IoT as a Service. 5th EAI International Conference, IoTaaS 2019, Xi’an, China, November 16-17, 2019, Proceedings

Research Article

Traffic Lights Detection Based on Deep Learning Feature

Download
58 downloads
  • @INPROCEEDINGS{10.1007/978-3-030-44751-9_32,
        author={Changhao Wang and GuanWen Zhang and Wei Zhou and Yukun Rao and Yu Lv},
        title={Traffic Lights Detection Based on Deep Learning Feature},
        proceedings={IoT as a Service. 5th EAI International Conference, IoTaaS 2019, Xi’an, China, November 16-17, 2019, Proceedings},
        proceedings_a={IOTAAS},
        year={2020},
        month={6},
        keywords={Traffic lights detection Deep learning Region proposal Classification},
        doi={10.1007/978-3-030-44751-9_32}
    }
    
  • Changhao Wang
    GuanWen Zhang
    Wei Zhou
    Yukun Rao
    Yu Lv
    Year: 2020
    Traffic Lights Detection Based on Deep Learning Feature
    IOTAAS
    Springer
    DOI: 10.1007/978-3-030-44751-9_32
Changhao Wang1, GuanWen Zhang1,*, Wei Zhou1,*, Yukun Rao1, Yu Lv1
  • 1: Northwestern Polytechnical University
*Contact email: guanwen.zh@nwpu.edu.cn, zhouwei@nwpu.edu.cn

Abstract

Traffic lights detection is an important task for intelligent vehicles. It is non-trivial due to variance backgrounds and illumination conditions. Therefore, a traffic lights detection system that can apply to different scenes is necessary. In this paper, we research the traffic lights detection based on deep learning, which can extract features with representation and robustness from input image automatically and avoid using artificial features. The approach of traffic lights detection proposed in this paper includes two stages: (1) region proposal and (2) classification of traffic lights. Firstly, we propose a region proposal method based on intensity, color, and geometric information of traffic lights. Secondly, convolutional neural network (CNN) was introduced for the traffic lights classification, obtaining 99.6% average accuracy. For detection, we evaluate our system on 6804 images of different scenes, the recall and accuracy of detection achieve 99.2% and 98.5% respectively.