Research On Traffic Sign Detection Based On Multiscale Awareness And Lightweight Algorithm

With the rapid development of object detection and deep learning,traffic sign detection has made tremendous progress.However,there are still challenges in current traffic sign detection,such as low accuracy in detecting small objects,missed detections,low real-time performance,and high computational costs.To address these issues,this thesis proposes a series of improvements to the YOLO v7 network model to achieve fast and accurate detection of traffic signs.The main contributions of this thesis are as follows:(1)To address the issues of low accuracy and missed detections in current traffic sign detection,this thesis proposes a method for improving the detection of traffic signs based on YOLO v7.Firstly,an improved attention mechanism combining CBAM and ECA is embedded in the YOLO v7 network to effectively capture interdependencies across channels and obtain information from both spatial and channel dimensions.Secondly,a multi-scale perception module is used to improve the receptive field of objects in the model by using a combination of 3x3 convolutions and shared convolution techniques,enhancing the feature fusion performance of the model.Finally,the FocalEIo U is used as the loss function for bounding boxes to achieve high-precision localization of targets.Experimental results on the CCTSDB(Chinese Traffic Sign Detection Benchmark)dataset show that the proposed method achieves an m AP(mean Average Precision)of 91.0%,which is a 4.2% improvement over the original YOLO v7,demonstrating the effectiveness of this approach.(2)To address the issues of low real-time performance and high computational costs in traffic sign detection,this thesis proposes a method based on YOLO v7 and knowledge distillation.Firstly,a lightweight network called Ghost Net is used as the backbone network for YOLO v7 to optimize the network model and accelerate network operations.Secondly,to further reduce the parameter count of the entire convolutional neural network architecture,Depth-wise Separable Convolution technology is employed to replace the regular convolution in the feature fusion network of YOLO v7.Finally,in the target classification stage,knowledge distillation is used to enable the student network to learn the powerful classification and regression capabilities of the teacher network.Experimental results demonstrate that the improved YOLO v7 algorithm in this paper has a slight decrease in detection accuracy of 1.5%,but the inference time of the model is reduced to one-third of the original.Figure[31] Table[9] Reference[80]...