IMPROVING TRAFFIC SIGN DETECTION BY COMBINING MSER AND LUCAS KANADE TRACKING

This paper presents the combination of Maximally Stable Extremal Region (MSER) and Lukas Kanade Tracking (LKT) for detecting traffic sign. The proposed approach employs the MSER to find the red circular sign in an image. Once the traffic sign is detected, it is tracked by LKT to predict the position in the next frame. The detected sign and predicted position are treated as the traffic sign candidates. To reduce the false positive, the traffic sign candidates are validated using a template matching technique based on the Histogram of Oriented Gradient (HOG) feature. Instead of using MSER and LKT individually, our approach combines both methods to find the best candidate according to the matching score obtained by the validation stage. The experimental results show that our proposed method provides the efficient detection with the Recall of 0.9353, the Precision of 0.9747, the Accuracy of 0.9331, and the frame rate of 47.950 fps.