Transcription

1 Glare reduction for Sky imaging cameras using tone mapping of simultaneous images Jose I. Zapata 1, Geoff Barton 1, Yuxin Liu 1 1 Research School of Engineering, The Australian National University jose.zapata@anu.edu.au Abstract Short-term cloud forecast systems detect clouds and track their movement by acquiring and processing digital colour images of the sky. The high intensity of solar irradiance can induce glare in images containing the solar disk. This paper proposes an approach using two inexpensive surveillance IP cameras fitted with neutral density filters to reduce solar glare. The cameras operate concurrently and capture 1600x1200 pixel images synchronously at 5 second intervals. Camera images are combined with a tone mapping algorithm into an image that exhibits a 40% reduction of the glare radius in the vicinity of the solar disk, and preserves the contrast of cloud and clear sky in the image. All image processing steps are implemented on a desktop computer with OpenCV 3.0, a comprehensive open source C++/Python image processing suite. This paper provides the design and implementation details of the sky imager system and the proposed glare reduction technique. 1. Introduction Short-term cloud forecasting has the potential to improve the operation of a concentrating solar thermal power (CSP) plant by estimating cloud cover over a concentrator field (Law et al. 2014, López-Martínez et al. 2002). Intermittent cloud cover drastically reduces the intensity of direct normal irradiance (DNI) intercepted at the receiver of a CSP plant. Intermittent DNI induces transients at the receiver that can lead to thermal stress and suboptimal plant performance (Camacho et al. 2012, Zapata 2015). Plant operation strategies can incorporate DNI forecasts to pre-empt and counteract receiver transients or to safely steer the system to stand-by mode and wait for cloud cover to dissipate. Short-term cloud forecasting systems employ a digital camera to acquire digital of the sky in the vicinity of the plant under consideration and aim to forecast cloud passage in timeframes between 1 and 60 min (Urquhart et al. 2015). Clouds in digital images are detected using image processing and machine learning algorithms that distinguish between blue (sky) and white (cloud) pixels, and track their movement across several frames. However, the high intensity of solar radiation can saturate camera sensors and create glare in lenses, producing a whiteout section around the solar disk in digital images. Approaches to deal with glare include shadow bands (Chow et al. 2011) and sequential photographs at different exposures (Cazorla et al. 2015). Shadow band approaches require moving parts and sequential exposure techniques require acquiring multiple images whilst varying the exposure in the camera. The approach presented in this study combines two low budget cameras filtered with neutral density filters at two different exposures to obtain low glare images of the sky simultaneously and without moving parts. Subsequently images are combined with image processing, giving high cloud/sky contrast for a low cost, no moving parts and a faster scan rate. Images are then combined using image processing algorithms implemented in open source software, with the aim to produce a low-cost, real time processing system. The rest of this paper describes the camera system hardware and image processing steps to combine images.

2 2. Hardware setup The sky imaging system consists of two fish eye surveillance cameras, a mounting frame, a network router and a desktop computer. The cameras are DLINK DCS-6010 surveillance cameras with a 180 fish eye lens, a maximum resolution of 1600x1200 pixels, a colour resolution of 24-bits, IP connectivity (e.g. UDP and TCP/IP protocols), and are powered with a 240VAC/5VDC power supply. Figure 1 (a) Camera system and alginment frame mounted on location at the ANU Solar Thermal Facilities. (b) Detail of leveling screws The mounting frame consists of a MDF platform where both cameras are mounted adjacent to one another, a swivelling base and levelling screws to ensure a consistent mounting. The network router creates a local area network that connects both cameras and the desktop computer using Ethernet cables. The desktop computer runs Ubuntu and VLC media player to retrieve high resolution images from both cameras at 5 second intervals. Images are stored in the computer hard drive for further processing. During operation, the sky imaging system is mounted in the rooftop test area of the Solar Thermal Facilities at the Australian National University, Canberra (see Figure 1). Figure 2 Internal and external neutral density filters for each camera Figure 3(a) and Figure 3(b) show the resulting images obtained from each camera after the applying the ND filter films. Figure 3(a) provides an image with high contras between clear sky and cloud, but exhibits substantial glare around the solar disk, whereas Figure 3(b) results in low glare but also low contrast. In addition, camera images differ slightly due to a difference between how each sensor is aligned in the camera mounting, and due to being

