Traffic Signs Detection and Segmentation Based on the Improved Mask R-CNN

Traffic signs detection and segmentation is one of the important parts of advanced driving assistance system. But there are predictable difficulties in detecting traffic signs from images or videos from car cameras owing to the next reasons: traffic signs are usually small-sized or medium-sized objects, and there is quantity imbalance between different traffic signs in the existed public data sets. Therefore, two main developments have been proposed in this paper. Firstly, an improved TT-100K-HHU traffic sign data set based on TT-100K is constructed. New images are collected from the Tencent Street View and labeled by Labelme software. Secondly, an improved Mask R-CNN is presented by revising the structure. More specific, feature pyramid network (FPN) is introduced into the backbone network of Mask R-CNN to achieve the fusion of feature maps at multiple scales, which can improve the representation abilities of network for objects with small or medium size. And in the prediction network, multiple cascaded Box Heads are applied to acquire more accurate location predictions and segmentation results. Experimental results show that the performance of the improved Mask R-CNN network is better than the existing algorithms.