Machine Learning Based Automated Left-Turn Vehicle Counts With Conventional Presence Mode Long-Loop Detectors: Alabama Case Studies

Abstract

Inductive-loop detectors are still used in many intersections because highway agencies have insufficient budget to replace them with new technologies or when accurate traffic monitoring is not necessary. Particularly, long-loop stop-bar detectors have been widely used for left-turning vehicles and are operated under “Presence” mode, which focuses only on detecting the presence of vehicles for signal operations, not for counting vehicles. As a result, a significant discrepancy exists between the number of vehicle detections and the actual vehicle observations. The present study developed machine learning (ML)-based methods to overcome this discrepancy so that conventional long-loop detectors can also be used for estimating hourly left-turn volume. Several ML classifiers were adopted to predict the difference (i.e., detection versus observation) for every vehicle detection event. Three predictors (detector occupancy time, left-turn phases when the detector is on and off) from high-resolution data were used as input of the ML models. Results showed that all the models are statistically significant with p-value <0.05. The developed models were then applied to estimate hourly left-turn volume, by adding its prediction result to left-turn detection counts by the detector. The hourly left-turn volume estimated with the proposed method accounts for 96% of the actual left turns for low left-turn traffic conditions, and 93% for heavy left-turn traffic conditions. This indicates that the proposed method can provide a good estimation of hourly left-turn volume for signalized intersections operated with long-loop detectors. The method is simple and based on data from existing vehicle detection technologies already in typical signalized intersections.