Context-based Trajectory Prediction with LSTM Networks

Abstract

Traditional target trajectory prediction model is generally trained on the previous trajectories purely while the context information of the trajectory is simply ignored. We assume that the trajectory pattern generally associates with a certain set of positions. For instance, the travelling trajectories of people of similar interest may be highly correlated. Such kind of context information provides more clues for trajectory prediction. As a result, context information should be utilized during trajectory predictions. Inspired by the above issue, we have designed an effective context-based trajectory prediction method with two types of LSTMs. The first type of LSTM model is specially built to predict the distinctive pattern that the trajectory follows while the other type of LSTM models are designed to predict the future positions of the trajectory given the context of the pattern it follows. First, we convert the real-valued target trajectories into discrete path sets with grids. And then we discover the distinctive patterns with hierarchical clustering. The context of the trajectory is modeled as the closest grid of the associated pattern. Later, we train the two types of LSTM models with the corresponding samples. Lastly, we apply the LSTM models for trajectory prediction. Experimental results indicate that our method outperforms the traditional LSTM neural networks significantly by making use of the context information of the trajectory.