TACS: A Calibrated Highway Surveillance Dataset for Traffic Analysis

Access to live streams of traffic surveillance cameras is becoming more common across the world. Despite this, many current traffic analysis algorithms do not utilize such cameras; rather, they rely on other sources of data. This can be attributed to the arduous process of extracting the necessary traffic information for such algorithms. The TACS (Traffic Analysis on Calibrated Streams) dataset aims to remove this barrier by annotating sig-nificant traffic information on realistic traffic surveillance video. We collect live video from the Virginia Department of Transportation and perform accurate camera calibration. After calibrating the camera, we obtain a top-down view of the video stream and extract the scene scale. From this, we calculate the real world position and speed of each vehicle. On top of such information, we calculate vehicle lane positions and two types of road congestion metrics. We collect between 10,000 and 15,000 frames from each of its 10 traffic video sources. Thus, the TACS dataset contains a wide variety of data and provides researchers with rich annotations for traffic-related algorithms.