Edge-Optimized Transformer for IoT-Enabled Vulnerable Road User Safety in Smart Transportation Systems

초록

The accurate detection of Vulnerable Road Users (VRUs) is necessary to ensure safe mobility in smart city environments. However, existing datasets focus on general pedestrian detection, which covers a limited VRU subgroup. Specifically, VRUs exhibit unique mobility patterns that require context-related awareness as well as adaptive control to ensure their own protection. Considering constraints of datasets available as well as models, we develop a specialized dataset tailored for six types of VRUs. Utilizing our dataset, we propose an IoT-enabled, edge-optimized Transformer framework for deployment in intelligent pedestrian signaling systems within connected transportation networks. Additionally, we apply post-training weight clustering and pruning and achieve a significant reduction in parameters compared to the baseline with a 94% mAP. Furthermore, we provide VRU prioritization and adaptive signal control protocols, considering the VRU's need for their safety. We aim to make the crosswalks smarter and support real-time communication between vehicles and urban traffic management units. Extensive evaluation across varied lighting and weather scenarios confirms the robustness of our model and suitability for low-latency within a smart transportation infrastructure. This study provides directions for future studies to focus on making the pedestrian signal intelligent and contribute to smart cities.

키워드