Naturalistic Driving Study Data for Signal State Detection Model Training

Project Description This collection was developed to be an input source for the identification of traffic signal heads and signal states using computer vision. A collection of signal annotations was compiled on a frame-by-frame basis for a small subset of intersection traversals. Annotations were completed on 63 intersection traversals and just over 8,100 frames of video using the Computer Vision Annotation Tool (CVAT). For a given frame of video, the annotation of a signal consists of a bounding box placed around the perimeter of any object that was a signal or closely resembled a signal. Objects closely resembling signals were recorded in order to reduce or eliminate false positive detections by allowing models to distinguish between them. Annotations begin when the signal head(s) can be clearly discerned, typically near the stop bar, and continue until the signal head(s) are no longer visible in the forward video view. Each bounding box had two associated attributes to denote the type or class of object and the signal state. Candidate objects for bounding boxes include signals in the vehicle travel directions and signal state, signals not in the travel direction of the vehicle and signal state, and objects resembling signals that are not signals. Data Request Scope and Specification This data set includes H.264 encoded MP4 forward video for each intersection traversal as well as signal annotations in csv and xml formats. The data files include the size of the video image, the pixel coordinates for all the bounding boxes, and a binary indicator of object occlusion. Pixel coordinates are based on an upper left corner origin.