Gait-based human classification using spectrograms analysis and deep learning

Gait is a routine activity performed daily by every able-bodied individual. Recent studies have demonstrated its potential as a reliable factor in behavioral biometrics for user recognition, offering high accuracy. This study presents a novel approach to human gait analysis leveraging spectrogram representations of accelerometer signals combined with deep learning techniques. In proposed methods time-series accelerometer data is transformed into spectrograms using the Short-Time Fourier Transform (STFT). This conversion enables the utilization of convolutional neural networks (CNNs), which are highly effective in image classification tasks, to identify distinctive gait patterns. The architecture of the neural network was carefully optimized through a series of empirical experiments, adjusting the number and order of layers to achieve the best classification performance. The effectiveness of the proposed solution was validated on both a custom dataset and a publicly available benchmark dataset. The proposed CNN-based approach achieved a prediction accuracy of 92.03%, outperforming traditional models based on raw sensor data and demonstrating competitive performance relative to state-of-the-art methods, such as ResNet-based CNNs, GRU networks, and the M-GaitFormer architecture. These results confirm the advantages of time-frequency feature extraction for gait recognition. This work demonstrates the potential of using everyday mobile devices equipped with accelerometers for biometric authentication based on behavioral traits. Future research directions include extending the model to multi-sensor fusion (e.g., incorporating gyroscope data), enhancing robustness across diverse walking conditions, and developing efficient implementations for real-time gait recognition on mobile platforms.