Research on Visibility Estimation Model Based on DenseNet

Visibility is a common indicator in meteorology, highway traffic and aircraft flight. The unit is usually meters. The main factors affecting visibility are fog and haze. As we all know, visibility is very important to highway safety. When visibility is very low, highway managers usually close the road in order to drive safety. However, in the aviation field, runway visibility is used to reflect the size of fog and haze near the airport, which is defined as the maximum distance at one end of the runway along the direction of the runway to identify the runway or the target object close to the runway (runway side lights at night). Normally, flights are banned when visibility at the airport is only about 400 meters. When visibility at the airport is only about 600 to 800 meters, flights can take off and land normally. However, due to security concerns, the airport will take temporary measures to control the flight flow, widening the flight departure interval, which is easy to cause flight delays. Therefore, visibility prediction is the highway management department and airlines are very concerned about the problem.

In recent years, video-based road condition (runway) visibility detection method [1,2] has attracted people’s attention, which to some extent overcomes the deficiency of laser visibility meter. Video visibility detection method is a combination of atmospheric optical analysis, image processing and artificial intelligence technology, through the analysis and processing of video images, establish the relationship between the video image and the real scene, and then calculate the visibility value indirectly according to the changes of image features. Wang Yaxue et al [3] established estimation and prediction models for visibility under different fog conditions. Through the establishment of multiple classification of multiple regression model. Han Huihui [4] introduced the ideas of closed-loop control and transfer learning into deep convolutional neural networks and designed a model of intelligent fog classification and visibility estimation based on deep transfer learning. Liu Jianlei et al. [5] analyzed the characteristics of anisotropy, continuity and horizontality of the inflection point line. Then, an inflection line detection filter is constructed according to these characteristics to improve the accuracy and speed of the detection. With the application of machine learning in various fields, the strong feature learning ability of deep learning network model has been applied to the classification and prediction of meteorological information, which has played a great role in promoting the improvement of forecast accuracy. In foreign countries, the artificial neural network model was initially used to forecast the visibility of winter weather in Milan, Italy [6], and the effect was better than the traditional method. By extending the standard probabilistic neural network method, Bremnes et al. [7] used multiple models to obtain an average output to improve the visibility forecast. Marzban et al. [8] combined the output of the numerical model with the ground observation data, and mapped the data to 39 different neural networks. The results showed that the performance of neural networks was generally due to logistic regression and MOS model. In China, Zhou Yongjiang et al. [9] used meteorological parameters such as temperature, air pressure, atmospheric precipitable water volume and PM 2.5 data of the same period to establish a haze prediction model integrating time series network and regression network to detect and forecast haze weather, proving that the fusion network model with meteorological parameters is more adaptable and accurate than the single network model.

In order to solve the visibility prediction problem in foggy weather, this paper constructs the visibility estimation model mainly through the method based on deep learning. The following is the introduction of deep learning-related issues and the main work of this paper.

The concept of deep learning stems from research on artificial neural networks. Deep learning is a feature learning method that transforms raw data into a higher level and more abstract representation through some simple but nonlinear models. A more abstract high-level representation (attribute category or feature) is formed by combining low-level features to discover distributed feature representation of data [9]. In the process of learning the internal rules and representation hierarchy of sample data, deep learning has the ability to analyze, learn and interpret the acquired text, image and sound data, making it achieve remarkable achievements in image recognition, speech recognition, natural language understanding, weather prediction, gene expression, content recommendation and other aspects. Nowadays, deep learning plays a crucial role in various tasks of image analysis by discovering or learning information features that describe the inherent laws or patterns of data.

