Traffic Counting System Using Machine Vision

Overview

Accurate vehicle counts on roadways are important for traffic departments and transportation agencies for transit planning, road maintenance, lane expansion, traffic alerts, and traffic light control. Prevalent methods such as pneumatic tubes, inductive coils, piezoelectric strips, and optical beams are disruptive, non-concurrent, inaccurate, and temporary. A statistically based machine intelligence approach using video cameras capturing traffic, that is highly accurate and autonomous, is developed. The computationally efficient process is nearly real-time and can be realized on an edge device co-located with the camera. The crux of the algorithm involves identifying a vehicle presence indicating a statistical signature in fixed partitioned areas in each of the lanes. The lane-wise traffic counts are incremented when the statistical metric in the partitioned areas exceeds certain dynamic thresholds. When applied to traffic video captures from vantage points above roadways under various lighting conditions, an overall traffic count accuracy of 96%, that is comparable to, or better than the published state-of-the-art performance, was achieved when counting about 2500 vehicles. The same algorithm and thresholding formula was used for different roadways. This method does not rely on horsepower-intensive and unnecessarily complicated image segmentation, machine learning-based vehicle identification, and motion tracking steps. Moreover, it can be mapped to software implementation on edge-computing embedded platforms, centralized traffic operation centers, or the cloud, enabling widespread deployment to cover the entire roadway network accurately, concurrently, continuously, autonomously, and economically.