Night-time traffic accidents pose significant challenges to road safety, making rapid detection and response crucial. In this research, we propose an efficient system for near real-time traffic accident detection using deep learning techniques on edge devices. Specifically, we leverage the attention-based Swin-Tiny Transformer model for accident classification on traffic surveillance data. We handle the dataset imbalance by utilizing appropriate loss function while training the classifier. For deployment, we utilize the Raspberry Pi 4 as an edge device, providing an affordable and efficient solution for near real-time accident detection at the point of occurrence. The system’s ability to process images locally allows for quicker response times, crucial for dispatching emergency services. Our system achieves an accuracy of 74.41% in challenging night surveillance scenarios.