An Efficient Driver Anomaly State Detection Approach Based on End-Cloud Integration and Unsupervised Learning

For current autonomous vehicles, real-time monitoring of driver states and prompt identification of abnormal behaviors during the driving process are of paramount importance for safety. This paper proposes an innovative edge-cloud fusion driver anomaly detection system to address the issues of unstable anomaly detection performance and high computational demands of existing driver anomaly detection systems. Our system achieves instantaneous anomaly detection on the edge side while transmitting crucial facial features of the driver to the cloud for long-term anomaly detection. The proposed method employs unsupervised learning techniques to identify challenging and ill-defined abnormal patterns. Through deployment testing on actual hardware platform, our edge-cloud detection system achieved a latency less than 100ms.