Robust video feature extraction invariant to natural lighting by using LBP techniques with adaptive thresholding

Real time applications in video processing require low computational cost algorithms that allow processing a considerable number, commonly 25, of frames per second. Particularly, in outdoor visual scenes, a challenge is to develop robust algorithms with environmental conditions such as natural lighting. We propose to compute an adaptive threshold based on de probability distribution of the differences in intensity between the pixels and the points on their neighborhoods when applying the LBP technique. We assume, and prove it experimentally, that such distribution is a generalized Gaussian distribution under normal conditions. We consider normal conditions a visual scene in an outdoor field composed of different objects, colors and textures. To compute the adaptive threshold, we first estimate the parameters of the generalized Gaussian distribution using the set of all differences in the image between the intensity values of pixels and points in the neighbourhood. We test the methods on four videos captures during day and night in different places in the city of Ibague. The results of this approach are of interest to determine patterns, identify objects or detect background in a further step. However, an extra step for blur correction must be still included, considering that the images of the frames at night are commonly blurred.