Traffic sign detection is a crucial task for autonomous driving systems.However,the performance of deep learning-based algorithms for traffic sign detection is highly affected by the illumination conditions of scenari...Traffic sign detection is a crucial task for autonomous driving systems.However,the performance of deep learning-based algorithms for traffic sign detection is highly affected by the illumination conditions of scenarios.While existing algo-rithms demonstrate high accuracy in well-lit environments,they suffer from low accuracy in low-light scenarios.This paper proposes an end-to-end framework,LLTH-YOLOv5,specifically tailored for traffic sign detection in low-light scenarios,which enhances the input images to improve the detection performance.The proposed framework comproses two stages:the low-light enhancement stage and the object detection stage.In the low-light enhancement stage,a lightweight low-light enhancement network is designed,which uses multiple non-reference loss functions for parameter learning,and enhances the image by pixel-level adjustment of the input image with high-order curves.In the object detection stage,BIFPN is introduced to replace the PANet of YOLOv5,while designing a transformer-based detection head to improve the accuracy of small target detection.Moreover,GhostDarkNet53 is utilized based on Ghost module to replace the backbone network of YOLOv5,thereby improving the real-time performance of the model.The experimental results show that the proposed method significantly improves the accuracy of traffic sign detection in low-light scenarios,while satisfying the real-time requirements of autonomous driving.展开更多
实时的吸烟行为监测对于保障工地安全具有重要的意义。研究针对隧道等低照明环境因光照强度低、光线分布混杂、点光源过曝光而导致的烟支检测精度低的问题,提出了一种基于改进YOLOv5s(You Only Look Once Version 5s)的吸烟行为检测模...实时的吸烟行为监测对于保障工地安全具有重要的意义。研究针对隧道等低照明环境因光照强度低、光线分布混杂、点光源过曝光而导致的烟支检测精度低的问题,提出了一种基于改进YOLOv5s(You Only Look Once Version 5s)的吸烟行为检测模型。首先,设计一种图像增强方法,旨在限制点光源过曝光产生的局部高光,增强烟支特征细节,改善图像对比度。其次,引入卷积块注意力模块(Convolutional Block Attention Module,CBAM),使模型更加聚焦烟支目标区域的内容信息和位置信息。最后,改进多尺度检测头,增加适用于烟支的更小检测层,提升模型对烟支的检测能力。试验结果显示,研究提出的针对低照明环境的检测模型可将平均检测精度从91.8%提升至95.9%,相较于原模型和其他经典模型,检测效果得到了显著提升,表明了方法的有效性。展开更多
基金National Natural Science Foundation of China,U20A20331,Long Chen.
文摘Traffic sign detection is a crucial task for autonomous driving systems.However,the performance of deep learning-based algorithms for traffic sign detection is highly affected by the illumination conditions of scenarios.While existing algo-rithms demonstrate high accuracy in well-lit environments,they suffer from low accuracy in low-light scenarios.This paper proposes an end-to-end framework,LLTH-YOLOv5,specifically tailored for traffic sign detection in low-light scenarios,which enhances the input images to improve the detection performance.The proposed framework comproses two stages:the low-light enhancement stage and the object detection stage.In the low-light enhancement stage,a lightweight low-light enhancement network is designed,which uses multiple non-reference loss functions for parameter learning,and enhances the image by pixel-level adjustment of the input image with high-order curves.In the object detection stage,BIFPN is introduced to replace the PANet of YOLOv5,while designing a transformer-based detection head to improve the accuracy of small target detection.Moreover,GhostDarkNet53 is utilized based on Ghost module to replace the backbone network of YOLOv5,thereby improving the real-time performance of the model.The experimental results show that the proposed method significantly improves the accuracy of traffic sign detection in low-light scenarios,while satisfying the real-time requirements of autonomous driving.
文摘由于低照度图像具有对比度低、细节丢失严重、噪声大等缺点,现有的目标检测算法对低照度图像的检测效果不理想.为此,本文提出一种结合空间感知注意力机制和多尺度特征融合(Spatial-aware Attention Mechanism and Multi-Scale Feature Fusion,SAM-MSFF)的低照度目标检测方法 .该方法首先通过多尺度交互内存金字塔融合多尺度特征,增强低照度图像特征中的有效信息,并设置内存向量存储样本的特征,捕获样本之间的潜在关联性;然后,引入空间感知注意力机制获取特征在空间域的长距离上下文信息和局部信息,从而增强低照度图像中的目标特征,抑制背景信息和噪声的干扰;最后,利用多感受野增强模块扩张特征的感受野,对具有不同感受野的特征进行分组重加权计算,使检测网络根据输入的多尺度信息自适应地调整感受野的大小.在ExDark数据集上进行实验,本文方法的平均精度(mean Average Precision,mAP)达到77.04%,比现有的主流目标检测方法提高2.6%~14.34%.
文摘实时的吸烟行为监测对于保障工地安全具有重要的意义。研究针对隧道等低照明环境因光照强度低、光线分布混杂、点光源过曝光而导致的烟支检测精度低的问题,提出了一种基于改进YOLOv5s(You Only Look Once Version 5s)的吸烟行为检测模型。首先,设计一种图像增强方法,旨在限制点光源过曝光产生的局部高光,增强烟支特征细节,改善图像对比度。其次,引入卷积块注意力模块(Convolutional Block Attention Module,CBAM),使模型更加聚焦烟支目标区域的内容信息和位置信息。最后,改进多尺度检测头,增加适用于烟支的更小检测层,提升模型对烟支的检测能力。试验结果显示,研究提出的针对低照明环境的检测模型可将平均检测精度从91.8%提升至95.9%,相较于原模型和其他经典模型,检测效果得到了显著提升,表明了方法的有效性。
