2025, Vol. 6, Issue 1, Part D

Gait recognition is a critical biometric approach used for identifying individuals based on their walking patterns. This paper presents an efficient gait recognition framework capable of addressing both known and unknown covariate conditions using Gait Energy Images (GEI). The proposed methodology integrates Convolutional Neural Networks (CNN) for known covariate conditions, where models are trained and tested on similar conditions such as normal walking. For unknown covariate conditions, which involve variations such as clothing, carried items like bags, and altered walking styles, features are extracted using Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and Haralick descriptors. These extracted features are then dimensionally reduced using Fisher Linear Discriminant Analysis (FLDA) and classified using Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and Random Forest classifiers.

The framework is implemented through a series of well-defined modules: the dataset upload module allows integration of the GAITB dataset; preprocessing ensures normalization and grayscale transformation of images for effective training; CNN is employed to train a model for recognizing person IDs under known covariate conditions; and advanced feature extraction techniques followed by FLDA are used to reduce feature sizes for subsequent training with SVM, MLP, and Random Forest classifiers. Performance metrics such as accuracy, precision, recall, and F1-score are utilized to evaluate the models. An additional comparison graph visually represents the accuracy of these classifiers, highlighting their effectiveness under different covariate conditions.