Prediction of pedestrian trajectory based on long short-term memory of data

Abstract

Recently, development of navigation robots and autonomous cars are rapidly progressing. When such robots become popular in our daily life, their collisions with humans should be avoided for safety. For that purpose, we predict pedestrian trajectories with LSTM (long short-term memory) networks and conventional neural networks, and we compare their results. In order to predict sequential data, we use the following two methods: (I) predicting n steps of data with n models, and (II) predicting n steps of data with a model by applying one-step prediction several times. By examining these two methods, it was found that the performances of the method I with the LSTM network and the conventional neural network are comparable, and the performance of the method II with the LSTM network is significantly better than that with the conventional neural network.