Gait classification through CNN-based ensemble learning

Abstract

Gait is a biological characteristic for video surveillance and many other applications, which can be used to identify individuals at a large distance. In this paper, a gait classification framework based on CNN Ensemble (GCF-CNN) is proposed, which includes three modules: 1) Feature extraction and preprocessing: use random sampling with replacement strategy to generate a serial of training sets from gait silhouette images; 2) Gait models training: construct and train primary CNN classifiers using different hyper-parameters, and train them a secondary classifier to combine them; 3) Gait classification: utilize the trained two-level classifier to achieve gait classification. In addition, the proposed classification framework is evaluated on the CASIA Gait Database and OU-ISIR Gait Database. And it is demonstrated by comprehensive experiments that the proposed classification framework can achieve outstanding performance in correct classification rate with respect to several state-of-the-art methods.