Single-and Multi-Sensor Techniques to Improve Accuracy of Urban Vehicle Classification

Vehicle traffic sensors are an essential part of urban traffic management systems. However, humans are still widely used when short-term traffic classification is required, because current traffic data collection systems do not meet requirements for high accuracy, rapid deployment, and low cost. This paper explores the use of a sensor network for vehicle traffic classification. We accomplish low cost and easy deployment by using simple embedded computers and inductive loop sensors that can be taped directly to the roadway. We explore several techniques to improve accuracy both at individual sensors and by combining readings from multiple sensors. Traditional approaches typically assume high-speed traffic; we instead focus on errors that are induced at low and varying speeds, since we expect sensors to often be deployed at locations where vehicles will slow or stop. The contributions of this paper are first, a comprehensive evaluation of approaches to reduce error: from feature definition, combining readings from a single sensor, and combining results from multiple independent sensors. Second, we evaluate these approaches against both on-line and off-line human observations, demonstrating sensor accuracy better than or slightly worse than on-line human classification depending on the similarity of the categories, and nearly optimal for length-based classification.