Detection And Recognition Of Road Signs And Lane Lines Based On Neural Network

The input of external road information is the source of decision-making and control of the automatic or assisted driving system,and ensuring the quality of the input information is the guarantee for the normal operation of the system.Therefore,extracting meaningful road condition information from the road environment has always been a research hotspot in the field of autonomous driving.This paper focuses on the task of road sign recognition and lane line detection.The main work is as follows.First,we researched the recognition of traffic signs,and developed a two-stage road sign detection method based on cascaded classifiers is implemented.This method uses different sizes of windows to slide to detect road scenes to extract Haar features,and import to the strong classifier which constituted by multiple weak classifiers.The strong classifier recognizes the positive and negative sample and provides the target area of interest for the subsequent cascaded convolutional neural network.The convolutional neural network is responsible for the identification of the type of road signs.This method has achieved better recognition results on the DF platform virtual environment traffic sign data set.Then the YOLOv3 single-stage target detection network is the best in industrial applications currently.We improved its backbone to make it have better small target detection capabilities,it improves 2.33 percent.For the TT100K traffic sign data set,the image size is too large and not conducive to training.The target is too small and the target clustering is high,an adaptive random image clipping method for this data set is proposed,and it has a certain data enhancement effect.Finally,the DFANet semantic segmentation network was simplified to achieve a faster separation of road signs and non-road areas.An instance segmentation network structure based on ERFNet is also proposed,which can segment and instantiate lane lines.Compared with methods such as SCNN,it has a slight loss of accuracy but has a faster detection speed.And after the semantic segmentation structure,the lane line classification neural network is cascaded to achieve 96% accuracy of virtual and real lane line classification.