Intelligent Traffic Video Analytics: Advancing Crash Event Understanding through State-Space Model Optimization, and Hybrid Architecture Development

This dissertation investigates the application of advanced artificial intelligence models to traffic surveillance footage analysis, addressing critical challenges in crash detection, narrative generation, and precise temporal localization. Through two interconnected research papers, this work bridges the gap between state-of-the-art AI capabilities and practical deployment in traffic safety applications. The first paper (Chapter 3) addresses the computational efficiency challenges of advanced sequence models by developing a novel unstructured pruning framework for Mamba state-space models. This approach achieves significant parameter reduction (up to 70%) while maintaining 91-97% of original performance across diverse benchmarks, making powerful sequence modeling viable for resource-constrained traffic monitoring systems. The second paper (Chapter 4) introduces HybridMamba, an innovative architecture combining vision transformers with Mamba-based temporal modeling to achieve superior fine-grained temporal localization of crash events. Evaluated on Iowa Department of Transportation footage, HybridMamba demonstrates a mean absolute error of 1.50 seconds in identifying crash times, outperforming existing video-language models by up to 3.95 seconds while using fewer parameters. Collectively, this research makes several contributions: (1) an efficient pruning framework enabling deployment of powerful sequence models in resource-constrained environments; (2) a specialized architecture for precise crash time localization; and (3) empirical evidence from extensive experiments demonstrating the effectiveness of the proposed methods. The findings have direct applications in emergency response systems, traffic management, and transportation policy, with potential to significantly enhance road safety through improved incident detection and analysis.

Committee: Anuj Sharma (co-major professor), Tichakorn Wongpiromsarn (co-major professor), Jack Lutz, Adarsh Krishnamurthy, and Soumik Sarkar