|
Construction and Use of Ground Truth Tool for Drivezone,
Lane Marking, and Hazard Classification in Roadway Video
Geoff Wright & Seth Teller
Abstract
The MIT DARPA Grand Challenge team, based at CSAIL, is building a large
perception codebase with a number of different algorithms related to finding
road lane markings and obstacles from video, radar, and LIDAR sensor data.
The team has also collected several logged data sets comprising video,
radar, LIDAR and IMU data. Our project has two aims:
- To create an efficient tool to label the logged video data with "ground
truth" locations of lanes, vehicles, hazards, drivable zones, and pedestrians
on a per-frame basis; and
- To use the tool to create labels for a number of logged videos.
Specifically, the types of items that are labeled by the tool are: Lane
Single, Lane Double, Lane Crosswalk, Lane Stopline, Lane Other, Vehicle
Stopped, Vehicle Approaching, Vehicle Retreating, Vehicle Other, Hazard
Up, Hazard Down, Hazard Other, Drivezone Smooth, Drivezone Rough, Drivezone
Other, Artifact Saturation, Artifact Occlusion, Artifact Other, Pedestrian
Stopped, Pedestrian Crossing.
The tool saves the labeled data to a text file in xml format according
to a specified grammar. Saved data includes the location of the nodes
that represent each item, in pixel coordinates, and associates each item
with a particular frame of the logged video. The tool also has functionality
to edit existing labeled data, remove or change labels, and to skip through
the video file either by single frames, ten frames or to skip to the next
labeled frame. The GUI superimposes labeled items onto the logged video
during viewing as shown in the screenshot below.
|

 |