Multi-View Gait Recognition Based on Generative Adversarial Network

摘要

We proposed a new multi-view gait recognition model based on the generative adversarial network (GAN). Gait recognition has great advantages in long-distance human body recognition, but it is also very challenging. The appearance-based recognition method performs gait recognition based on the outer contour of the human body, but the appearance may change due to factors such as clothing and carrying objects. The method based on view transformation is recognized by converting the gait image to a 90° normal state. However, with the increase of the gait viewing angle difference, a larger transformation error will be produced, which will result in the loss of gait features, and a good recognition effect will not be achieved. In this paper, a GAN was used to convert the gait image of any states and views to the normal state of 54°, 90°, 126°, and realized the gait recognition based on the extracted invariant features. In order to reduce the loss of characteristic information during the view transformation, this paper introduced a residual structure in the network, which could retain more identity information while realizing the view transformation. Finally, the fusion model was used to fuse the recognition results of the three views in the recognition and decision-making stage. In order to verify the effectiveness of the model in this paper, we used the CASIA-B gait dataset to evaluate the performance of the model. The experimental results showed that the model in this paper had achieved a good recognition effect in the sequence of bag and coat. Compared with advanced networks such as SPAE, GaitGANV1 and GaitGANv2, the performance of this model was better, and it was more robust to changes in viewing and clothing.