基于视频的雾天能见度实时监测方法研究

On the real-time monitoring method for the fog-cloud visibility based on the video frequency

目前我国高速公路能见度监测主要依靠能见度观测仪,该方法覆盖区域有限且成本较高,而我国高速公路图像采集设备应用广泛,因此提出了一种基于监控视频图像测量雾天能见度的方法。该方法将图像进行分窗格处理,通过相关性分析筛选出最优兴趣窗格的亮度均方差特征矩阵,建立BP神经网络修正线性残差组合模型。结果表明：残差修正模型监测效果优于单一线性回归模型,残差修正模型的决定系数R2为0.977;在光线充足的情况下,残差修正模型的相对误差在10%以下,模型精度相对稳定;最后应用此方法监测高速公路雾天能见度,模型的正确率在80.48%以上。验证了用该方法测量雾天能见度具有可行性和有效性。

The safety of the fog-cloud visibility monitoring system in China has been mainly dependent on the ground visibility observation,though the said method is expensive and confines itself in the spatial resolution. As compared with the said visibility monitoring system,the image acquisition method should serve as one of the most effective and efficient alternatives to the visibility monitoring system and in turn has been widely adopted by the highway transportation and monitoring authorities. Therefore,in this paper,we would like to introduce a method of measuring visibility in fog,based on the monitoring video. And,for this purpose,we would like to divide,first of all,the images of the fog-cloud into a few so-called window panes,and then try to work out the calculating mean square deviation of the pixel value in the pane as specific features,which can be relevant in terms of its brightness and contrast in the panes of the best interest,which can be extracted through the correlation analysis. And,secondly,it would be possible for us to build up a function model between the features of the interest and the visibility by training the linear regression model,which can be worked out by calculating the residue of the observed simulated data. Hence,correspondingly,the BP neural network can be adopted and applied to correcting the residuals of the linear model with the results being compared with the linear regression model. And,then,comparing the two above models and applying the determination coefficient（ R2） to the results are supposed to enable us to verify that the residual correction model can behave much better than that of the single linear regression. The accuracy R2 of the residual correction model can be made to go up to the accuracy of 0. 977,whereas the BP neural network can further improve the accuracy and generality of the said model.Hence,comparing the different kinds of the light intensity,it can be found that the residual correction model tends to be comparatively stable under the condition of the sufficient light illumination,with a relative error being no more than 10%. Thus,the model can be said qualified enough to demonstrate its general functional performance at the low light intensity,with a relative error being merely 15. 314%. Besides,the method can also be used to monitor the visibility in the highway scenario under traffic and transportation,with the accuracy rate being over 80. 48% and with the leakage rate—below 2. 86%,correspondingly and respectively. Therefore,it can be concluded that the method under study can be taken as feasible and effective in measuring the fog visibility and suitable for the real-time monitoring of the fog visibility with the moving condition in the highway system.