DeLTR: A Deep Learning Based Approach to Traffic Light Recognition

Yiyang Cai, Chenghua Li, Sujuan Wang, Jian Cheng

Abstract: Traffic light recognition is crucial for the intelligent driving system. In the application scenarios, the environment of traffic lights is very complicated, due to different weather, distance and distortion conditions. In this paper, we proposed a Deep-learning based Traffic Light Recognition method, named DeTLR, which can achieve a reliable recognition precision and real-time running speed. Our DeTLR system consists of four parts: a skip sampling system, a traffic light detector (TLD), preprocessing, and a traffic light classifier (TLC). Our TLD combines MobileNetV2 and the Single Stage Detector (SSD) framework, and we design a small convolutional neural network for the TLC. To run our system in real-time, we develop a skip-frames technique and make up the delay of the time in the final response system. Our method could run well in complex natural situations safely, which benefits from both the algorithm and the diversity of the training dataset. Our model reaches a precision of 96.7% on green lights and 94.6% on red lights. The comparison to the one-step method indicates that our two-step method is better both in recall and precision, and running time’s difference is only about 0.7 ms. Furthermore, the experiments on other datasets (LISA, LaRA and WPI) show a good generalization ability of our model.