Title: Compressive Origin-Destination Matrix Estimation

Abstract: The Origin-Destination (OD) Matrix Estimation (ME) is one of the main tasks in transportation planning. Recent approaches for OD ME include estimation based on link traffic counts. In most of the situations, however, the ME is cast as an underdetermined inverse problem with infinitely many solutions as usually the number of available individual link traffic counts is much smaller than the total number of possible paths in a traffic network. As a result, there is no unique solution and one should use other available information in the ME procedure. In this paper, we aim to perform ME under the assumption that path allocations are suitably sparse (i.e., among all alternative paths that exist for a given OD pair, only a few of the paths are taken in a trip). Our work, called Compressive Origin-Destination Matrix Estimation (CODE), is inspired by Compressive Sensing (CS) which is a recent paradigm in signal processing for sparse signal recovery. We show that in cases where the true path allocation is suitably sparse, it is possible to perform the OD ME from a small number of link flow observations. The main technical novelty of our approach is in casting the OD ME problem as $\ell_1$-recovery of a sparse signal $\boldsymbol{x} \in \mathbb{R}^{N}$ from measurements $\boldsymbol{y} = A \boldsymbol{x} \in \mathbb{R}^M$ with $M < N$, where $\boldsymbol{y}$ contains link traffic counts, $A$ is a binary matrix whose structure is dependent on the topology of the network and is assumed to be known, and $\boldsymbol{x}$ is the path allocation vector with $\|\boldsymbol{x}\|_0 \leq S < N$. The path allocation vector $\boldsymbol{x}$ contains all OD pair flows and path splits for each OD pair as unknown variables. In addition to synthetic examples, we consider applying CODE to real data taken from a region in East Providence.