Research on Closed-Loop Anomaly Detection Based on Kalman Filter During Driving and Its Impact on Traffic Safety

To enhance driving safety in longitudinal car-following scenarios, this study proposes a closed-loop anomaly detection and safety-mode switching framework designed to mitigate risks caused by sensor noise and potential sensor faults. The framework integrates Kalman filtering, χ2-based statistical consistency checks, and a cumulative-sum (CUSUM) change-point detector. It incorporates stochastic braking behaviors of the lead vehicle, ACC-based following control of the ego vehicle, measurement noise and injected faults (bias, drift, scale errors, and packet drops), as well as a safety mode triggered upon anomaly detection—featuring an increased desired time headway, adjusted control gains, and modified braking limits. Using minimum gap, time-to-collision (TTC), time-headway (THW), low-TTC/low-THW exposure ratios, false-alarm and missed-detection rates, and detection delay as quantitative evaluation metrics, we compare a baseline “no-detection” scheme with the proposed “closed-loop detection (KF-based)” approach. Experimental results demonstrate that the Kalman filter significantly improves residual distinguishability and suppresses noise-induced false alarms, while CUSUM offers heightened sensitivity to small but persistent anomalies. In drifting range-sensor fault scenarios, enabling the detection framework substantially reduces hazardous headway exposure and effectively prevents collisions under multi-segment stochastic braking conditions of the lead vehicle.