Research And Application Of Pedestrian Trajectory Prediction Method Based On Scene-constrained GAN

With the development of science and technology,the application of intelligent devices is more and more popular in daily life.These intelligent devices generate a huge massive of trajectory data,so it is in urgent need to sufficiently analyze and leverage these data.Learning historical trajectory modes and understanding pedestrian motion patterns are the key issues of the trajectory prediction,which is a challenging problem in the area of self-driving,socially conscious robots,etc.At present,researchers have proposed a variety of trajectory prediction methods,including the method based on linear function,statistics or deep learning.Among them,the method based on linear function may only ensure the accuracy in a short predicted period,and cannot model for complex motions.The statistical method can fuse the context information of trajectory,but it is difficult to process high-dimensional data and design a universally applicable method.In recent years,the methods to predict the future trajectory based on deep learning are mainly focused on two aspects:the influence of physical environment on pedestrian trajectory,and the social interaction between target pedestrians.However,few methods consider both the physical environment and social interaction.In order to solve these problems,this paper proposes a trajectory prediction model with higher accuracy and stronger robustness.This model considers the target motion characteristics,the constraint information of scene information on pedestrian's trajectory decision,and the interaction between the pedestrian and other pedestrians simultaneously,and uses the GAN training mode to generate clearer and true samples.In the generator,a coder-decoder is built to process the pedestrian's trajectory data by using the LSTM,because the LSTM can learn the movement pattern of the pedestrians from the trajectory data,and its hidden layer can "remember" the historical track information for a long time.Secondly,considering the interaction between the target pedestrian and other pedestrians,a pooling module is constructed,in which the interaction is simulated through the relative position between the pedestrians.Finally,to get the constraint information of scene information on pedestrian's trajectory decision,this paper uses the convolutional neural network to extract the semantic features of the scene,and obtains the high-dimensional vectors by fusing the scene feature semantics with feature vectors which contain the pedestrians motion patterns and interaction between pedestrians.By using the high-dimensional vectors,the results of trajectory prediction would be more accurate.In addition,based on the trajectory prediction method proposed in this paper,a pedestrian trajectory prediction system based on the scene constraint GAN is designed and implemented.The design of the system,from the whole to the part,achieves the functional modules of model training and trajectory prediction,and develops the functions of the operation interface,data verification and error prompt for the two modules to perfect the system and simplify the operation.The experimental results show that the results of GAN trajectory prediction based on scene constraints proposed in this paper are more accurate than others under the standards of both tie displacement error and final displacement error.