TraffiPrompt: Network-aware multimodal prompt learning with motion-aware temporal modeling for driving video anomaly detection

Driving video anomaly detection is crucial for intelligent transportation systems, yet remains challenging because raw video streams are high-dimensional, expensive to transmit, and hard to process under real-time latency and bandwidth constraints. To address this issue, we propose TraffiPrompt, a network-aware multimodal framework that combines normality-guided prompt learning with motion-aware temporal modeling under coarse temporal supervision. The framework first builds a compact multi-source visual representation for feature-level transmission, then enhances it with a motion-aware branch and ego-motion compensation to capture dynamic irregularities. It further introduces prompt-guided semantic alignment with anomaly-category prompts, normal prompts, traffic-context auxiliary prompts, and video captions to strengthen cross-modal interaction. We also adopt stagewise self-training with pseudo-label refinement to improve temporal localization. Experiments on DoTA show that TraffiPrompt achieves over 0.90 AUC while maintaining strong recall at tIoU thresholds of 0.30 and 0.50. Additional efficiency and delay analyses further clarify the practical characteristics of the proposed feature-level temporal head.