Validating the accuracy of sensors and methods is an essential step in the collection of traffic speed data. The accuracy of automated speed data has been evaluated in both small- and in large-scale testing efforts using multiple technologies and methods, as documented in existing literature. In these studies, an important challenge is the creation of a ground truth speed data set that represents actual traffic history. Though inductive loops are standard for data collection, the use of non-intrusive traffic data collection technologies has become increasingly popular. Video-based detectors have demonstrated the ability to substitute conventional detection devices. Computer vision systems and video tracking software provide a wide variety of data, including conventional traffic parameters such as flow and velocity, while preserving a complete record of events. Though existing literature documents several issues associated with extracting vehicle speeds from video, the analysis of speed data, especially at the microscopic or individual level has been limited. The purpose of this paper is to evaluate the accuracy of a video-based detection system, comprised of commercially available video cameras and the Traffic Intelligence video analysis software system. To provide robust calibration, several camera orientations were tested along two types of facilities in Montreal, Canada. Video was collected on an urban arterial and a highway section, with cameras oriented both perpendicular and parallel to traffic direction. After calibrating the feature-tracking software, a semi-automated vehicle tracking process was used to extract the vehicle speeds. Comparison to manually observed speeds was undertaken to evaluate the quality of the extracted speeds. Although the traditional mean error approach led to unacceptable results, a new approach was proposed for the evaluation of traffic detection technologies. The proposed segregated error approach divides the mean error values into separate values representing accuracy and precision errors. In doing so, several of the camera orientations exhibited precision error values within the accepted range speed data quality (5%). Even with large errors, video data can be calibrated to acceptable levels of accuracy so long as precision error is minimized through appropriate selection of camera position and orientation.
