Small Traffic Sign Detection Based On Region Convolution And Weak Edges

Autonomous traffic sign detection and recognition is significant in autonomous driving.As a preliminary technique of traffic sign recognition,traffic sign detection plays an important role in intelligent decision making.Nowadays,computer vision-based object detectors perform well on public datasets.However,traffic sign detection in crowded urban scene is challenging,since most sign instances are small with distortion,occlusion and blur.Additionally,complex foreground background and real-time speed requirement make traffic sign detection more difficult.Recently,deep learning-based methods for image classification,object detection and edge detection have achieved significant progress.Hence,the region convolution networks and edge detection networks are employed in this paper,and proper modification has been proposed to detect small traffic signs.Experimental results demonstrate that the proposed methods achieve superior detection performance,with 50% AP on BDCI dataset and considerable computational consumption.Main work and contributions are listed as following.1.To improve RPN’s mediocre detection performance on small traffic sign,number and percentage of positive training samples are increased to augment region proposals’ quality,via proposed shrinking of reference window size and resolution zooming factor from input to last convolution layer.2.To reduce the model size and running speed of Faster R-CNN,layer’s identical transformation from fully-connected to convolution is employed,with dilated convolution and channel sampling tricks.By doing so,both model size and running speed can be reduced with acceptable accuracy.3.To detect traffic sign in crowded scene,RCF as weak edges network for multi-instance region proposal generation is proposed,and an evaluation metric is presented to evaluate multi-instance region proposal.4.For multi-shot on each multi-instance region,RPN is utilized to train a localization model.5.To refine the region proposals from RCF-RPN network,a fast R-CNN is employed on each single instance region proposal,forming the complete weak edges guided object detector,which can naturally deal with small and common scale traffic sign instances.Based on the above improvements,the proposed algorithms boost the performance of small traffic sign detection and keep an affordable computational efficiency.It is also promising to be used on general small object detection.