Cross-Model Temporal Cooperation Via Saliency Maps for Efficient Recognition and Classification of Relevant Traffic Lights

Recognizing the state of relevant traffic lights when a vehicle is approaching an intersection is a challenging task for advanced driver-assistance systems (ADAS) that requires the continuous processing of frames recorded by onboard cameras. This type of task can be energy-intensive and may result in significant battery consumption and overheating. To address this problem, we propose a novel architecture, called Cooperative Frame Classifier (CFC), that efficiently exploits temporal redundancies between consecutive frames to maintain high performance while greatly reducing the average number of operations needed to process the video frames. The architecture consists of two convolutional neural network models with different parameter sizes and input resolutions. Each frame is processed by only one of the models. The model with lower input resolution and parameter size leverages the saliency maps generated by the larger model. Experiments show that our approach is not only more accurate than existing ones based on object detection, but it is also an order of magnitude more efficient, and therefore it is better suited for practical applications.