A Deep Learning Approach to Person Identification via Gait Analysis in Adverse Covariate Scenarios*

In computer vision, the recognition of the person across multiple cameras is considered an immensely complex process, including the recognition of the same individual in various scenarios. Human gait recognition is a biometric approach that depends on walking patterns to recognize an individual from a long distance. However, gait recognition systems face several challenges, known as covariate factors, that decrease the accuracy of the system. These covariate factors include wearing conditions, view angles, walking speeds, carrying conditions, and low light. To handle covariate factors, we developed a model that starts with a fine-tuned pose estimation system to extract joint information. Gamma correction is applied to handle low-light conditions, then uses CNN and LSTM networks to process spatial and temporal gait features. The features are classified using C-SVM, Naïve Bayes, BG Trees, and XGBoost. The model is tested on CASIA B, CASIA C, and OUMVLP-Pose datasets. The model achieves 98.2% accuracy on CASIA A, 97.2% on CASIA B, and 97% on OUMVLP-POSE.