Real-time Traffic Signs Recognition using MSERs and CNN
Eunsoo H Park, Cheolyong Jang, Byeonghak Lim, Hakil Kim · Journal of Institute of Control Robotics and Systems · 2017
This paper proposes a traffic sign recognition algorithm that is robust in various environments. Color information is an important element in the traffic sign recognition system, as the performance depends on variations in weather conditions, illumination, and type of cameras used. Besides the above factors, traffic signs also differ across countries. To overcome these problems, our approach involves traffic sign detection, classification, and tracking. In the detection module, color enhancement with maximally stable extremal regions is performed to improve the extracting candidate regions of traffic signs. Support vector machine classifiers with distance to border and histogram of oriented gradient feature vectors are used to detect the traffic signs. Detected traffic signs are thereby classified using convolutional neural networks with fine-tuning. Additionally, Kalman filter-based multi-target tracking not only verifies traffic sign detection but also optimizes the detection of regions of interest. The result of traffic sign detection is 95.67% when trained on the Belgium Traffic Signs Dataset for Detection (BTSD) training dataset and tested on the Germany Traffic Signs Detection Benchmark test dataset. Moreover, while using the BTSD training dataset, the area under the curve of our method is 89.56%. In classification, the performance of INHA Traffic Signs Classification is increased to 97.48% by adding transfer learning.