Head Pose Feature for Prediction Pedestrian Intention to Crossing the Road Using LSTM

Understanding pedestrian behaviour when crossing the road is an important key to the development of autonomous vehicles. Because pedestrians are considered Vul-nerable Road Users (VRUs), they are likely to be killed if they are involved in an accident. To ensure their safety, it is then necessary to predict the pedestrian's intention based on their behaviour. In this experiment, we propose head pose observation for predicting their intention, by observing pedestrians' head pose data, we can then predict their intention to cross the road. To achieve this purpose, we use human head detector and head pose extraction feature, and the resulting yaw, pitch and roll as three head pose features. To select the most optimal feature is important for predicting pedestrian intention, then we make 7 combination scenarios based on these three features and compare it with the same model. Based on this scenario, it is proved that all these three data are optimal to observe pedestrian intention. There are three behavioural annotation that have been used, there are crossing, not crossing and will crossing. We derive will crossing from the annotation looking and not crossing while waiting at the roadside. Prediction of pedestrian behaviour is done by using LSTM model, and the resulting precision on crossing and not crossing with 0.98, while will crossing is 0.94.