Research On Pedestrian Trajectory Prediction Algorithm Based On GRU And Attention Mechanism

In the long-term evolution,human being has been fully socialized.They can realize the state of the environment in advance and navigate autonomously before crossing a congested area.They can choose a suitable road to avoid danger or collision.This kind of early warning capability,the pedestrian trajectory prediction model simulated by computer simulation should be able to simulate the prediction ability of human beings in a crowded environment.Various intelligent systems based on this type of prediction model can greatly reduce the probability of equipment collision in complex environments,and can plan routes for specific targets.At present,there are many achievements in this field,but the traditional pedestrian trajectory prediction methods are based on manually set features to model pedestrian behavior.In view of the fact that the pedestrian movement scene is full of various influencing factors,the modeling method based on artificial features is no longer suitable for the complex and changeable environment of pedestrians,and can only adapt to specific environments.The emergence of artificial neural network provides a solution to this kind of problem.Through machine learning,the features of large amounts of data can be used to extract and use.Machine learning has proved to be an excellent method for solving sequence problems such as speech recognition,language translation,and text generation.As a typical sequence-to-sequence problem,the pedestrian's future trajectory can theoretically be predicted using machine learning.In this problem,a pedestrian's historical movement trajectory is used as the input sequence of the neural network,and the pedestrian's movement trajectory in the future is the output sequence.At present,although the research on the pedestrian trajectory prediction task has achieved certain results,the extraction of pedestrian states and the limitations that are restricted due to the choice of scenes have yet to be resolved.Due to the problems above,this dissertation starts from the attention mechanism and uses a recurrent neural network to construct a pedestrian trajectory prediction model.First,the models of this dissertation encode all the pedestrians in the scene with recurrent neural network to obtain the hidden features and behavior habits of each person.Second,the models add a module containing a single attention mechanism or a double attention mechanism to perform the motion state of the pedestrians in the scene.Finally the models input the vector through the attention mechanism module to the decoder to obtain the final pedestrian prediction trajectory.After comparative analysis of the experiments on the public experimental data set,the algorithm model proposed in this dissertation reduces the prediction error and is effective and feasible.The research work in this dissertation includes the following three points:(1)Aiming at the problem of incomplete extraction of pedestrian states by traditional methods and restricted by scene selection,this dissertation proposes a GRU pedestrian trajectory prediction method based on scene states.This method adds a scene state processing layer between the encoder-decoder architecture to extract the relative spatial state of pedestrians and filter pedestrian features.Finally,the model's ability to grasp the movement of pedestrians in the scene is improved.(2)In order to continue to improve the accuracy of the model for pedestrian trajectory prediction,this dissertation attempts to combine dual attention mechanism for pedestrian trajectory prediction.By focusing on the pedestrian's historical movement habits and scene state,the model's prediction ability has been further improved.(3)Aiming at the problem of the traditional method's insufficient grasp of pedestrian motion characteristics in certain specific scenes,and in order to improve the model's more accurate correlation with the pedestrian's potential features,this dissertation introduces a two-way GRU as a pedestrian encoder,respectively from positive sequence and reverse order to extract the hidden state of pedestrian movement and weighted fusion,so as to enhance the model's ability to describe the movement scene.