Pedestrian Trajectory Prediction Based On LSTM And Perceptron

With the development of economy,the scale of cities is expanding,and the traffic scenes in urban clusters are more complex and changeable.Vehicles,pedestrians,and other traffic participants put forward higher requirements for predicting the direction of the route in the near future.It has become a very challenging task to predict the trajectories of various traffic participants by means of science and technology.Compared with vehicles,pedestrians are less constrained by rules in most scenes,such as pedestrian street,subway station,etc.,which leads to more complex and changeable pedestrian path.With the continuous development of deep learning,the research of pedestrian trajectory prediction based on various neural networks emerges in an endless stream,and has achieved good results,especially the popular long short term memory(LSTM)has become one of the popular networks for pedestrian trajectory prediction due to its combination of long-term and short-term memory of temporal information.Although LSTM encodes position and displacement information well,it ignores the relationship between pedestrians.In this paper,the relationship between pedestrians is used to optimize the cell state of LSTM.In the process of LSTM back-propagation,the pedestrian relationship module is constructed by organically combining the original independent pedestrian trajectories.The experimental results show that the prediction results of the optimized model are better than those of the crowd interaction deep neural network(cidnn)model on GC,ucy,ETH three public data sets.Secondly,as an important information to describe the state of pedestrian movement,velocity plays an important role in pedestrian trajectory prediction.In this paper,the perceptron is used to encode the pedestrian speed and construct the pedestrian movement module.Experimental results show that the motion module performs well on UNIV data sets of GC,ucy and eth,Finally,based on the advantages and disadvantages of the former two modules in pedestrian trajectory prediction task,this paper designs a pedestrian trajectory prediction fusion model which combines the pedestrian movement module and the pedestrian interaction module.The fusion model is evaluated on the open data sets of GC,ucy and eth.The experimental results show that the average Euclidean distance and the final Euclidean distance prediction results of the fusion model are better than the cidnn model on the three data sets,and the prediction effect is significantly improved on the HOTEL data set of eth.We also compare the fusion model with the state refinement for long short term memory(sr-lstm),cidnn and other pedestrian trajectory prediction methods.The experimental results show that the final European distance prediction results of the fusion model are better than these models.