Comparative Study of Landmark Detection Techniques for Airport Visibility Estimation

Reliable and exact assessment of visibility is essential for safe air traffic. In order to overcome the drawbacks of the currently subjective reports from human observers, we present an approach to automatically derive visibility measures by means of image processing. It is based on identification of visibility of individual landmarks and compiling an overall visibility range. The methods used are based on concepts of illumination compensation as well as structural (edges) and texture recognition. Validation on individual landmarks showed a reliable performance of 96% correct detections. Furthermore, a solution for compiling the overall visibility report is presented, that resembles the currently used standard in air traffic management. 1. Motivation In order to guarantee for safe air traffic controll ers are relying on precise forecasts and measurements of the current weather situation. The exact acquisition and specification of the atmospheric condition builds the basis for any fore cast and thus is elementary for any aviation weather service. While weather minima 1 that still allow for efficient and safe air traffi c are continuously lowered, they still have huge impact o n air traffic. E.g. thunderstorms [13, 2] along with all their weather phenomena have severe influe ce on efficient and economic handling of air traffic. This is also reflected in various delay-st atistics [1, 13]. Figure 1: Classic airport sensor system: meteorological instrumentation and equipments (left), ceilometer (middle), sketch of a visibility sensor (right). 1 Visual meteorological conditions: http://en.wikipe dia.org/wiki/Visual_meteorological_conditions All major airports operate dedicated sensor systems to assess the current weather situation. Besides “classic” parameters like wind, pressure, humidity, and temperature, there are point-like measurements of visibility and cloud cover informat ion (Figure 1). One of the essential parameters is a precise measur ement of the visibility in the airport range. Currently human observers compile visibility report s2 every 30 minutes based on visual observation of known landmarks (prominent structures like build ings, mountain tops, etc...) in a so-called landmark map (Figure 5 and Figure 6). These landmar ks have attached the distances to the observation point, and by identifying which of thos e landmarks are still visible, and which not, a visibility estimate is derived. This naturally is very subjective to the individual observer and error-prone, thus an objective measurement is highly sought by all operators to al l w for efficient flight and tactical air traffic planning as well as for operational handling of air traffic. The automated measurement with a visibiliy sensor (Figure1, right) is still a rough estimate only as it calculates the overall long-ran ge visibility from close and very local observations. The approach presented in this paper is aiming at e mulating the human observer procedure, by deriving the visibility of already established land marks with automatic image processing methods. As the landmarks completely differ from each other in their appearance, structure, size, as well as by illumination variations (time of the day, changi meteorological conditions such like rain, snow and fog), it is not possible to set-up a unique met hod that can cope for all landmarks. Thus, differen t approaches to illumination compensation, structure detection, and classification are applied and evaluated, in order to find the most discriminating method for each individual landmark. The rest of the paper is structured as follows. The methods applied to landmark detection are presented in Section 3 following a short overview o f the relevant literature in this field (Section 2) . By analysis of a comparative study (Section 4) it i s shown how to choose, for each landmark separately, the best suited method for its recognit ion. This is followed by displaying the accuracy of the method on several landmarks and a discussion on the strategy in assessing the prevailing visibility based on the visibility of each singular ndmark. Section 5 concludes with final remarks. 2. Related Work Current visibility sensors employ the sender and re ceiver principle: a ray of light is emitted by a projector and caught either by a photodetector (e.g . scatter meter) or by a digital camera. The literature review below shows that the usage of cam er s for the measurement of visibility became more and more popular over the last six years. In 2005, Luo et al. [9] measured visibility by anal yzing the intensities of grey level images. Due to the fact that high-frequency information depends on the brightness and the texture in urban images they developed a model to establish a relationship between frequency components and urban visibility. They showed that using a Sobel operator or FFT (high pass filtering) is adequate for extracting high frequency components and thus for m nitoring visibility. Furthermore it was proven that the results of both methods correlate with eac h other as well as with human observations. Also Raina et al. [11] investigated the usage of co ntrast for the measurement of visibility. Unlike [9], instead of investigating the whole image, only regions of interest were employed. Their experiments are based on a network of webcams where contrast values of acquired images are compared to clean day conditions. Statistical evalu ations finally allow for the sought classification. 2 METAR: international standard format for reporting weather information Another approach was presented by Kim et al. [7] wh o investigated the relationship between the optical measurement and HSI colour differences. Th eir goal was to analyze air pollution based on visibility. The idea for the approach itself is bas ed on the fact that the colour of sky depends on th e light scattering (e.g. blue for small aerosol parti cles and white for larger particles) and especially the colour of haze varies with the optical properties o f aerosol. By measuring the difference of the HSI space between a target image and the clear sky refe rence image it is possible to estimate the status of visibility by the usage of the developed grading visibility level. Poduri et al. [10] went one step further trying to make sky analysis available for mobile phones. Their approach is based on the generation of an ana lytic model of the sky as a function of appearance. Visibility is finally estimated by comp arison of a new image with this model. The main drawback of this method is that it works for cloud free sky only. All those approaches have in common that their main goal is to “see” the amount of pollution in the air. The basic idea is always to develop a referenc e model and compare it to the newly acquired image. In contrast to our work we are not intereste d in the long distance visibility only but especially in the maximum visible distance represen ted by the visibility of previously defined regions of interest. Furthermore, we do not use a “ perfect weather image” as reference image but employ several occurring views. 3. Landmark Detection Our visibility estimation is based on recognition o f ground landmarks scattered around airports. Due to different sun illumination throughout the da y and varying weather conditions like rain, snow and fog, the same landmark can display drastically different appearances. Several approaches have been proposed for solving the variable illumination problem in image processing. As shown in [12] for face recognition under various lighting conditi ons these approaches can be classified into three main categories: normalization, invariant features extraction and modelling. The first category of approaches ( normalization) includes image pre-processing algorithms that are employed to compensate and normalize the illuminati o . Since most of these algorithms do not require any training or modelling they can be consi dered as general purpose image pre-processing algorithms like histogram equalization, gamma corre ction, and logarithmic transforms. The second category ( invariant features extraction) aims at extracting illumination invariant features from the image and applies the recognition on those. Edge maps and different texture descriptors (Gabor filtering, Local Binary Patterns , etc...) belong to this category. The last category of approaches ( modelling) generally uses low-dimensional linear subspaces f or modelling image variations under different lighting conditions. For instance, the Principal Component Analysis (PCA) falls among this category. These approaches generally require a training set of images representing the object unde r a lot of different illumination conditions. 3.1. Illumination normalization The Retinex theory [8] yields an algorithm for extr acting an illumination-normalized representation of the image. The theory is based on the reflectanc e illumination model of human vision which assumes that it is sensitive to scene reflectance a d local change of contrast while being insensitive to illumination conditions and global brightness le vels. The basic definition for a pixel at position (x, y) is: ( ) ( ) ( ) ( ) y x I Ke y x I y x R c r , log , log , / 2 ⊗ − = − (1) with 2 2 1 πσ = K , 2 2 2 y x r + = and 2 2σ = c This is the logarithmic difference (e.g. the logari thm of the quotient) between the image and a version of it convolved with a low-pass filter (i.e . Gaussian filter), the low frequency component being considered as the illuminance of the input st imulus (Figure 2). Figure 2: original image (left), Retinex (right) The constant c is referred as the scale of the Retinex and by var ing it, Jobson et al. [6] proposed an extension to the original Retinex algorithm that in stead of giving the result at a single scale, outpu ts an image constructed as the weighted sum of single scaled Retinex images. That implementation displays a better balance of dynamic compression an d colour re