Abstract: Automatic incident detection not only plays an important role in traffic safety management, but also contributes to the operation of intelligent transportation systems. Although the emerging information technologies and artificial intelligence approaches are paving the way for high-precision incident detection, existing incident detection methods fail to handle the unbalanced incident data with excessive zero observations. Also, issues related to network delays and privacy leakage of centralized computing are prevalent. To fill the above gaps, this study proposes a novel automatic incident detection paradigm using an edge-cloud collaboration mechanism. In particular, a Spatio-Temporal Variational Digraph Auto-Encoder model is developed to distinguish the incidents in dynamic traffic flows. To be specific, the model encoder includes two components. The first module, deployed in an edge server, is designed to extract the local contexts from the real-time traffic flow. The dynamic traffic flows will be projected into a spatio-temporal digraph, and in turn addressed by a graph convolutional network for extraction of the deep-seated features. Similarly, the second module is deployed in a central server to capture the spatio-temporal global contexts from historical traffic flows. Finally, the above-concerned contexts are integrated and fed into a model decoder to measure the likelihood of incidents. To testify the proposed paradigm and model, real-world datasets were applied. The experimental results revealed the proposed model outperforms state-of-the-art models in terms of detection accuracy, achieving 26.3% improvement over the best-performing baseline. Furthermore, the proposed paradigm is more efficient in respondence compared with traditional centralized computing, realizing 8x processing speed for the same detection task.