Table 4. The Recognition Accuracy of Each Technique in Our Model.
CHAPTER 5
CONCLUSION
The proposed algorithm includes ROI extraction, HOG feature design, and stacked CapsNet. We detailed the component of HOG feature, which could effectively obtain the pixel-level features of the image from any angle, and speed up the learning of the model. Experimental results on a real-world traffic sign dataset show that our model ran faster, occupied less memory, and required fewer parameters than baseline methods. The superiority over the classic CNN comes from the CapsNet, which fully utilizes images of different angles and directions with the aid of vectors.In future we will increase the number of hard negative samples during the training stage and expand the GTSDB training dataset by generating real-world pictures containing traffic signs in situations such as lighting, weather changes. And also we will try to train the model to detect the traffic sign in video.
REFERENCES
RESEARCH PAPER
By XIE BANGQUAN , AND WENG XIAO XIONG
https://ieeexplore.ieee.org/document/8698449
By JIANMING ZHANG ,ZHIPENG XIE, JUAN SUN , XIN ZOU, AND JIN WANG
https://ieeexplore.ieee.org/document/8986614
By RUI CHEN , LEI HEI , AND YI LAI
https://ieeexplore.ieee.org/document/9233325
By ZHIGANG LIU , JUAN DU , FENG TIAN , AND JIAZHENG WEN
https://ieeexplore.ieee.org/document/8701699
By SHOUHUI HE , LEI CHEN , SHAOYUN ZHANG , ZHUANGXIAN GUO , PENGJIE SUN , HONG LIU , AND HONGDA LIU
https://ieeexplore.ieee.org/document/9353477
Share with your friends: |
