Fast Sign Recognition with Weighted Hybrid K-Nearest Neighbors Based on Holistic Features from Local Feature Descriptors

Abstract

Sign recognition is crucial not only for road asset inventory but also for intelligent vehicles. Fast and robust sign recognition is still an open problem especially in varying and complex road environments. Speeded-up robust features (SURF) and oriented FAST and rotated BRIEF (ORB) are two key point detectors and local feature descriptors widely used for image feature point matching. In this paper they are both used to compute sign holistic features from their local feature descriptors. A novel method called weighted hybrid K-nearest neighbors (WH-KNN) is proposed to fuse the extracted holistic features for fast and robust recognition. The proposed method can achieve less than 1.4% false negative rate and less than 0.2% false positive rate for all the three datasets. On average it took less than 1.5 ms for holistic feature extraction and less than 0.5 ms for sign feature matching on a low-profile laptop with a 2.4 GHZ CPU and 4 GB RAM. The results from three data sets demonstrate that the proposed method is accurate and fast for real-time road sign recognition.