Map matching based on multi-layer road index

Highlights

• A method (EICN) is designed to build a semantic multi-layer road index (MRI) system.

• A map matching algorithm (SMRI) is proposed based on a well-established MRI system.

• TA real-world experiment shows the reliability of EICN and the efficiency of SMRI.

• The MRI system can support urban analysis at directional road segment (DRS) level.

Abstract

This study develops a new map matching algorithm targeting off-line applications. The algorithm takes a holistic view of the entire GPS trajectory and finds its match by first dividing it into several segments. This segmentation is made possible through creating a multi-layer road index system for the original road network. For each segment, a global map matching strategy is employed to identify the best match. The algorithm is compared against three state-of-the-art map matching algorithms from the literature. To get ground truth data, we design and perform numerous test drives with predefined paths that have a total length of 234 km. GPS trajectories recorded during the test drives are used to evaluate the algorithms. Our numerical experiments show the proposed algorithm improves match efficiency by up to two order of magnitude compared to the benchmark algorithms. Importantly, it achieves this remarkable speedup with negligible losses in matching accuracy.

Introduction

The application of continuous location data in transportation research has become commonplace. It creates an increasing need for georeferencing, i.e. associating these data with other geographical objects for further analysis. One of the most frequently performed georeferencing tasks is to match GPS trajectories to road networks, often known as map matching (MM) (Zheng, 2015). In the on-line mode, the MM operation forms the foundation for navigation systems that have become an integrated part of driving experience (Hashemi and Karimi, 2014). In the off-line mode, it is mainly used to identify routes consisting of a sequence of intersections and roads. Once the routes are recovered, they can be used to examine route choice behavior (Giannotti et al., 2007, Liu et al., 2014, Zheng, 2015, Xie et al., 2017), determine travel times on road segments (Wang et al., 2014), and estimate the delay incurred at intersections (Liu et al., 2009), among other applications. It is the off-line mode that concerns us here.

To the best of our knowledge, no commercial software exists that performs map matching “automatically”, partly because road networks often come with vastly different formats and quality. Consequently, transportation researchers who are thrilled with the availability of GPS trajectories often face the daunting challenge of developing an MM algorithm to meet their own needs. The purpose of this paper is to address this challenge by proposing a new and efficient MM algorithm.

At first glance, MM seems a rather straightforward operation that typically consists of two steps: (1) identify the right road segment for each GPS point and (2) knit the matched segments together to form a continuous path. The devil, however, is in the details. On the one hand, GPS data for civilian applications is subject to a limit of location precision.¹ In addition, any particular set of GPS trajectories may have errors generated by humans and instruments (Skog and Handel, 2009, Bhowmick and Narvekar, 2018). These errors often lead to irregular sampling intervals (SIs) that could turn a seemingly easy task into a difficult guesswork (Oran and Jaillet, 2017). On the other hand, road networks or maps available to researchers may be of limited quality. In some cases, these networks are simply out of date. More often, however, they may not be designed to cover all existing roads. For example, networks used for urban travel forecasting typically ignore minor local roads (Quddus et al., 2007, Yuan et al., 2012). If a portion of a GPS trajectory happens to lie on these roads, mismatching is inevitable. Hence, the key is to detect and resolve these data quality issues while performing map matching.

Interestingly, we humans can often correctly identify a path from a GPS trajectory littered with missing and/or erroneous data. One advantage of human judgment is the ability to rule out the “impossible” sub-routes created by the problematic GPS points, by taking a holistic view of the trajectory. Indeed, a path typically consists of just a few main distinctive segments which are often easily recognizable from such a holistic view (Gotsman and Kanza, 2015).

This insight leads to a class of “global” MM algorithms attempting to process multiple points together, sometimes even the entire trajectory, to infer the correct path. An early example in this class is the local look-ahead strategy (see e.g., Bernstein and Kornhauser, 1998, White et al., 2000). The idea is to simultaneously process several points to create a piecewise linear curve, which is then compared against the candidate curves formed by individually matched road links. Chawathe (2007) proposed a global algorithm that recognizes spatial heterogeneity in the positional accuracy of GPS points. The algorithm divides a trajectory into segments based on a measure of similarity between the trace points and nearby links. Those segments with higher similarity measures will be given priority when the actual matching operation takes place. Effectively, they are used as “anchors” to match low-similarity segments.

In this paper we propose a MM algorithm that utilizes topological information created through a network pre-process. The method first groups the links in the road network into segments, based on a set of rules called “Cross Continuity Conditions” (3C). The goal is to include in each segment the links that may share a street name and/or have similar orientation and geometry characteristics. For example, a long arterial street traversing through many city blocks may be represented as a single segment with a unique index. Accordingly, we introduce a semantic map called a multi-layer road index (MRI) system, in which segments are represented in the highest layer and links are in the lowest. Using the MRI system, the proposed algorithm, dubbed segment-based MRI (SMRI), first divides each trajectory into segments, and then applies a global strategy to find the best match for each segment. The SMRI algorithm is compared against three state-of-the-art MM algorithms from the literature, namely the hidden Markov matching (HMM) algorithm (Jagadeesh and Srikanthan, 2017), the multi-criteria dynamic programming matching (MDP) algorithm (Chen et al., 2014) and (3) the locality-based map-matching (LBMM) algorithm (Chandio et al., 2015). Two metrics are employed to evaluate the map matching quality: (1) the accuracy ratio of matched points, or ARP (i.e., the ratio of successfully matched trajectory points against manually verified ground truth trajectories); and (2) the time consumed in processing, or TCP (i.e., the average time required to process one thousand points). To get ground truth data, we design and perform numerous test drives with predefined paths that totals 234 km. GPS trajectories recorded during the test drives are used to evaluate the algorithms. The experiment results show that the SMRI algorithm consistently outperforms the benchmark algorithm in terms of both matching accuracy and efficiency.

Fig. 1 provides an overview of the study, which consists of three parts: preprocessing the map data to create the multi-layer road index system; preprocessing GPS data and segment-based matching. The rest of this paper is organized as follows. Section 2 discusses related studies. Section 3 defines the Multi-Layer Road Index (MRI) System and introduces the Cross Continuity Conditions (3C). It also develops an Extended Intersection Continuity Negotiation (EICN) algorithm. In Section 4, we present the SMRI algorithm and Section 5 reports the results of numerical experiments designed to evaluate the performance of the SMRI algorithm. The last section concludes the paper and discusses the directions for future research.