Research On Multi-scale Traffic Sign Detection Based On Semantic Segmentation

Since the advent of driverless cars,it has been widely concerned by domestic and foreign automakers.A reliable visual perception system is an important support for driverless technology.Traffic signs carry important road driving information.the system that detects traffic signs accurately is conducive to guiding the driver to operate the vehicle safely,and have a great significance for improving the ability of scene understanding of driverless cars.Under the promotion that the hardware computing ability is improved and a large number of scholars devote to relevant researches,deep learning has developed rapidly in recent years.Especially in the field of computer vision,deep convolutional neural networks create many breakthrough achievements.Research methods based on convolutional neural networks occupy an important position in visual tasks such as image recognition,object detection and image semantic segmentation.This paper focuses on the application and improvement of image semantic segmentation technology in traffic sign detection tasks.(1)In real detection scene,the scale of traffic signs that appear on images varies.Although image semantic segmentation based on convolutional neural networks can predict objects at the pixel-level,how to accurately segment multi-scale object has always been a difficulty of research.An effective scheme that enhance the segmentation performance is to fuse multi-scale feature generated by convolutional neural network.However,information of some scale features is important component for semantic representation of a particular scale object,information of other scale features affects features expression slightly.Therefore,based on the Deep Lab segmentation framework,a scale feature weighting method is introduced to weight the multi-scale features generated by the atrous spatial pyramid pooling module,so that the convolutional neural network can adaptively select specific scale features and effectively improve features representation and pixel prediction.The experiment demonstrates that the scale feature weighting strategy is beneficial to the performance improvement.(2)Aiming at the time-consuming problem of image segmentation process in which the traffic sign detector based on fully convolutional network extracts the region of interest(ROI),a two-stream decoder network is proposed for semantic segmentation of ROI.The speed of image segmentation is improved by incorporating a lightweight network named ENet as the backbone of the network,In the upsampling stage,a one-step decoding branch is constructed for decoding the information of multi-scale features layer.Combining multi-step and one-step upsampling implements two-stream decoding way.The prediction result of ROI is fused from the output of different discriminative decoding branches.segmentation accuracy is improved.Region proposal methods is guided to the ROI to generate more accurate detection boundary.A classifier based on convolutional neural network is used to recognize the traffic sign in the region.The experimental results show that the traffic sign detector integrating with proposed network segments ROI with 23 ms per image,and maintains a high detection accuracy.