The increasing frequency of road accidents highlights the urgent need for intelligent systems capable of real-time incident detection to support rapid emergency response. This paper presents a lightweight spatiotemporal accident detection framework that integrates a single-stage object detection model with an auxiliary motion-analysis stream based on dense optical flow. By jointly exploiting spatial appearance information and short-term temporal motion variations, the proposed system aims to improve robustness in complex traffic scenes while maintaining real-time performance. Unlike conventional approaches that rely solely on frame-wise object detection, the framework captures motion irregularities surrounding collision events to mitigate false alarms caused by normal traffic dynamics. A multi-source dataset was curated from publicly available traffic surveillance images and accident-related video clips obtained from heterogeneous sources, encompassing diverse viewpoints, traffic densities, and environmental conditions. The system was evaluated against representative object detection baselines using standard detection metrics, along with inference time analysis to assess deployment feasibility. Experimental results demonstrate that the proposed fusion-based approach achieves improved detection consistency with low computational overhead, making it suitable for real-time surveillance applications. The study highlights the effectiveness of combining spatial detection with simple temporal motion cues for practical accident monitoring in intelligent transportation systems, while also discussing current limitations and directions for future enhancement.