A Method of Pedestrian Trajectory Prediction Based on LSTM

Abstract

In the public scene, different pedestrian walks on different paths to avoid colliding with obstacles or others. Any small vehicle navigation in such a scenario should be able to anticipate the approximate position of the people around it at the next moment, and adjust its path to avoid collisions based on the predicted results. Such a problem of trajectory prediction can be regarded as the task of sequence generation, and we are interested in how to predict the future trajectory of pedestrians based on their past trajectory. In recent years, Recurrent Neural Network (RNN) model has been successful in sequence prediction tasks. So, this paper proposes a model combining an attention mechanism and Long Short-Term Memory (LSTM) artificial neural networks, to solve this question. This model can predict a pedestrian's future trajectory by learning his past trajectory. Experiments shows the model work well on multiple datasets, and the test results show that it has a very good effect.