Abstract : Predicting intersection turning movements is an important task for urban traffic analysis, planning, and signal control. However, traffic flow dynamics in the vicinity of urban arterial intersections is a complex and nonlinear phenomenon influenced by factors such as signal timing plan, road geometry, driver behaviors, queuing, etc. Most current methods focus on predicting turning movement counts using data at coarser aggregations in the order of minutes or above. Important details such as platoon movements may be lost at such coarse resolutions. In this work, we propose machine learning approaches to imputing turning movement counts at intersections using data at subcycle resolutions, from 5 seconds to 375 seconds. In particular, we show that deep neural networks are capable of directly learning an abstract representation of intersection traffic dynamics using detector actuation waveforms and signal state information. We generate a large dataset of 30 million cycles by approximately replicating real-world traffic arrival patterns from archived loop detector data in a microscopic traffic simulator. We extensively evaluate our models and show that our models predict turning movement counts with greater accuracy when higher resolution data are provided.
Turn movement counts (TMCs) are used for a wide variety of applications related to intersection analyses, intersection design, and transport planning. Recently, several advances in traffic control technologies and applications are driving additional needs for continuous, real-time, quality TMC data. These technologies include:
- Adaptive control technologies: These systems dynamically change the signal phasing pattern locally at an intersection based on traffic demand.
- Regional Integrated Corridor Management System (R-ICMS): This system will perform a mesoscopic simulation of the network using TMCs to validate the improvement to the network if a diversion route response plan is implemented or an optimized set of signal timing plans is deployed.
Our contributions can be summarized as follows:
- We show that using detector waveforms instead of aggregated traffic volumes leads to better accuracy for turning movement counts prediction. Our neural network models perform well both under unsaturated and oversaturated conditions.
- We have developed neural network models for predicting turning movement counts for an intersection, given information from stop bar detectors and advance detectors of the intersection and advance detectors of upstream intersections. Our approach uses a novel modular neural network architecture for predicting turning movement counts. By using a common weight matrix for all four directions in the first hidden layer of the network, the size of this network is much smaller than a typical feed-forward network, thereby reducing the uncertainty in predictions.
- We have developed a system for generating synthetic datasets with traffic distributions close to real world. We used detector controller logs from 300 intersections in Orlando to construct inflow waveforms and fed them into SUMO to run simulations.