A neural network based method to determine initial object positions for segmentation

This paper presents a new neural network based method for the segmentation of sky from video data collected from a vehicle with a fixed camera position. Using a combination of spatial information, a neural network and thresholding, a high degree of success has been achieved with the images tested. Having the approximate location of the sky allows for an initial starting point for segmentation to be determined. By training a neural network on various sky pixel data, it is possible to find starting locations for thresholding despite the effects of different lighting conditions which significantly affect the colour of the sky in an image. Using this information, thresholds based on colour difference can be employed to discover sky connected pixels. Due to the similar colour of poles to sky, these must then be subtracted from the discovered sky pixels using an edge detection algorithm. The results have been compared with both an SVM and exclusive thresholding technique and comparative analysis is presented in this paper.