Real-time Traffic Light Detection and Recognition based on Deep RetinaNet for Self Driving Cars

Self-driving cars are getting more popularized nowadays due to its safe, convenient and congestion free transportability. Most of the real-time challenges for autonomous driving like recognizing traffic lights, traffic signs, pedestrians are being accurately addressed by the newer state-of-the-art algorithms based on Deep Learning. Recent technological advancements in cloud computing and the availability of high-end cloud-based Graphical Processing Units (GPU) accelerated the development of AI algorithms substantially. For real time detection and recognition of traffic lights, we propose RetinaNet (a deep neural network architecture) based model through transfer learning. The deep neural network RetinaNet was used as model and the system was implemented in Keras with TensorFlow backend in Google Colaboratory cloud platform. The RetinaNet model was trained and evaluated on Bosch Small Traffic Light Dataset containing traffic light images of resolution 1280 by 720 pixels, which falls under four type of classes. The model achieved improved accuracy of detection and classification than other deep learning methods for real-time operation.