Research On Multi-type Traffic Signs Detection Based On Deep Learning

With the rapid development of artificial intelligence and the advent of the 5G era,people’s requirements for the safety,intelligence,and operability of automobiles are constantly increasing.The control methods of traditional cars are undergoing tremendous changes.People try to develop assisted driving systems and driverless cars based on the traditional automobile industry.Assisted driving or unmanned driving can be achieved through the control of the computer system.In this way,the mental burden of the driver is reduced,the occurrence of traffic accidents is reduced,the safety of the vehicle is improved,and the driver is relieved from the heavy driving work.Whether it is driver assistance or self-driving cars,it is inseparable from the detection of the environment outside the cab.Because the traditional algorithm needs to manually extract the color,edge,texture and other features of the traffic sign when detecting the traffic sign.This makes this kind of method slow in detection and poor in robustness,making it difficult to accurately detect traffic signs in complex real traffic scenarios.In order to achieve timely and accurate detection of multiple types of traffic signs such as traffic signs,traffic lights and stop lines,this paper proposes multi-class traffic sign detection algorithm T-YOLO and TM-YOLO based on the deep learning object detection algorithm YOLO v3 algorithm.The algorithm first uses the Darknet-53 convolution network,and adds several convolutional layers and 3 upsample layers at the end of the Darknet-53 network to form a feature extraction network to complete multi-class traffic sign feature extraction.Secondly,the 16-fold,8-fold,and 4-fold down-sampled feature maps of the original image obtained by three times upsampling are merged with the corresponding size feature maps extracted by Darknet-53.The fused feature maps are used to locate multiple types of traffic signs and class prediction.Finally,a non-maximum suppression algorithm is used to screen a large number of prediction results obtained in the previous step,select the prediction frame that is closest to the true location of the multi-class traffic signs and correctly classified,and complete the multiclass traffic sign detection task.In order to reduce the amount of parameters of the network algorithm,the TM-YOLO algorithm is proposed to ensure that it meets the design indicators of the self-made multi-class traffic sign data set SUTDB.The TM-YOLO algorithm adjusts the feature extraction network and convolution method in the T-YOLO algorithm.TM-YOLO uses the Mobilenet v1 network for feature extraction,and uses deep separable convolution instead of the standard convolution in the T-YOLO algorithm to achieve the compression of the parameter amount.In this paper,experiments are conducted on four datasets,such as traffic signs,traffic lights,stop lines,and multiple types of traffic sign datasets.Experimental results show that the proposed T-YOLO algorithm has a recall of 97.31% and an average precision of 94.55% on the traffic sign dataset TT100K;On the traffic signal dataset LaRA,the recall is 98.76%,and the average precision is 94.47%;On stopline dataset,the recall is 86.47%,and the average precision is 84.28%;On the multi-type traffic sign dataset SUTDB,the average precision of three types of traffic signs is 92.93%.And the detection speed on the four datasets is higher than 30 frames per second,the parameter is about 247 MB.TM-YOLO obtained an average accuracy mAP of 90%,a detection speed of 41 frames per second,and a parameter of 29 MB,on a multi-type traffic sign dataset SUTDB.Experiments prove that T-YOLO and TM-YOLO algorithms have a good detection effect on multiple types of traffic signs,and meet the design indicators proposed in this paper.