Gait Recognition from Surveillance Footage Using Deep Learning

Gait recognition is a biometric method that identifies individuals based on their unique walking patterns. This study aims to develop a gait recognition system from surveillance scenarios using public gait datasets. using a deep learning approach based on Long Short-Term Memory (LSTM) networks. Two datasets are utilized: OU-ISIR MVLP and CASIA-B, each offering different characteristics in terms of subject diversity, viewpoint variations, and data collection conditions. The system includes stages of silhouette image preprocessing, feature extraction, model training, and testing through a web-based interface. Training results indicate that the LSTM model can recognize gait patterns with varying accuracy depending on the dataset and viewing angle. In tests using the CASIA-B dataset, the model achieved correct predictions with a maximum confidence score of 0.7412 at certain angles, while the OU-ISIR MVLP dataset produced less consistent predictions with lower average confidence scores. The system was implemented using a web-based architecture built with Next.js and Flask. This research provides an early overview of the potential and limitations of applying LSTM models in silhouette-based gait recognition.