The safe operation of automatic driving system in complex road environment requires the cooperation of many kinds of sensors. However, the sensor will produce wrong data due to its own fault, bad weather and algorithm problems during operation, which will lead to the system misjudging obstacles, thus causing safety problems. This paper analyzes the inherent limitations of cameras, lidar, millimeter-wave radar and other sensors, and puts forward effective methods of anomaly detection and the application of multi-sensor fusion in automatic driving safety early warning, in order to provide technical reference for improving the safety of automatic driving.