Traffic Sign Detection Algorithm Based On Multiple Features

With the development of society,scientific and technological progress,and the use of cars has become the main way for people to travel.The surge in the number of cars has led to traffic congestion and frequent traffic accidents,which has put enormous pressure on social traffic.In order to alleviate traffic pressure,protect people ’s living conditions and provide a more comfortable and safe living environment,smart driving is gradually being put on the agenda.The intelligent driving system avoids unnecessary traffic accidents by autonomously detecting and identifying road traffic conditions.As an important part of the road,traffic signs have become the primary research object of intelligent driving systems.How to efficiently and correctly detect and identify traffic signs in natural scenes has become a hot issue in current research.Aiming at the low accuracy of detection and recognition caused by small-scale traffic signs in real scenes,this thesis proposes an improved method based on YOLO v3 multi-scale traffic sign detection.The method calculates the corresponding receptive field size according to the feature map,and reasonably adjusts the scale of the feature map to make it more adaptable to the change of traffic signs at different scales in the real scene.At the same time,k-means clustering is used to obtain the a priori box size.Three kinds of prior frames are set for each downsampling scale,and 12 kinds of size prior boxes are clustered to improve the detection accuracy of smaller scale traffic signs.Due to the large scale change of traffic signs,the method uses four kinds of feature maps to detect traffic signs,and has good detection effects on traffic signs with various complex backgrounds or illumination situations.The experimental results show that the recognition rate of the proposed method is up to 98.77%,which is about 2.97%higher than other current deep learning methods,and the average is about 27.37%higher than the traditional method.In order to ensure the accuracy of small-scale traffic sign detection,this thesis proposes a small-scale traffic sign preference detection method.The method reduces the proportion of the coordinate error in the whole regression function by setting the coordinate error co-efficient,thereby avoiding the inaccuracy of the detection accuracy of the small-scale traffic detection target by the large-scale detection frame.The experimental results show that by adjusting the IOU reduction ratio coefficient,the recognition rate increases by 0.05%,and the detection effect of small-scale traffic signs has been improved.The Recall value and Pre-cision value are 5%higher than FastR-CNN.Compared with YOLO v3,the area of curve is significantly improved.The comprehensive research results show that the improved method of multi-scale traf-fic sign detection can significantly increase the detection and recognition rate of traffic signs.The multi-feature traffic sign detection method with increased small-scale preference can ef-fectively improve the detection and recognition rate of traffic signs,and has better robustness for the detection and recognition of small-scale traffic signs.