Recognition of surrounding vehicle anomalous intent in the vehicle-road-cloud environment for autonomous driving

This study addresses the challenge of recognizing anomalous intent in surrounding vehicles for autonomous driving, proposing the ADID+ framework. The framework employs dynamic frequency band division and dual-stream attention mechanisms to integrate vehicle-road-cloud collaborative data, achieving high-precision, low-latency anomaly detection. Evaluated on 12,586 multi-modal annotated samples covering six anomaly types (such as malicious cut-ins and illegal overtaking), ADID+ demonstrated an F1-score improvement of 6.5 %-18.6 % over mainstream methods. It also reduced the average detection delay to 132 ms, thereby meeting active safety system requirements. This approach overcomes the bottlenecks of traditional time-domain models through frequency-domain feature enhancement, revealing distinct spectral representation patterns for abnormal behaviors like malicious cut-ins (5.2 Hz) and illegal overtaking (3.5/4.2 Hz). Even in scenarios with missing multi-modal data or extreme parameters, the framework maintains a detection accuracy above 72 % and a false alarm rate below 3.7 %. Limitations of this study include not accounting for communication latency jitter and online frequency band optimization. Future work will focus on developing dynamic latency compensation mechanisms and lightweight operators to enhance robustness under non-ideal conditions.