A model-free gait recognition approach for person identification with covariates

Abstract

In the present scenario, the identification of abnormal activity in the crowd plays a vital role in the detection of criminal and mental activities. Gait can be utilized for person authentication and can be used in surveillance where a large number of people pass through. Gait recognition is an evolving biometric technology, which involves people being recognized purely through walking manner. It is an interesting research topic as gait can be obtained without a person's cooperation, and gait has a unique nature. At present, the methods being used for gait recognition can be broadly divided into two categories, model-based, and model-free approaches. The model-based approach requires prior modeling to generate gait features, which requires high computational costs, whereas the model-free approach directly extracts features from silhouette frames. The main challenge faced by a model-free approach is that it is sensitive to covariate conditions such as clothing, carrying objects, and walking surfaces. This paper presents a modelfree approach that is very effective at handling known and unknown covariates. In this paper, two popular methods, namely, Convolutional Neural Network (CNN) and discriminative feature-based classification methods, are used for gait recognition. The CNN-based method is used for known covariates, and the discriminative feature-based classification method is used for unknown covariate conditions. The Chinese Academy of Sciences (CASIA) Gait Database B from the Center for Biometrics and Security Research (CBSR) is used to train the models and their performance is evaluated in terms of Accuracy, Precision, Recall, and F1-Scores.