Scalable Framework for Enhancing Raw GPS Trajectory Data: Application to Trip Analytics for Transportation Planning

Abstract

Transportation analysts and planners are beginning to leverage GPS trajectory data to draw additional insight into travel behavior and enhance data-driven decision-making capabilities. However, raw trajectory data cannot be utilized directly; they require extensive processing prior to analysis. This paper presents a scalable approach for enhancing raw GPS trajectory data by snapping and routing waypoints along a user-defined target road network that may have discontinuities and missing links, thus enabling trajectory datasets to be used in conjunction with the types of non-routable road networks often employed by transportation agencies. The proposed approach fuses a well-established map matching solution with a custom waypoint conflation procedure, and provides a framework to execute the trajectory processing in parallel to efficiently leverage available computing resources for large GPS datasets. To demonstrate its capability, four months of 2018 trajectory data from Maryland (2.5 billion waypoints from 46 million trips) are processed in this manner and assigned to a Traffic Message Channel road network. The enhanced trajectory data are then used to demonstrate a real-world use case, identifying key travel patterns along the I-270 spur in Maryland—a key commuting corridor currently being considered by the Maryland Department of Transportation for a congestion mitigation investment.

Appendix: OSRM Map Matching

For completeness, we provide a brief description of the method from Newson and Krumm (2009) that is implemented by the Open Source Routing Machine (OSRM), which is used in Step 1 to match raw GPS points to the road network. The map matching algorithm employed by the OSRM uses a Hidden Markov Model (HMM) to find the most likely road route represented by a time-stamped sequence of latitude/longitude pairs (Newson and Krumm 2009). Given a sequence of GPS points $\{z_1,z_2,\ldots ,z_t,\ldots ,z_T\}$ and the road segments in the OSM network $\{r_1,r_2,\ldots ,r_i,\ldots ,r_N\}$, the HMM calculates the measurement probability $p(z_t|r_i)$ that the GPS points $z_t$ would be observed if the vehicle was actually on road segment $r_i$ as

$$\begin{aligned} p(z_t | r_i) = \dfrac{1}{\sqrt{2\pi }\sigma _z}e^{-0.5\left( \dfrac{\Vert z_t-x_{t,i}\Vert }{\sigma _z}\right) ^2}, \end{aligned}$$

(1)

where $\sigma _z$ is the standard deviation of GPS measurements and $\Vert z_t-x_{t,i}\Vert$ measures the great-circle distance between $z_t$ and the closest point $x_{t,i}$ to $z_t$ on the road segment $r_i$.

Based on (1), the probability of each GPS point matched to a list of candidate road segments can be obtained. The next step requires determining the probability of a vehicle moving between candidate road segments for any two temporal adjacent trace points. The transition probability function is defined to measure the probability of a vehicle moving between the candidate road matches,

$$\begin{aligned} p(d_t) = \dfrac{1}{\beta }e^{-\frac{d_t}{\beta }}, \end{aligned}$$

(2)

in which

$$\begin{aligned} d_t = |\ \Vert z_t - z_{t+1}\Vert - \Vert x_{t,i} - x_{t+1,j}\Vert \ |, \end{aligned}$$

(3)

where $x_{t,i}$ and $x_{t+1,j}$ represents the closest point on road $r_i$ and $r_j$ for GPS point $z_t$ and $z_{t+1}$, respectively, and $\Vert x_{t,i} - x_{t+1,j}\Vert$ denotes the driving distance between point $x_{t,i}$ and $x_{t+1,j}$.

Given the two probability functions, the Viterbi algorithm is used to quickly find the most plausible path that maximizes the product of the measurement probabilities and transition probabilities (Newson and Krumm 2009).