An Improved Variational Auto-Encoder with Reverse Supervision for the Obstacles Recognition of UGVs

The obstacles detection plays an important role in the field of unmanned ground vehicle (UGV). This paper proposes a semi-supervised learning model with reverse supervision based on Variational Auto-Encoder (VAE) to recognize the terrain obstacles of UGVs. The proposed model compresses terrain data to latent space and casts the abnormal observations to invalid white noise in order to perform more accurate fitting on marginal likelihood of normal observations. In addition, the proposed model adopts the convolutional layer instead of fully connected layer of VAE to extract data features. Gaussian Mixed Model (GMM) is used to fit the latent distribution of normal terrain data. The improved VAE could learn the actual potential distribution of target data with the reverse supervision of abnormal data, it can achieve better performance in generating ability and discriminating ability compared with existing generative models. The superiority and effectiveness of the proposed model are illustrated and validated by an application in the shooting range of UGVs. Besides, Cthe proposed model has the promising potential for some other applications, it can be used for military operations, robot rescue, and terrain exploration in dangerous environment, etc.