Traffic Sign Recognition Based On Scene Understanding And Deep Learning

The acceleration of modernization and the increase in the number of vehicles have made the problems of road traffic safety and transportation efficiency,such as intensified traffic congestion and frequently occurring traffic accidents,increasingly prominent.The road system of traffic sign recognition(TSR)is an important part of an intelligent vehicle.This system collects and recognizes the information of the traffic signs appearing on the road in the course of driving,gives directions or warnings to drivers in time,and controls vehicles directly to ensure smooth traffic flow and prevent accidents.The existing vision-based traffic sign recognition methods have the following problems:(1)Existing methods focus on extracting features of traffic signs and ignore the constraints of spatial positional relationships between traffic signs and other objects in the scene.This way results in incorrectly detecting other similar objects as traffic signs and failing to detect very small traffic signs.(2)The existing methods have poor recognition performance for small traffic signs.Aiming at the problem that the existing methods are prone to misidentification,this paper proposes a candidate region extraction algorithm based on scene analysis.This method first uses semantic segmentation framework DepLabv3+to perform semantic segmentation for scene objects in the driving environment,including road,traffic sign,sidewalk,building and other target segmentation.In order to eliminate similar objects in the scene,a scene structure model based on the constraints of spatial positional relationships between traffic signs and other objects is proposed to establish trusted search regions.According to the rules for placing traffic signs in real word,three search areas are defined in the scene structure model,which are located on the left,right,and above the road.Location information can be used to filter out erroneous candidates outside the scene structure model and retain valid traffic sign candidate regions.Aiming at the problem that existing methods performed poorly on small target recognition tasks,a new network multiscale densely connected object detector(MDCOD)based on densely connected style,multiscale feature fusion,and Jaccard K-means clustering is proposed in this paper.This model constructs Detail-attention DenseNet as the base network,and introduces dense connections to overcome the problems of gradient disappearance and excessive parameters in deep networks.Multi-scale feature fusion is accomplished by method of learning half and reusing half,which greatly improves the shortcomings of insufficient shallow features in the feature extraction process.In the network,ten feature maps are used for objects prediction,which are suitable for the task of small object recognition.In the convolution prediction,the aspect ratio of default box close to the ground truth will help improve the accuracy of traffic sign location.To this end,this paper proposes a Jaccard K-means algorithm to select the best aspect ratio for the default box for ensuring that their scale meets the distribution rules of traffic signs.The above methods have been comprehensively tested in this paper.Some of the results have been published in SCI journal papers.Experimental results show that the proposed candidate region extraction algorithm based on scene understanding can distinguish traffic signs and similar targets well.The proposed method is tested on Tsinghua-Tencent 100 K and German Traffic Sign Detection Benchmark datasets and achieves accuracies of 92.3%and 99.90%,respectively,outperforming the existing methods.