Roadside monitoring equipments can obtain traffic flow and speed real-time data more accurately, so that traffic management departments can significantly improve the perception of traffic anomalies. By analyzing the characteristics of traffic state and traffic flow data, this paper establishes a set of real-time traffic accident detection system and early warning process based on traffic flow sequence data. First of all, in terms of traffic state perception, this paper introduces the Attention mechanism to capture important traffic condition features based on the Seq2Seq model. Secondly, in terms of anomaly detection of traffic condition, the auto-encoder is used to realize the reconstruction of input sequence data, and the structural reconstruction error is obtained by comparing the original data, and then the real-time detection of traffic accidents and the classification of accident warning levels are realized according to the set threshold. Finally, based on the traffic and speed data of the Shanghai Yanán elevated road, the thresholds for evaluating traffic accidents in different time and space conditions are determined respectively, and the feasibility of the real-time traffic accident detection model is demonstrated by the confusion matrix evaluation method.