PTPGC: Pedestrian trajectory prediction by graph attention network with ConvLSTM

Abstract

With the rapid development of intelligent systems, such as self-driving vehicles, service robots and surveillance systems, pedestrian trajectory prediction has become a very challenging problem. How to perceive, understand and predict the motion patterns of pedestrians in a highly crowded and chaotic environment in order to prevent future collisions becomes a top priority. The motion of pedestrians is not only affected by their own factors, but also by surrounding neighbors. To solve the above problems, we propose a model named PTPGC based on graph attention and convolutional long short-term memory (ConvLSTM) network to predict multiple reasonable pedestrian trajectories. Firstly, the pedestrians are represented in a dynamic graph by setting a Euclidean distance threshold. Then, a graph attention network is used to learn the spatial interaction relationship of all pedestrians in each time step, and a temporal convolutional network (TCN) is used to encode the pedestrians’ own factors. Finally, we use the ConvLSTM to iteratively predict the multiple reasonable and feasible future trajectories of pedestrians. Experiments show that our model has a higher prediction accuracy on two public pedestrian data sets (ETH and UCY) compared with the existing baselines for pedestrian trajectory prediction, and the generated trajectories are more in line with social rationality and physical constraints.