Covariate-invariant gait recognition using random subspace method and its extensions

Compared with other biometric traits like fingerprint or iris, the most significant advantage of gait is that it can be used for remote human identification without cooperation from the subjects. The technology of gait recognition may play an important role in crime prevention, law enforcement, etc. Yet the performance of automatic gait recognition may be affected by covariate factors such as speed, carrying condition, elapsed time, shoe, walking surface, clothing, camera viewpoint, video quality, etc. In this thesis, we propose a random subspace method (RSM) based classifier ensemble framework and its extensions for robust gait recognition. Covariates change the human gait appearance in different ways. For example, speed may change the appearance of human arms or legs; camera viewpoint alters the human visual appearance in a global manner; carrying condition and clothing may change the appearance of any parts of the human body (depending on what is being carried/wore). Due to the unpredictable nature of covariates, it is difficult to collect all the representative training data. We claim overfitting may be the main problem that hampers the performance of gait recognition algorithms (that rely on learning). First, for speed-invariant gait recognition, we employ a basic RSM model, which can reduce the generalisation errors by combining a large number of weak classifiers in the decision level (i.e., by using majority voting). We find that the performance of RSM decreases when the intra-class variations are large. In RSM, although weak classifiers with lower dimensionality tend to have better generalisation ability, they may have to contend with the underfitting problem if the dimensionality is too low. We thus enhance the RSM-based weak classifiers by extending RSM to multimodal-RSM. In tackling the elapsed time covariate, we use face information to enhance the RSM-based gait classifiers before the decision-level fusion. We find significant performance gain can be achieved when lower weight is assigned to the face information. We also employ a weak form of multimodal-RSM for gait recognition from low quality videos (with low resolution and low frame-rate) when other modalities are unavailable. In this case, model-based information is used to enhance the RSM-based weak classifiers. Then we point out the relationship of base classifier accuracy, classifier ensemble accuracy, and diversity among the base classifiers. By incorporating the model-based information (with lower weight) into the RSM-based weak classifiers, the diversity of the classifiers, which is positively correlated to the ensemble accuracy, can be enhanced. In contrast to multimodal systems, large intra-class variations may have a significant impact on unimodal systems. We model the effect of various unknown covariates as a partial feature corruption problem with unknown locations in the spatial domain. By making some assumptions in ideal cases analysis, we provide the theoretical basis of RSM-based classifier ensemble in the application of covariate-invariant gait recognition. However, in real cases, these assumptions may not hold precisely, and the performance may be affected when the intra-class variations are large. We propose a criterion to address this issue. That is, in the decision-level fusion stage, for a query gait with unknown covariates, we need to dynamically suppress the ratio of the false votes and the true votes before the majority voting. Two strategies are employed, i.e., local enhancing (LE) which can increase true votes, and the proposed hybrid decision-level fusion (HDF) which can decrease false votes. Based on this criterion, the proposed RSM-based HDF (RSM-HDF) framework achieves very competitive performance in tackling the covariates such as walking surface, clothing, and elapsed time, which were deemed as the open questions. The factor of camera viewpoint is different from other covariates. It alters the human appearance in a global manner. By employing unitary projection (UP), we form a new space, where the same subjects are closer from different views. However, it may also give rise to a large amount of feature distortions. We deem these distortions as the corrupted features with unknown locations in the new space (after UP), and use the RSM-HDF framework to address this issue. Robust view-invariant gait recognition can be achieved by using the UP-RSM-HDF framework. In this thesis, we propose a RSM-based classifier ensemble framework and its extensions to realise the covariate-invariant gait recognition. It is less sensitive to most of the covariate factors such as speed, shoe, carrying condition, walking surface, video quality, clothing, elapsed time, camera viewpoint, etc., and it outperforms other state-of-the-art algorithms significantly on all the major public gait databases. Specifically, our method can achieve very competitive performance against (large changes in) view, clothing, walking surface, elapsed time, etc., which were deemed as the most difficult covariate factors.