embedded inertial sensors
• Crowd monitoring and soft biometric profiling in public safety systems