Deep detection network for real-life traffic sign in vehicular networks

The challenge for real-life traffic sign detection lies in recognizing small targets in a large and complex background, making state-of-the-art general object detection methods not work well in both detection speed and precision. The existing deep learning models for traffic signs detection fail to use the fixed feature of the targets. This paper proposes a novel end-to-end deep network that extracts region proposals by a two-stages adjusting strategy. Firstly, we introduce an AN (Attention Network) to Faster-RCNN for finding all potential RoIs (Regions of Interest) and roughly classifying them into three categories according to colour feature of the traffic signs. Then the FRPN (Fine Region Proposal Network) produces the final region proposals from a set of anchors per feature map location extracted by the AN. We also modify the model by (1) adding a deconvolutional structure to convolutional layers to fit the small size of targets, and (2) replacing the classifier with three softmax corresponding to three coarse categories obtained by the AN. Our method is evaluated on two publicly available traffic sign benchmarks which are collected in real road condition. The experiments show our method generates only 1/14 of the anchors generated by Faster-R-CNN so the detection speed is increased by about 2 fps with ZF-Net and it reaches an average mAP of 80.31% and 94.95% in two benchmarks, 9.69% and 7.88% higher than Faster-R-CNN with VGG16, respectively.