Аннотация:Human gait is one of the biometric characteristics that a person can be identified by. However, the wide applicability of gait recognition in real life is prevented by a great variety of conditions that affect the gait representation, such as different viewpoints. In this work, we present a novel view resistant approach to overcome the multi-view recognition challenge. The new loss function is proposed to increase the stability of the model to view changes. Besides this, the cross-view embedding of the gait features is made to enhance their discriminant ability which improves the recognition accuracy as well. The proposed approaches show a significant gain in quality and allow to achieve the state-of-the-art accuracy on the most common benchmark and outperform the most successful model on the majority of the views and on average.