Predicting Future Locations of Moving Objects by Recurrent Mixture Density Network

Rui Chen,Wanli Li,Naikun Guo,Mingjian Chen

doi:10.3390/ijgi9020116

Abstract

Accurate and timely location prediction of moving objects is crucial for intelligent transportation systems and traffic management. In recent years, ubiquitous location acquisition technologies have provided the opportunity for mining knowledge from trajectories, making location prediction and real-time decisions more feasible. Previous location prediction methods have mostly developed on the basis of shallow models whereas shallow models are not competent for some tricky challenges such as multi-time-step location coordinates prediction. Motivated by the current study status, we are dedicated to a deep-learning-based approach to predict the coordinates of several future locations of moving objects based on recent trajectory records. The method of this work consists of three successive parts: trajectory preprocessing, prediction model construction, and post-processing. In this work, a prediction model named the bidirectional recurrent mixture density network (BiRMDN) was constructed by integrating the long short-term memory (LSTM) and mixture density network (MDN) together. This model has the ability to learn long-term contextual information from recent trajectory and model real-valued location coordinates. We employed a vessel trajectory dataset for the implementation of this approach and determined the optimal model configuration after several parameter analysis experiments. Experimental results involving a performance comparison with other widely used methods demonstrate the superiority of the BiRMDN model.

Highlights

Location prediction refers to estimating the anticipated location at several following timestamps based on recent trajectory and the current status [1]
There are still some bottlenecks that exist in the current representation-based methods. It is tricky for deep learning algorithms such as the recurrent neural network (RNN) [14] and convolutional neural network (CNN) [15], which are good at performing discrete prediction, to predict real-valued location coordinates
The bidirectional recurrent mixture density network (BiRMDN) integrating the long short-term memory (LSTM) network and mixture density network (MDN) has the ability to learn the contextual information from a recent part of the trajectory and exploit the probability distribution function to model the uncertainty of future locations

Summary

Introduction

Location prediction refers to estimating the anticipated location at several following timestamps based on recent trajectory and the current status [1]. There are still some bottlenecks that exist in the current representation-based methods It is tricky for deep learning algorithms such as the recurrent neural network (RNN) [14] and convolutional neural network (CNN) [15], which are good at performing discrete prediction, to predict real-valued location coordinates. Motivated by the limitations of the existing methods in predicting long-term location coordinates, we dedicated this paper to a location prediction approach based on deep learning The core of this approach is the construction of a bidirectional recurrent mixture density network (BiRMDN) model. The main superiority of BiRMDN model lies in the following: (1) It is better at improving the prediction accuracy than the traditional machine learning methods due to the utilization of long-term spatiotemporal dependency and contextual information in trajectory data.

Related Work

Methods

Experimental Results

Dataset and Experiment Setup

Analysis of Parameter Effect

Comparative Study with Other Approaches