Traffic Light Detection in Autonomous Driving Vehicles
Abstract
Autonomous driving is an important field of research, especially now since the world is moving away from gas-operated vehicles, and towards electric vehicles. Since the operation of these vehicles heavily depends on the algorithm running behind the scenes, it’s imperative to ensure that these algorithms are highly accurate and efficient in detecting obstacles on the roads (pedestrians, construction signs, etc.) and making smart driving decisions (stopping, speeding up, turning, etc.). Although there has been prior work in detecting traffic lights, the existing algorithms are either not efficient enough to run in real-time or require high-power computing capabilities. There also isn’t much prior work around detecting flashing lights, especially since their inconsistent frequency makes the problem more challenging. Hence, the goal of this thesis is to create an efficient and accurate real-time detection algorithm for traffic light signal state, and its color and shape. It also aims to determine if the detected traffic light is flashing. This research explores the domain of autonomous driving and traffic light detection, investigates potential solutions to tackling these challenges, and implements and tests the most efficient approach. The algorithm resulting from this research will be tested for feasibility with real-time inferences, as well as compatibility with the Robot Operating System (ROS). By testing with various public open-source datasets, as well as images collected from the local Bryan/College Station area, the proposed neural network will be trained and tested against various types of traffic lights.
Citation
Srivastava, Dakshika (2023). Traffic Light Detection in Autonomous Driving Vehicles. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /199658.