Speed sign recognition in complex scenarios based on deep cascade networks

摘要

Speed sign is one of the most important instant indications for drivers to adjust the speed of their cars. In the literature, almost all of the existing methods for speed sign recognition are based on static pictures with clean images. When dealing with the traffic signs in complex environments, these existing approaches often have inaccurate detected areas, which lead to the difficulty of recognition. The authors propose a deep cascade network to improve the recognition of the speed signs with a structure of cascade subnetworks. The proposed network is composed of a localisation subnetwork and a classification subnetwork. The difficult issue in complex scenarios is the detection of the speed sign due to its small resolution, occlusion, colour fading etc. The proposed localisation subnetwork can improve the localisation accuracy by borrowing the idea of locating the targets from coarse to fine. Ultimately, the classification sub-network extracts more effective features for speed sign recognition. The experimental results illustrate that the proposed method outperforms the YOLOv2 or YOLOv3 model in identifying the speed sign in complex scenarios with at least 6% higher in terms of area under curve, and this will promote the improvement of recognition significantly.