In the field of computer vision, convolutional neural network (CNN) has become the most mainstream method. A milestone event in the history of CNN is the emergence of ResNet model, which can train deeper CNN model and achieve higher accuracy. The core of the ResNet model is to train a deeper CNN network by establishing a “short-circuit connection” between the front layer and the back layer, which facilitates the back propagation of the gradient during the training process. The DenseNet(Densely connected convolutional networks) model has the same basic idea as ResNet, but it establishes the dense connection between all the front layers and the back layers. Another feature of DenseNet is feature reuse through the concatenation of features on channels. These features allow DenseNet to achieve better performance than ResNet with fewer parameters and computational costs,

In this paper, the mapping relationship between image coordinates and road coordinates is established for the video images collected by the camera, and the image distance information is converted into road distance information. The mapping relationship between image coordinates and road coordinates is completed by camera self-calibration technology, and its working steps are as follows: 1)

Establish the road condition camera imaging model, as shown in Figure 3, in which three coordinate systems are defined, among which the ground coordinate system XW-YW-ZW and the camera coordinate system XC-YC-ZC are used to represent the three-dimensional space; Image plane coordinate system Xf -- Yf represents the imaging plane. The world coordinate system is established, whose origin is the intersection point of the camera optical axis and the ground. The camera coordinate system is established, and the origin is the camera optical center position. Set the distance between the optical center of the camera and the origin of the world coordinate system as l, the pitch Angle of the camera as t, the declination Angle as p, and the rotation Angle as s, and the area between parallel lines on the ground plane corresponds to the highway pavement in the camera’s view field.

Figure 1.

Figure 2.

Figure 3.

Further, more accurate and effective training can be performed if the convolutional network contains shorter connections between the input layer and the output layer. In this paper, we accept this observation and introduce a dense convolutional network (DenseNet) that connects each layer to each other in a feedforward manner. A traditional convolutional network with L layers has L connections -- one connection between each layer and the one that follows -- and our network has L (L + 1) / 2 direct connections. For each layer, the feature maps of all previous layers are used as inputs, and its own feature maps are used as inputs for all subsequent layers. DenseNets have several compelling advantages: they alleviate the problem of disappearing gradients, enhance feature propagation, encourage functional reuse, and dramatically reduce the number of parameters.

The structure of the DenseNet network is shown in the figure 2. Each layer of the network is connected to all subsequent layers in a dense connection way.

In the training of DenseNet model, error information of training samples can be transmitted Back to the hidden layer through Back Propagation (BP) algorithm to realize constant updating and iteration of weight matrix between hidden layers until network convergence.

BP network is a feedforward neural network, which was proposed in 1985. In this kind of network, there are two kinds of signals flowing: one is the working signal, which is the input signal applied and propagated forward until the actual output is generated at the output end of the signal, is a function of the input and weight; The second is the error signal. The difference between the actual network output and the theoretical output is the error, which starts from the output end and propagates back layer by layer. The output of the JTH unit at the output end in the NTH iteration is y_j (n), then the error signal of the unit is e_j(n) = d_j(n) − y_j(n), and the square error of unit j is defined as 12yj2(n)${1 \over 2}y_j^2(n)$. The instantaneous value of the total square error of the output is: 8 ε(n)=12∑jej2(n) \[\varepsilon (n)=\frac{1}{2}\underset{j}{\mathop \sum }\,\text{e}_{j}^{2}(n)\]

Densenet can not only make efficient use of multi-dimensional feature information through the “feature re-calibration” strategy, but also slow down the attenuation of error terms of each hidden layer through the reverse transmission mechanism of the network itself, ensure the stability of gradient weight information, and enhance the learning and expression ability of the deep network, so as to further improve the network performance.

The data set used in this paper is the airport video and airport AMOS observations provided by the 2020 Modeling Competition. Firstly, the original airport video information and meteorological data were cleaned. Using python data analysis technology to carry out data fusion of features; In addition, when using DenseNet for feature extraction, direct input of disordered discrete data will mislead the internal training mechanism of the network and confuse the importance of features. Discrete data is encoded by assigning values in turn according to the frequency of variable occurrence. For continuous data, feature values are mapped to [0,1] by minmax normalization to remove dimensional differences between features.

The classification prediction module mainly consists of three layers of network structure: global pooling layer, full connection layer and Softmax classifier. Through the combined operation of the three-layer network structure, the n×n×k dimension size matrix is converted into the probability corresponding to the occurrence of each category. Figure 4 below shows the integrated visibility analysis Organization.

Figure 4.