Accurate traffic light detection using deep neural network with focal regression loss

Abstract

This paper proposes a method that uses a deep neural network (DNN) to detect small traffic lights (TLs) in images captured by cameras mounted in vehicles. The proposed TL detector has a DNN architecture of encoder-decoder with focal regression loss; this loss function reduces loss of well-regressed easy examples. The proposed TL detector has freestyle anchor boxes that are placed at arbitrary locations in a grid cell of an input image, so it can detect small objects located at borders of the grid cell. We evaluate the proposed TL detector with a focal regression loss on two public TL datasets: Bosch small traffic light dataset, and LISA traffic lights dataset. Compared to state-of-the-art TL detectors, the proposed TL detector achieves 7.19%–42.03% higher mAP on the Bosch-TL dataset and 19.86%–49.16% higher AUC on the LISA-TL dataset.