Research On Data-driven Pedestrian Trajectory Prediction Diction Methods

Pedestrian trajectory prediction refers to predicting pedestrians’trajectory in the future given the trajectory of pedestrians in a historical period and the scene information in a given scene.Pedestrian trajectory prediction has important applications in many aspects,such as autonomous driving,robot navigation,and intelligent transportation systems.A critical task of autonomous vehicles and robots in traveling is to analyze the movement intentions and tendencies of other users on the road,especially for pedestrians in a disadvantaged position,to avoid possible collisions.The difficulty of the pedestrian trajectory prediction mainly consists of three parts.First,pedestrians’ interaction will affect each other’s movement,so predicting the target pedestrian’s trajectory needs to consider other pedestrians around.Besides,the influence of different pedestrians on the target pedestrian is different.The influence is dynamically changing,so it is difficult to model pedestrians’ influence with simple hand-crafted rules.Second,pedestrians’ movement will be affected by the surrounding scenes.Pedestrians will walk on the road and try to avoid trampling on the lawn.Pedestrians will not collide with buildings.Pedestrians will stop at red lights and start to move when lights turn green,and those who want to drive will move towards the vehicle.Therefore,the influence of the surrounding scene is also significant.Third,because many factors such as pedestrians’ goals and mental states are unknown when making trajectory prediction,it is reasonable that there will be multiple possible future trajectories for each historical trajectory;that is,pedestrian trajectory prediction is a multi-modal problem.We focus on the problem of pedestrian trajectory prediction.Firstly,a new quantitative evaluation method is proposed because of the shortcomings of the existing crowd trajectory prediction method evaluation system.In addition to the accuracy of the trajectory,the rationality of the trajectory is also considered.We propose to use the average collision times(ACT)to statistically count the collision of predicted trajectories,which is a consideration of the rationality of predicted trajectories.Secondly,to fully consider the interaction between pedestrians and their influence,and reduce the collision of predicted trajectories,a model based on attention mechanism and generative adversarial network is proposed.Its attention mechanism directly infers each pedestrian’s importance to the target pedestrian from the relative distance and relative speed between them at each prediction moment.Inspired by PixelGAN and PatchGAN,we optimize the discriminator of the model.The experimental results show that the ACT-avg of CoL-GAN on the ETH-UCY dataset is 17.917%lower than that of STGAT,and ADE(average displacement error)is 1.845%lower.Finally,for scenes with a large number of pedestrians,frequent interactions,and long prediction time(such as the PEDWALK dataset),a model based on graph convolutional neural network is proposed.In order to better model the pedestrian motion interaction process,it builds a multi-level spatio-temporal directed graph to model pedestrians’interaction.By dividing pedestrians in different distance ranges into different levels of the graph,the interactive effects of pedestrians in different distance ranges are modeled differentially.In addition,the spatio-temporal directed graph can not only model the temporal and spatial characteristics of pedestrian interaction but also model the asymmetry of pedestrian interaction.Experiments show that the model achieves a 7.765%lower ADE and a 4.172%lower FDE(final displacement error)on the PEDWALK dataset than STGAT.