Decision-level Gait Fusion for Human Identification at a Distance

Abstract

Gait Recognition is one of the latest and most attractive biometric techniques currently under research, due to its potential application in identification of individuals at a distance, unobtrusively and even using low resolution images. In this thesis a comprehensive study of the gait problem is presented, covering gait databases and the different approaches used for preprocessing, feature extraction and classification. The objective is to achieve a robust technique that performs well, independently of the input data and the many covariates that affect this behavioral biometric technique. Firstly, gait data is processed using three gait representation methods as the features sources; Accumulated Prediction Image (API) and two novel gait representations namely; Accumulated Flow Image (AFI) and Edge-Masked Active Energy Image (EMAEI). Secondly, each of these methods is tested using three matching schemes; Image Projection with Linear Discriminant Functions (LDF), Multilinear Principal Component Analysis (MPCA) with K Nearest Neighbor (KNN) classifier and the third method: MPCA+ Linear Discriminant Analysis (MPCALDA) with KNN classifier. Gait samples are fed into the MPCA and MPCALDA algorithms using a novel tensor-based form of the above-mentioned gait representations. We endup having nine recognition modules which are analyzed individually using four different experimental setups and compared to the results reported in six of the most recent papers that used the same database and the same experimental setups. Finally, decisions from the nine recognizers are fused using decision-level (majority voting) scheme. A comparison between unweighted and weighted voting schemes for final decision is also shown. The experimental results show clearly that the proposed approach outperforms the state-of-the-art gait approaches used in the literature, and reports the highest recognition rates known to the date of writing this report. As a result of the comprehensive study and extensive experiments, we conclude that model-free gait approaches, particularly spatio-temporal and energy-based methods, are the best choice in a gait recognition system used for human identification. We also note that single classifiers may not be reliable and robust to deal with gait recognition, and a fusion scheme that combines the power of each of the base classifiers is needed.