Efficient Neuro-Symbolic Predictive Modeling for Near-Miss Accident Detection in High-Velocity Video Streams

Phuong Thi Mai Nguyen, Minh-Son Dao, Swe Nwe Nwe Htun, Koji Zettsu

Abstract: Real-time detection of near-miss accidents from high-velocity dashcam video streams is a critical Big Data challenge for intelligent transportation systems, demanding both predictive capability for proactive safety and extreme computational efficiency for edge deployment. Traditional deep learning models and large multimodal architectures (MLLMs) are often too resource-intensive and lack the structured output necessary for interpretable, real-time forecasting. To address this, we propose an Efficient Neuro-Symbolic Predictive Modeling (NSPL) framework. Our model learns to compress unstructured video data into a sequence of discrete, interpretable symbolic templates (e.g., [bicycle, cut-in, left, intersection, suddenly]), achieving significant reduction in data dimensionality. A novel transformerbased architecture then leverages these structured sequences to predict future event templates within a 0-5 second horizon, enabling the forecasting of near-miss events before they fully unfold. Empirically validated on a real-world driving dataset, our model achieves superior predictive accuracy while operating at 30 FPS on a Raspberry Pi with less than 1 GB memory, significantly outperforming MLLM baselines. This work demonstrates a scalable paradigm for predictive analytics on video streams, combining the efficiency of lightweight deep learning with the interpretability and forecasting power of symbolic reasoning for safety-critical transportation applications.

External IDs:dblp:conf/bigdataconf/NguyenDHZ25