Methodologies for Estimating Traffic Flow on Freeways Using Probe Vehicle Trajectory Data

Probe vehicle data are increasingly becoming the primary source of traffic data. As probe vehicle data become more widespread, it is imperative that methods are developed so that traffic state estimators such as flow, density, and speed can be derived from such data. In this dissertation three different methodologies are proposed for predicting traffic flow or volume on a freeway. All of the proposed methodologies exploit several different traffic flow theories in conjunction with probe vehicle data to predict traffic flow. The first methodology takes advantage of the fundamental diagram or speed-flow relationship. The relationship states that flow can be estimated when speed is known. In this case, flow is traffic volume and speed comes from probe vehicles. Flow is predicted for four different models of fundamental diagrams and is analyzed at different time aggregation intervals. Results show that of the four fundamental diagrams, Van Aerde’s Model is the best performing model with the lowest average percent error. It is also observed that flow prediction is more accurate during low speed (congestion) compared to high speed (free-flow) conditions. The second methodology exploits the shockwave theory, which pertains to the propagation of a change (discontinuity) in traffic flow. From probe vehicle trajectories, shockwave is estimated as the boundary between free-flow and congested regimes of traffic flow. After clustering the traffic regimes into free-flow and congested periods, the traffic flow during congestion is estimated using the Northwestern congested-regime fundamental diagram. From this estimation, the flow during free-flow is then predicted. Analyses show that the percent error of the predicted flow during free-flow ranges from -9 to 1%. The third methodology is the car-following approach which relies on the spacing or distance between a leader and follower which can be directly measured from the trajectories. Based on a set of known probability distributions, the position of the follower vehicle with respect to the lead vehicle is estimated given that the spacing between the two random probe vehicles is known. A framework is developed to automatically process probe trajectories to extract relevant probe data under stop-and-go traffic conditions. The model is tested based on NGSIM datasets. The results show that when vehicle spacing is small the prediction of follower position is very accurate. As spacing increases the error in predicted follower position also increases. Though there exists some estimation error, all three approaches can reasonably predict flow for freeways using probe vehicle data