In view of the influence of pedestrians on vehicle driving in traffic scenes and the problem that the assisted driving needs to avoid the danger for pedestrians, a pedestrian hazard assessment method based on vehicle-mounted monocular cameras is proposed. Based on the characteristic environment of Chinese cities, this paper divided the driving environment into three scenarios: regular roads, crosswalks, auxiliary roads. Different risk assessment methods are used for each type of scenarios. A convolutional neural network is used to detect and identify pedestrians, auxiliary police, signal lights and sidewalks on the road in the video. Then, it detects the key points of the pedestrian and uses the multi-target tracking method to generate the time series of the pedestrian skeleton. Pedestrian behavior and trend are obtained through LSTM（Long-Short Term Memory Neural Network) analysis of posture sequences. Finally, a pedestrian hazard assessment model is built to realize the pedestrian hazard assessment in multi road scene by synthesizing the video information, pedestrian information and scene information. The experimental results show that the scene classification makes the assessment results of the hazard model more consistent with the actual pedestrian hazard. The proposed model can effectively evaluate the pedestrian hazard and assist the drivers to drive safely.