HUMAN GAIT RECOGNITION FOR BIOMETRIC IDENTIFICATION USING DEEP NEURAL NETWORK

Abstract

Gait recognition has become an increasingly valuable solution for the non-invasive identification of any individual, particularly in the domains of security and healthcare. Despite its advantages, the variations in viewing angles, clothing, and carried objects continue to hinder the robustness of the model in real-world scenarios. Therefore, in this study, we present an automated gait recognition system that addresses these limitations through a series of deep learning-based novel enhancements. Firstly, a sequential contrast enhancement technique is used to improve the visual quality of the image, followed by another technique named, mean-based contrast enhancement. The resultant input channels are then further fused into a unified representation using a streamlined integration method. In addition, the Bayesian hyperparameter optimization technique is employed to optimize the performance of the model. The proposed system also incorporates a Parallel Feature Extraction and Fusion (PFxF) method to enhance the spatial and temporal information, along with an improved Scale and Rotation Invariance Optimization (SRIO) technique for a robust selection of features. Furthermore, we evaluated multiple models to perform classification tasks, including L-SVM, C-SVM, Q-SVM, BG Tree, and MD-NN models on the CASIA-B dataset. The performance of each classifier was assessed using standard metrics such as True Positive Rate (TPR), error rate, precision, recall, accuracy, and Area Under the Curve (AUC). The results show that the BG Tree classifier consistently outperformed its counterparts across all covariate conditions. It recorded the highest TPR of 98.40%, 96.20%, and 95.80% for the BAG, COAT, and NORMAL classes, respectively. Moreover, it achieved the leading precision of 97.57%, recall of 97.57%, accuracy of 97.60%, and AUC of 97.67%, while maintaining the lowest error rate of 2.40%. These outcomes demonstrate the reliability and adaptability of our proposed approach in diverse conditions, making it an effective solution for future gait-based biometric applications