Robust traffic light and arrow detection using optimal camera parameters and GPS-based priors

Abstract

In this paper, we propose a vision-based traffic light and arrow detection algorithm for intelligent vehicles. We detect all three traffic light colours along with the arrow direction robustly for varying illuminations and traffic lights. A fine-tuned convolutional neural network is used in an offline phase to localise the traffic light region-of-interest within a given camera image. Given the constrained region-of-interest, we detect the red, yellow and green traffic lights. The robustness and accuracy of the traffic light detection is significantly enhanced by identifying optimal camera parameters used during the scene perception. We identify the optimal camera parameters for varying traffic light types and illumination conditions. Apart from the traffic light detection, we also detect the arrow direction using a novel bounding box method. We perform a comparative analysis with baseline algorithms and demonstrate high detection accuracy with very few false positives in varying conditions.