Research On Atmospheric Visibility Measurement Method Based On Machine Vision

Haze weather brings a lot of inconvenience to people's daily life. Forecasting atmospheric visibility timely and accurately is important for aviation, river and highway transport and so on.In recent years, atmospheric visibility measurement method based on the digital image has been widely used, for it is more in line with the human eye observation. Besides, measurement method based on the digital image is easy to operate and has high detection accuracy. In this paper, we use the digital cameras to collect images and to measure the atmospheric visibility from the perspective of machine vision, the main research contents contain:(1) visibility measurement method based on dark channel prior theory was realized. How to select the defogging coefficient in the dark channel prior theory and how to select correct target area to calculate the transmittance were given a detailed analysis. Experiments were carried out in two different scenarios and the experimental results were compared with the data measured by the forward scatterometer.The experiments show that 93.5% of the experimental data relative error is within ± 15% and is not subject to weather conditions and venues.(2) visibility measurement model based on image single region was established. Through extracting the image region of interest and partitioning the ROI to small area, each small area was analyzed separately. By using dark channel prior theory and the local contrast to extract image feature in each small area, different visibility measurement model was set up, and the model was solved by the least squares method. The experimental results show that the accuracy of the measurement model based on single region is 70.67% and the error is relative large, but it provides a new idea for the improvement and perfection of the measurement model.(3) visibility measurement model based on image multi-region was built. The image was partitioned by different sizes of grid. Visibility measurement model was established by merging the individual regions of the ROI and extracting the image feature. The experiments were carried out by combining different learning methods with feature extraction methods, then the experimental results were compared with the true values. The experimental results show that using improved Sobel filter operator with M5 'model tree has the highest accuracy, and 93.5%of the sample data relative error is within ±20%, which fully validate the feasibility and the accuracy of this method.The accuracy of the visibility measurement based on the dark channel prior theory is high,which accords with the visibility input standard of MOR, but it needs to select the target area manually. The measurement accuracy based on image single region is poor, but it is significantly improved by merging the individual regions of the ROI.