Traffic light detection and recognition in autonomous vehicles

Abstract

Traffic Light Detection and Recognition for Autonomous Vehicles Autonomous Vehicles (AVs) face the challenge of recognising active traffic lights under harsh environmental conditions. Standard cameras and computer vision algorithms also face the same challenge. In this paper, we built a small-scale system to mitigate this challenge. First, we developed a light controller and a dataset builder script. The light controller and dataset builder script were then used to build a dataset of traffic lights with different lights activated. Bounding boxes were annotated on the traffic light dataset using dlib's imglab software. The dataset uses the HOG with Linear SVM object detector. An RGB histogram approach is adopted to train a logistic regression model on the feature vector data to recognise which light is "on" among the training images. Finally, a robot control script is developed and tested. The script uses both the object detector and colour recogniser for its detection and recognition. Our results show 89% acc