Fast Road Sign Recognition Based on ORB-Encoded Holistic and Local Features
Zhaozheng Hu, Yuezhi Hu, Yicheng Li · Transportation Research Board 95th Annual MeetingTransportation Research Board · 2016
ORB (Oriented FAST and Rotated BRIEF) is a fast and robust local feature detector with excellent performance in image feature point matching. However, in this paper, ORB is used not only to encode sign local features but also holistic features. Based on the ORB-encoded holistic and local features, the authors develop a very fast and accurate method for sign recognition. In this method, each input sign image is resized into two standard images with the resolutions of 63×63 and 200×200 (in pixel), respectively. From the small-size image, the image center is used as the ORB point position. The ORB descriptor is thus extracted as sign holistic features. From the large-size image, ORB points are localized automatically and the corresponding descriptors are extracted. By using the classic Bag-of-Word (BOW) method, ORB descriptors are further encoded with visual words from the BOW dictionary as sign local features. The holistic and local features are finally fused by proposing a novel method Hybrid K- Nearest Neighbors (H-KNN), for sign recognition. The proposed method has been validated with both public dataset and dataset collected in the field. It can achieve over 96.5% and 94.0% recognition rates on these two datasets, respectively. In average, it took less than 19ms for feature extraction and less than 1ms for sign type matching on a low-profile laptop computer with 2.1GHZ Intel I3 CPU and 2G RAM. The results demonstrate that the proposed method is both accurate and fast for real-time sign recognition.