3 adjacent from one another. The next section details the steps to merge these images, to obtain the image resulting in Figure 3(c). Figure 3 (a) Sky image with 6-fstop reductions (-1.8EV) NDF (b) Sky image with 8 f-stop reductions (-2.4EV) NDF (c) Tone-map fused image displaying cloud details and reduced glare 3. Processing images to form a tone-map fused image Images acquired with the camera system are processed in order to produce a single image with reduced solar glare and high contrast between cloud and sky. The image processing software has been developed in C++ using OpenCV 3.0, a high performance open source library featuring a complete set of image processing and computer vision algorithms. Images captured by one camera are aligned to the simultaneous capture from the second camera so that the sky and surrounding landmarks (i.e. trees, buildings, etc.) overlap when superimposing the images. The process that rotates, translates and stretches an image for the alignment required for the camera is termed an affine transformation. The camera system employs an iterative image alignment algorithm that compares the intensity profiles in both images to calculate the amount of transformation required to align them (Evangelidis & Psarakis 2008). As the alignment between cameras does not change, the amount of transformation (termed the transformation matrix) is calculated a priori and only the affine transform is only applied during operation, to save computation time. Aligned images are subsequently combined using an exposure based tone-mapping algorithm (Mertens et al. 2007). The algorithm blends photographs with different exposures, based on selected portions of each image that satisfy an adjustable criteria for saturation and contrast. In the case of the camera system, the parameters are tuned to reduce glare and increase

4 contrast between contrast and clear sky. A sample tone-map fused image is shown in Figure 3(c). The glare around the solar disk covers approximately 40% less pixels in the tone-map fused image than in Figure 3(a), and exhibits better cloud sky contrast than Figure 3(b). 4. Cloud detection results on tone-map fused images The tone-map fused images produced by the camera system are processed further, so that pixels corresponding to cloud cover can be identified. This is done by first ruling out pixels that are neither cloud nor sky, and then classifying the remaining pixels based on colour differences. The fish eye lenses in the camera system project a circular image onto the square sensor, leaving the outside portion of the image unused. In addition, landmarks surrounding the camera system such as buildings and trees appear at the edge of the fish eye image. All these pixels are ignored by maintaining a mask, a matrix with the same number of elements as pixels in the image, and in each element containing a value of zero to ignore the corresponding pixel in the image or value of one otherwise. The remaining pixels are processed with a thresholding algorithm to differentiate cloud from sky. A thresholding algorithm evaluates the red, green and blue (RGB) information contained in each pixel with a predetermined rule, and determines if a pixel is above or below the threshold set by the rule. For cloud detection, thresholding methods evaluate the colour difference between white/grey clouds and predominantly blue skies. Thresholding rules for cloud detection include the difference between the red and blue channels (RBD), the ratio between red and blue channels (RBR) and colour saturation (Chauvin et al. 2015, West et al. 2014). Figure 4 (a) Tone-mapped image of sky with occluded solar disk (b) Cloud detection algorithm with a saturation level of 0.7, and highlighted area where glare still affects cloud detection Initial trials consisted of evaluating tone-map fused images obtained by the camera system using RBR, RBD, normalised RBD, saturation and Euclidian Geometric Distance thresholding rules; and manually varying the threshold values for each rule. These trials showed that the best performing algorithm is a saturation based rule (Souza-Echer et al. 2006), which calculates the saturation of a pixel in the range [0, 1]. A threshold value of 0.7 produces a satisfactory cloud outline across an entire image.

