Traffic Light Detection by Fusing Object Detection and Map Info
Name
Tetiana Shtym
Abstract
To share streets with human drivers, self-driving cars must locate traffic lights and recognize their states. While for human drivers recognizing a relevant traffic light does not require much effort, it is a challenging task for self-driving cars. Although, for state-of-the-art object detection methods detecting traffic lights is simple, identifying to which lane they apply is non-trivial.
The most common approach relies on precise locations of traffic lights on the high-definition map, localization of the car, and camera position with respect to the car. When a vehicle approaches a traffic light, traffic lights from HD-map are projected to the camera images. Then, regions that include traffic lights, or ROI for traffic lights are extracted and fed to the classifier. To mitigate localization errors, ROI need to be enlarged. However, this can lead to imprecise classification as the bounding box might not capture traffic light adequately.
In this thesis, the problem is addressed by introducing traffic light recognition by fusing object detection and HD-map information. The process is divided into three phases: get 2D traffic lights' ROI by projecting 3D bounding boxes from the map to the camera image; perform traffic light detection on the image to get 2D bounding boxes and traffic light states; associate traffic lights with lanes by matching detected bounding boxes with ROI using Intersection-over-Union metric.
The proposed method was integrated into Autoware.AI and tested on prerecorded routes in Tallinn and Tartu. The approach achieved an accuracy of 93% and outperformed the approach currently used by Autoware.AI.
The most common approach relies on precise locations of traffic lights on the high-definition map, localization of the car, and camera position with respect to the car. When a vehicle approaches a traffic light, traffic lights from HD-map are projected to the camera images. Then, regions that include traffic lights, or ROI for traffic lights are extracted and fed to the classifier. To mitigate localization errors, ROI need to be enlarged. However, this can lead to imprecise classification as the bounding box might not capture traffic light adequately.
In this thesis, the problem is addressed by introducing traffic light recognition by fusing object detection and HD-map information. The process is divided into three phases: get 2D traffic lights' ROI by projecting 3D bounding boxes from the map to the camera image; perform traffic light detection on the image to get 2D bounding boxes and traffic light states; associate traffic lights with lanes by matching detected bounding boxes with ROI using Intersection-over-Union metric.
The proposed method was integrated into Autoware.AI and tested on prerecorded routes in Tallinn and Tartu. The approach achieved an accuracy of 93% and outperformed the approach currently used by Autoware.AI.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Tambet Matiisen, Meelis Kull
Defence year
2021