Domain-Robust Pedestrian-View Intersection Classification

Pedestrian-view intersection classification is one of the key components in building navigation systems for small autonomous vehicles maneuvering on pedestrian paths. However, until recently, such systems operate only in one domain, such as only in outdoor environments. In this work, we propose to train a model that is robust across multiple domains, i.e., indoor and outdoor. In order to achieve such robustness, the network is trained to extract only intersection classification related features while not extracting the domain information. Such objective is achieved by using two-branched network: intersection classifier and domain predictor. The network is then encouraged to fail predicting the domain while successfully predict the intersection categories. Extensive experiments demonstrate the superiority of our proposed model against the baseline trained using the dataset containing samples from both domains.