Research On Pedestrian Trajectory Prediction In Complex Environment

Predicting the future trajectory of pedestrians is a basic problem in the research of human-computer interaction technology or intelligent systems in various fields.It has a very wide range of applications in self-driving vehicles,robot navigation systems,service robots,intelligent transportation system and video surveillance systems.However,due to the complexity of human activities in crowded environments,accurately simulating and predicting the trajectory of pedestrians has always been a very challenging problem.Usually,the solution of traditional pedestrian trajectory prediction is based on the method of artificially extracting features.Artificial feature extraction does not have good versatility and scalability since it is difficult to establish a complex mathematical model and the computational complexity is pretty high.However,the method of deep-learning which is based on data-driven can effectively make up for these defects.In order to improve the accuracy of trajectory prediction,it is important to model the interaction between pedestrians.Therefore,this thesis will use the method of combining the traditional ideas of social-force model with deep-learning to carry out our model of pedestrian trajectory prediction based on Long Short-Term Memory(LSTM).This thesis studies the existing trajectory prediction methods and proposes the improve methods.The first is to study the social-force model that artificially simulates the interactions of pedestrians,the second is to model and analyze the existing deeplearning methods that focus on the important influences of the social interactions between pedestrians,including the Social-LSTM model based on LSTM network,the attention mechanism in the field of Natural Language Processing(NLP)and the Generative Adversarial Networks(GAN).After comparing the disadvantages of each method,we propose an optimized plan for the disadvantages and then propose a social intension model named IS-L2 and IS-GAN to achieve the trajectory prediction that conforms to social norms in a complex environment.We train and test different models on the synthetic dataset for interactions of pedestrians created by the social-force model,and compare the interaction performances of the existing methods with the IS-L2 proposed in this thesis.Furthermore,we simulate and analyze the trajectory prediction methods of different models,and visualize the prediction effects of these models.In view of the multimodality of pedestrian trajectory prediction and the lack of training data for multiple-future prediction in existing large public datasets,this thesis uses a 3D simulator CARLA to create a new dataset that is reproduced with the static scenes and dynamic elements in ETH/UCY and VIRAT/Act EV video datasets,and we use the simulator to artificially annotate multiple possible future trajectories in the videos as training data for the model.In addition,we generate data for each training trajectory and scene images from multiple camera views to improve the robustness of the model by creating enhanced samples in datasets.Finally,we use the multi-view simulation dataset to train these models on the simulated dataset and and test them on the real environment datasets in order to quantitatively and qualitatively analyze the performances of different methods.