Traffic Signal Classification with Cost-Sensitive Deep Learning Models

Abstract

Deep learning has many successful real-world applications including traffic signal recognition, which are used in driver assistance systems and autonomous vehicles. Accurate detection of traffic signal indications is critical to ensure safety under autonomous driving. Many past studies have been completed on traffic signal recognition including detection, classification and tracking with datasets which are typically highly imbalanced due to the nature of traffic signal displays. However, most studies simply ignored the minority classes and did not consider cost-sensitive information inherent to traffic signal indications. This paper evaluated several cost-sensitive techniques applicable to deep learning models in traffic signal classification. A convolutional neural network (CNN) was used in the evaluation as the baseline model. Cost-sensitive techniques including cost-proportionate rejection sampling and the use of cost-sensitive loss function was then applied to the baseline CNN model to evaluate and compare the effects of using cost information in traffic signal classification. Arbitrary cost information was assumed in this evaluation, but the resulting models did not improve accuracy in prediction. Future studies may consider more carefully crafted cost information and/or other cost-sensitive techniques.