Research and Implementation of Mountain Road Traffic Sign Identification System

• Main Authors: WU, SHIUE-LING, 吳雪玲
• Other Authors: WU, CHIEN-CHUNG
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/kctp49

碩士 === 南臺科技大學 === 資訊工程系 === 106 === It is helpful to enhance traffic safety by the driving information which is provided by the roadside traffic signs. However, as the traffic signs information cannot be seen immediately while driving, drivers always ignore a lot of important driving information. In order to solve this problem, an automatic traffic sign identification system which takes mountain road as an example is proposed. The system is developed with image recognition technology and can provide the identification result of related driving information to drivers. The system proposed in this study is divided into two parts. Firstly, through the brightness of the image, the possible traffic sign image and its position can be quickly detected with the classifier trained by Haar Cascade in Haar-like and AdaBoost method. The second, the detected images can be classified by the convolutional neural network (CNN) classifier design with GoogLeNet Network Architecture in Tensorflow library and Python program language. If the top-1 class recognized result of possible traffic sign image probability is above the threshold, the result will be displayed on the screen for drivers to read. To speed up the recognition process and the correctness of recognition result, a method is proposed in this study to avoid analyzing those areas without traffic signs. In the method, the collected mountain road videos are divided into 12 regions to find the non-traffic sign areas. After this method, the recognition speed is improved 11.24 frame/sec from to 12.04 frame/sec. Besides, through the method proposed in this study, the precision of limit detection area can be improved from 80.9694% to 86.3612% by Haar-like method. Through the system proposed in this study, the precision can be improved from 97.4720% to 98.0052% with the method composed of Haar-like and CNN.