Research Of Gait Recognition Based On Self-Training

Biometrics is a technology that uses the physiological or behavioral characteristics of people to authenticate their identities. Gait recognition is a new kind of biometrics recognition method that aims to identify an individual based on his/her style of walking. Compared with other biometric features such as fingerprint, face and iris, gait has some unique advantages. Gait is effective for recognition at a distance and it is non-contact, difficult to disguise or conceal. Therefore, gait is the most potential biometric feature when identifying individuals at a distance. With the growing demand for a full range of visual surveillance and monitoring systems in security-sensitive environments, gait recognition has received great attention from computer vision researchers because of its potential application in visual monitoring field.Existing gait recognition algorithms are almost based on supervised learning that uses only labeled samples to train learners. Obviously, it is difficult for learners to have strong generalization ability if they are trained with a limited number of labeled samples. On the other hand, it is a wasteful if we just use a limited number of labeled samples without using a large amount of unlabeled samples. Thus, how to make full use of the large amount of unlabeled samples to optimize the performance of learners while the labeled samples are limited has been a hot spot in machine learning field. Semi-supervised learning addresses this problem by using a large amount of unlabeled data, together with the labeled data, to build better classifiers without manual intervention. Considering the problem above, we proposed a semi-supervised gait recognition algorithm based on self-training. The main contributions of this thesis are as follows:We proposed a semi-supervised gait recognition algorithm based on self-training, which uses a large amount of unlabeled gait sequences, together with a small number of labeled gait sequences to build better classifiers. Firstly, PCA training is performed on the small number of labeled gait sequences to obtain feature transformation matrix and compute the gait features of sequences (i.e., trained database), and then we use the trained database to classify the unlabeled sequences. Typically, for each class, the most confident unlabeled sequence, together with its predicted label, is added to the labeled training set. Then PCA training is rerun and the procedure repeats for a given number of times or the trained dataset keeps unchangeable. Experimental results show that the proposed algorithm has an encouraging recognition performance even when each class has only one labeled sequence at the beginning.Construct a new gait dataset. For verifying the effectiveness of the semi-supervised gait recognition algorithm, each class must have enough sequences. However, all the public gait datasets do not satisfy this condition. Thus, we construct a new gait database in order to verify the effectiveness of the algorithm. The biggest characteristic of our dataset is each class has40sequences, which makes the experimental results persuasive. The experimental results on this database show that the proposed algorithm can not only make full use of the large amount of unlabeled sequences to reduce the cost of labeling samples, but also make the gait recognition system optimize itself to satisfying performance after the self-training process.Although the experimental results prove the effectiveness of the proposed algorithm, it still has some shortages, such as the fixed view angle and the simple moving background. Thus, the main work is to improve the robustness of the algorithm in the future. For example, adopting multi-camera technology to recover the three-dimensional data of human gait in order to make the algorithm unrelated to view angles; developing moving detection algorithms that is more robust to accomplish the silhouette separation better under true environment; fusing different gait features to improve the robustness of algorithm.