Research On Traffic Sign Detection Model Based On YOLO

Object detection is an important research topic in computer vision,main task of which is to locate and classify targets of interest in pictures.Traffic sign detection,as a subdivision of object detection,plays an important role in the advanced driver-assistance system and automatic driving,and has got peoples attention more and more.In a real-world environment,traffic signs are often in complex backgrounds,and the target sizes are small.These factors can bring difficulties to detect traffic signs efficiently.At the same time,due to the constraints of on-board computing and storage resources,the detection model must be lightweight to have better real-time performance.For this,the dissertation proposes two traffic sign detection models to accommodate the traffic sign detection in vehicle applications,as follows:(1)The dissertation introduces a MSA＿YOLOv3 detection method,which can accurately locate and classify small traffic signs in real time.First,data augmentation is achieved using image mixup technology.Second,a multi-scale spatial pyramid pooling block is introducedinto the Darknet53 network to enable the network to learn object features more comprehensively.Finally,a bottom-up augmented path is designed to enhance the feature pyramid in YOLOv3,and the result is to achieve accurate localization of objects by utilizing fine-grained features effecti3 clvely in the lower layers.This method effectively overcomes the defect that the usual deep learning object detection model is not good for the small object detection: after the small target passes through several convolutional layers,it is difficult to find the details on the feature map of the last layer.According to the tests on the TT100 K dataset(which is a dataset for traffic sign detection),the performance of the proposed MSA＿YOLOv3 is better than that of YOLOv3 in detecting small traffic signs.The detection speed of MSA＿YOLOv3 is 23.81 FPS,and the m AP(mean Average Precision)reaches up to 86%.(2)The L-YOLO object detection model,which is suitable for on-board equipment,reduces the complexity and computation of the model by using partial residual connection on Backbone Network.L-YOLO uses Gaussian box localizaton loss at the detection laye.It reduces FP and improves TP by predicting the uncertainty of bounding box positioning during the detection process.On TAD16 K data set,the m AP of L-YOLO is 2.1% higher than Tiny YOLOv3,45% less parameter amount,46% less calculation amount and 20% higher detection speed.MSA＿YOLOv3 can accurately locate small traffic signs by improving the network structure.L-YOLO is a lightweight traffic sign detection model,which can be combined with multi-object tracking technology to apply to intelligent traffic system as a component of edge computing.