All-sky Image Prediction Based On Local Features And Particle Filtering

Because of the strong impact of cloudiness on surface solar irradiance, an accurate description of the temporal development of the cloud distribution is essential for the large-scale application of solar energy. Meanwhile, the accurate prediction of all-sky image is of important value in atmosphere analysis. It is computational expensive to make study using meteorology technologies, which are not suitable for short-term prediction of cloud. On another hand, cloud image analysis with the help of computer vision technology, which shows high performance on accuracy and efficiency, are becoming a hot research topic.In this paper, we deal with four parts:camera calibration, cloud detection, cloud matching, and cloud movement estimation. The details are as follows:1. A camera model and calibration method for fisheye lens is proposed. To collect the training data, we make use of the relation of solar zenith angle and the position of the sun in the image innovatively, which could make up the data conveniently.2. Two most distinctive features including red-blue difference and color invariant are extracted, and a hierarchical adaptive segmentation method which combines adaptive Mixture Gaussian Model and LDA-based linear classifier is designed, which are used in the classification of cloud and sky. Experimental results show that the good performance of our proposed method.3. With the comparison of the robustness of some classical color features, SURF features combined with RGB color information are used in cloud matching in image sequences. The proposed Color-SURF feature not only enriches the quantity of detected keypoints, but also increases the accuracy of matching.4. A hierarchical movement estimation model is designed. In traditional affine-transformation model, six pairs of matches at are indispensable to get the motion parameters. In the coarse to fine model, with two pairs of key points detected on the cloud, we are able to estimate the motion of it. This model could estimate not only the displacement, but also variation of scales and rotation of the cloud.5. Particle Filtering is adopted to update the movement of cloud in order to reduce the estimation errors. Experimental results on the real captured all-sky images show the high accuracy of the proposed method.