A pedestrian road-crossing behavior prediction method for a plane intersection, comprising the following steps: step 1: designing an instant reward function; step 2: establishing a fully convolutional neural network-long-short term memory (FCN-LSTM) model to predict an action reward function; step 3: training the (FCN-LSTM) model on the basis of reinforcement learning; and step 4: predicting pedestrian crossing behavior and carrying out a danger early warning. In the technical solution, a complex pedestrian motion model does not need to be established, and a large number of labeled data sets do not need to be prepared. Autonomous learning of pedestrian road-crossing behavior features at a plane intersection is achieved, and behavior such as walking, stopping and fast running are predicted. In particular, pedestrian road-crossing behavior when dangers such as human-vehicle collisions and scrapes are induced are predicted in real time. Therefore, a danger early warning is carried out on pedestrians crossing the road and passing vehicles, which is beneficial for reducing the traffic accident rate of key road sections such as plane intersections, and ensures the safety of the pedestrians in a traffic environment.