Domain-Enhanced Dual-Branch Model for Efficient and Interpretable Accident Anticipation

2025-07-16

AI Summary

To address the low accuracy, high computational overhead, and poor interpretability of real-time traffic accident prediction in autonomous driving, this paper proposes a domain-enhanced dual-branch multimodal fusion model. The model separately processes driving videos (using Long-CLIP) and structured accident texts (semantically parsed via GPT-4o prompt engineering), enabling efficient cross-modal collaborative modeling through cross-modal feature alignment and a lightweight fusion mechanism. Innovatively, it incorporates traffic-domain knowledge-constrained prompt templates and a hierarchical attention aggregation strategy, significantly reducing inference latency while preserving model interpretability. Evaluated on three major benchmarks—DAD, CCD, and A3D—the method achieves state-of-the-art performance with fewer parameters, improving average accuracy by 4.2% and F1-score by 5.7%. This work establishes a new paradigm for real-time, trustworthy accident early warning systems.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multimodal Learning
• Application Domains: Mobility, Driving & Flight

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Developing precise and computationally efficient traffic accident anticipation system is crucial for contemporary autonomous driving technologies, enabling timely intervention and loss prevention. In this paper, we propose an accident anticipation framework employing a dual-branch architecture that effectively integrates visual information from dashcam videos with structured textual data derived from accident reports. Furthermore, we introduce a feature aggregation method that facilitates seamless integration of multimodal inputs through large models (GPT-4o, Long-CLIP), complemented by targeted prompt engineering strategies to produce actionable feedback and standardized accident archives. Comprehensive evaluations conducted on benchmark datasets (DAD, CCD, and A3D) validate the superior predictive accuracy, enhanced responsiveness, reduced computational overhead, and improved interpretability of our approach, thus establishing a new benchmark for state-of-the-art performance in traffic accident anticipation.

Problem

• Develop efficient accident anticipation for autonomous driving
• Integrate visual and textual data for improved prediction
• Enhance interpretability and reduce computational overhead

Innovation

• Dual-branch architecture integrates visual and textual data
• Feature aggregation method uses GPT-4o and Long-CLIP
• Prompt engineering enhances feedback and archives