Aiming at the problem that the highway management personnel rely on personal experience to open the emergency lane, and it is difficult to scientifically regulate the traffic flow in real time, a highway traffic flow congestion model is proposed. This model processes the video data based on the YOLO-V5s algorithm, extracts the traffic flow parameters, uses polynomial regression to fit their change rules, establishes the model with the common time period index as the training set, and adopts the weighted linear extrapolation method to issue congestion warnings. The results show that for the road sections where the predicted congestion exceeds 30 min, the warning can be issued 10 min in advance. The prediction accuracy rate of the traffic flow density reaches 85.76%, and the prediction accuracy rate of the vehicle speed reaches 88.19%. This provides a new idea for highway congestion warning.