Enhanced Vision-Based Person Re-Identification via Gait Recognition Using Siamese Model Based on Long Short-Term Memory Networks

Nowadays, gait recognition has become a promising technology for the long-term re-identification of individuals due to its non-invasive characteristics and potential for real-world applications. Gait is a type of biometric feature that represents the unique walking pattern of a person, including stride length, posture, and movement. This study investigates the effectiveness of gait patterns as unique characteristics for the long-term re-identification of individuals. First, the skeletal joint points of a person are extracted from video sequences using the OpenPose tool. Then, feature extraction is performed on the extracted skeletal joint points. Joint angle features are extracted to obtain the spatial representation of gait patterns. For the classification and re-identification tasks, a Siamese network with a Long Short-Term Memory (LSTM) model is trained using the sequence of extracted features as input data. To analyze the system performance, experiments are conducted using a dataset comprising video sequences of walking individuals recorded in real-world environments. The experimental results demonstrate a reasonable person re-identification accuracy of 87.96% on the test data. Furthermore, challenges such as variations in walking speed and clothing are addressed through robust feature selection techniques. This study aims to contribute to the development of practical applications in security and video surveillance systems.