Self-driving vehicles are being tested to make them more road-ready and safer for a real traffic environment. Automobile giants: like Tesla, Waymo, Toyota, etc., are working exhaustively to cater to the needs of futuristic smart vehicles using deep learning methodologies. For a self-driving car to avoid collision situation, it must be able to accurately detect and classify traffic lights. After performing exhaustive experiments, we chose to compare the feature extraction capabilities of various pretrained CNN-based transfer learning models like VGG16, ResNet50, AlexNet, DenseNet121, InceptionV3, and Xception on freely available Lara & Lisa traffic light datasets. We segregated the Lisa traffic light dataset into day and night subsets and then manually separated the images into various traffic light classes like dayRed, dayYellow, nightYellow, nightGreen, road and traffic lights, LeftGreenArrow, and RightGreenArrow etc. We used the random forest classifier to identify the color of the detected traffic light. DenseNet121 achieved a top accuracy of 100% on the Lara traffic light dataset and the maximum accuracy of 99.88% on the Lisa Day-light dataset and 98.89% accuracy on the Lisa Night-light dataset.
Automatic Traffic Light Detection for Self-Driving Cars Using Transfer Learning
Lect. Notes in Networks, Syst.
2022-01-03
10 pages
Article/Chapter (Book)
Electronic Resource
English
Traffic Light Detection System in Self-Driving Cars
IEEE | 2023
|Deep learning using computer vision in self driving cars for lane and traffic sign detection
BASE | 2019
|