基于视觉的夜间细颗粒物浓度估计

Vision-based night-time fine particulate matter concentration estimation

基于视觉的细颗粒物浓度(PM_(2.5))估计技术依据成像时悬浮细颗粒物对光线散射和吸收的整体影响来评估其浓度。这类技术具备良好的普适性,可实时检测广阔区域。已有研究依赖大气光均匀且充足的日间场景,无法适用于缺乏大气光且光照不均匀的夜间场景。本文提出首个基于视觉的夜间PM_(2.5)浓度估计方法,通过图像处理捕获人造光源在不同散射方向的光强分布,并以此特征拟合浓度值。该方法创新地将人造光源及周边光晕区域视为夜晚雾霾信息的主要来源。由于夜间自然光照强度相对人造光源较低,其主导的区域往往趋于漆黑,导致日间雾霾信息的主要来源(自然光照下像素颜色随着景深增加而逐渐接近“大气光/天空”颜色)在夜间的作用相比光源处要小很多。该方法明显优于日间PM_(2.5)估计方法,平均误差(MAE)为6.187μg/m^(3),决定系数(R^(2))为0.857,对比最新的端到端的神经网络方法在MAE和R^(2)上分别有20.69%、13.36%的相对提升。

The technique for estimating the concentration of fine particulate matter(PM_(2.5))based on visual cues relies on assessing its concentration by considering the overall effect of suspended fine particles on light scattering and absorption during imaging.These technologies demonstrate robust generalizability and real-time detection capabilities across large-scale areas.Existing studies depend on daytime scenes with uniform and sufficient atmospheric light,limiting their applicability to nighttime scenario with insufficient atmospheric light and uneven illumination.This paper introduces the pioneering vision-based nighttime estimation method of fine particulate matter concentration,which captures the intensity distribution of an artificial light source in various scattering directions through image processing,and utilizes the feature to correlate with fine particulate matter concentration.The proposed method innovatively leverages the artificial light source and its surrounding glowing area as the main source of nighttime haze information.Since areas dominated by natural lighting typically appear black at night,the conventional basis for daytime haze estimation(i.e.,pixel value approaching the color of“atmospheric-light/sky”as the depth of field increases),is barely useful at night.This method outperforms existing daytime haze estimation methods,achieving a mean absolute error(MAE)of 6.187μg/m^(3) and a correlation coefficient(R^(2))of 0.857.Compared to the popular end-to-end neural network method,it shows a relative improvement of 20.69% in MAE and 13.36%in R^(2).