A continuous similarity recurrence plot fusion approach for collision instant identification in autonomous vehicle accidents

In traffic accidents involving autonomous vehicles, the participation of multiple agents complicates the determination of liability, where the precise identification of the collision instant plays a decisive role in attributing responsibility. To address the limitations of existing methods that rely on manually set thresholds and exhibit insufficient robustness, this study proposes a novel method termed the Continuous Similarity Recurrence Plot Fusion with Comprehensive Collision Index (CSRP-CCI). This approach achieves automatic and precise identification of the collision instant by constructing a continuous similarity recurrence plot, implementing a multi-component fusion strategy of triaxial acceleration data, and defining a comprehensive collision indicator. Validated on six real-world autonomous vehicle accident cases, the study employed three different phase-space reconstruction parameter determination methods—namely AC + FNN, MI + FNN, and C-C—and compared them against the standard threshold method based on EDR/DSSAD data. The results demonstrated that the proposed CSRP-CCI method achieved detection deviations of fewer than 12 sampling points (<120 ms) across all cases, showing improvement over traditional threshold-based methods. Among the three parameter determination approaches, the C-C method exhibited the best stability and achieved earlier identification of the collision instant than the reference instant calibrated by forensic investigators in Case 3 and Case 6. Its results showed high consistency with those of the standard threshold method, indicating that it more closely approximated the initial physical contact instant of the collision. This study provides a highly accurate, interpretable, and robust technical solution for the forensic investigation of autonomous vehicle accidents.