Multi-Adversarial In-Car Activity Recognition Using RFIDs

In-car human activity recognition opens a new opportunity toward intelligent driving behavior detection and touchless human-car interaction in future smart transportation systems. Among the multiple sensing technologies such as camera and WiFi, radio frequency identification (RFID) exhibits unique advantages given its low cost, few privacy concerns, and easy deployment. Existing RFID-based approaches are mostly confined to working well in stable indoor space. The in-car recognition, however, is much more complex with multiple impact factors. On one hand, the fast-changing external environment such as moving cars and pedestrian may have interference on the sensing signal. On the other hand, different cars and human subjects vary much in space configurations and motion habits, respectively, which inevitably introduce activity independent signal patterns. The included extraneous information can pollute the activity related features, rendering a poor performance for existing solutions. In this paper, we present RF-CAR, an RFID-based in-car activity recognition framework that is able to remove activity irrelevant features from RF signals and only retain the activity-specific features. Our framework integrates a deep learning architecture and a novel multi-adversarial domain adaptation network for training and prediction. Our extensive experiments further demonstrate the superiority of RF-CAR, even in an untrained environment