Safety analysis of autonomous vehicles based on target detection error

Abstract

Connected and automated vehicles (CAVs) rely on their perception systems to detect traffic objects, with the uncertainty in detection results significantly influencing the safety of their decision-making and control mechanisms. This paper introduces a safety potential field for CAVs that accounts for target detection errors. Initially, the paper categorizes errors arising from target detection into classification, labelling, and positioning categories. Subsequently, an elliptical model-based safety potential field is developed, incorporating potential field line optimization using safety thresholds and lane lines. This approach facilitates the determination of critical values and safety distribution for the potential field. The paper then proceeds with coefficient calibration and experimental analysis to validate the reliability of the proposed model. Findings indicate that as target detection errors increasingly manifest, the safety potential field area for CAVs becomes more restrictive, enhancing the field's sensitivity to these errors. The critical safety value for CAVs is maintained within the range of [0 m, 7 m], providing a stable basis for decision-making and control. Additionally, the safety value for CAVs falls between [15, 25], favouring the improvement of safety gradient distribution under the calibrated safety potential field values.