5 Figure 4 shows a tone-map fused image with only a sliver of the solar disk behind clouds, alongside the resulting map of detected clouds. It can be seen that cloud is correctly identified throughout the image, even when the shades of cloud range between bright white and dark grey. Similarly, clear sky portions have been identified for different shades of blue, and even in small portions surrounded by cloud. In Figure 4(b), a region around the solar disk has been manually shaded in light blue to highlight an area of the image identified as cloud, but can be seen to correspond to clear sky in Figure 4(a). This is a result of the selected thresolding value that is sensitive around the occluded solar disk. It is possible to correctly identify this portion of the image as clear sky using a threshold of 0.6, but this yields poor cloud detection elsewhere in the image. Figure 5 (a) Tone-map fused image of sky with visible solar disk (b) Cloud detection on tone-map fused image with a saturation level of 0.5 (c) Single camera image of sky with visible solar disk and glare (d) Cloud detection on single camera image with a saturation threshold of 0.5 Figure 5 shows another cloud detection result of a tone-map fused image compared against the cloud detection of single camera shot. Figure 5 shows that at a reduced threshold value of 0.5, the tone-map fused image produces a smaller cloud than the single shot image with glare, demonstrating the potential of the glare reducing technique proposed in this study. This result also shows how reducing the saturation value may improve the distinction of cloud, clear sky around the solar disk but it is detrimental elsewhere in the image.

6 5. Summary and future work The present work has shown the design and construction of a low cost camera system intended to track the passage of clouds over a concentrating solar thermal system to forecast changes in direct normal irradiance. The camera system consists of two cameras with different neutral density filters, in order to capture a simultaneous double exposure image of the sky. The low exposure image provides reduced glare around the solar disk whereas the higher exposure contributes better contrast between clouds and sky. Images are successfully combined with an exposure tone-map fusion algorithm, and processed with a saturation based threshold algorithm to identify and classify pixels in the image that are cloud, clear sky or surrounding landmarks. Future work in this project will focus on refining the classification of clouds and clear sky around the solar disk. The work includes tracking the position of the solar disk in the images using solar position algorithms, and processing different regions of the image with different threshold values (i.e. adaptive thresholding). References Camacho, E., Berenguel, M., Rubio, F. & Martinez, D. (2012), Control of Solar Energy Systems, 1st edn, Springer-Verlag London Limited. Cazorla, A., Husillos, C., Antón, M. & Alados-Arboledas, L. (2015), Multi-exposure adaptive threshold technique for cloud detection with sky imagers, Solar Energy 114, Chauvin, R., Nou, J., Thil, S., Traoré, A. & Grieu, S. (2015), Cloud detection methodology based on a sky-imaging system, Energy Procedia 69, International Conference on Concentrating Solar Power and Chemical Energy Systems, SolarPACES Chow, C. W., Urquhart, B., Lave, M., Dominguez, A., Kleissl, J., Shields, J. & Washom, B. (2011), Intra-hour forecasting with a total sky imager at the uc san diego solar energy testbed, Solar Energy 85(11), Evangelidis, G. & Psarakis, E. (2008), Parametric image alignment using enhanced correlation coefficient maximization, IEEE Transactions on Pattern Analysis and Machine Intelligence 30(10), Law, E. W., Prasad, A. A., Kay, M. & Taylor, R. A. (2014), Direct normal irradiance forecasting and its application to concentrated solar thermal output forecasting - a review, Solar Energy 108(0), López-Martínez, M., Vargas, M. & Rubio, F. R. (2002), Vision-based system for the safe operation of a solar power tower plant, in IBERAMIA 2002: Proceedings of the 8th Ibero- American Conference on AI, Springer-Verlag, London, UK, pp Mertens, T., Kautz, J. & Van Reeth, F. (2007), Exposure fusion, in Computer Graphics and Applications, PG th Pacific Conference on, IEEE, pp Souza-Echer, M. P., Pereira, E. B., Bins, L. S. & Andrade, M. A. R. (2006), A simple method for the assessment of the cloud cover state in high-latitude regions by a ground-based digital camera, Journal of Atmospheric and Oceanic Technology 23,

7 Urquhart, B., Kurtz, B., Dahlin, E., Ghonima, M., Shields, J. E. & Kleissl, J. (2015), Development of a sky imaging system for short-term solar power forecasting, Atmospheric Measurement Techniques 8(2), West, S. R., Rowe, D., Sayeef, S. & Berry, A. (2014), Short-term irradiance forecasting using skycams: Motivation and development, Solar Energy 110(0), Zapata, J. (2015), Full state feedback control of steam temperature in a once-thorough direct steam generation receiver powered by a paraboloidal dish, Journal of Solar Energy Engineering 137(2), Acknowledgements The authors wish to acknowledge the Australian Renewable Energy Agency (ARENA) and the ANU Research School of Engineering for their financial support to conduct this study.