Estimation of pedestrian crossing intentions in in-vehicle video

In recent years, there has been a lot of research on mathematical models of traffic flow, and various models have been proposed for vehicles and pedestrians. Such mathematical models are often used to simulate traffic congestion and disaster evacuation because they behave close to reality under certain conditions. In such simulations, the destination and travel path states of individual agents are often limited. However, when looking at actual traffic flow, individual vehicles and pedestrians often have a large degree of freedom in their destinations and travel paths. Estimating such an initial state is not easy because it requires consideration of not only physical quantities such as the speed and direction of movement, but also the pedestrian's intentions, such as which direction they are trying to go. Therefore, this study addresses the inference of the intentions of the persons in the video. In this study, we assume in-vehicle video and classify whether the pedestrian in the video is attempting to cross the roadway or not. Experimental results showed that classification of in-vehicle video datasets using a 3D-extended convolutional neural network is possible with an accuracy of more than 70%. Prediction using multiple frames with the 3D-CNN was also shown to be more accurate than prediction using only one previous frame.