Real-Time Multi-view Face Mask Detector on Edge Device for Supporting Service Robots in the COVID-19 Pandemic

摘要

The COVID-19 pandemic requires everyone to wear a face mask in public areas. This situation expands the ability of a service robot to have a masked face recognition system. The challenge is detecting multi-view faces. Previous works encountered this problem and tended to be slow when implemented in practical applications. This paper proposes a real-time multi-view face mask detector with two main modules: face detection and face mask classification. The proposed architecture emphasizes light and robust feature extraction. The two-stage network makes it easy to focus on discriminating features on the facial area. The detector filters non-faces at the face detection stage and then classifies the facial regions into two categories. Both models were trained and tested on the benchmark datasets. As a result, the proposed detector obtains high performance with competitive accuracy from competitors. It can run 20.60 frames per second when working in real-time on Jetson Nano.