Map Matching for Taxi GPS Data with Extreme Learning Machine

Abstract

In this work, we explore a new map matching method through mining historical GPS data collected by taxis. The principle behind is that the map matching can be regarded as a pattern recognition if there are enough historical GPS points labelled with road network information. Supervised learning algorithms are feasible for this situation. However, the learning speed of conventional learning techniques is often not satisfactory, especially facing enumerous classes (road labels). Considering the matching (classifying) speed and accuracy, we employ the Extreme Learning Machine (ELM) as a multi-class classifier for its excellent performance in the learning speed. Furthermore, we propose MapReduce based GPS trajectories training data preprocessing algorithms and an optimal ELM parameter selection algorithm. Extensive experimental results show that, compared to the SVM-based approach, the ELM-based approach achieved faster learning speed and matching speed, while got close to similar performance on matching accuracy.