Efficient Horizon Line Detection using Clustering and Fast Marching Method

Abstract

Video analysis of maritime scenarios typically includes detection of horizon line for reference. The horizon line is the imaginary line, which separates water and sky as well as water and land. The horizon line plays a major role in terms of demarcating the water region in the video frame for further analysis. Considerable research has been aimed at horizon line detection. Various approaches have been reported including (i) Canny based edge detection followed by Hough transform, (ii) machine learning combined with statistical methods. However, the Hough transform has several limitations, in terms excessive analysis time, deviation of estimated line from the actual horizon line, sensitivity to presence of floating objects on the horizon, error due to presence of large number of edges. Present paper describes an efficient method for detecting the horizon line for analysis videos obtained by cameras mounted on floating vessels such as unmanned surface vehicle in maritime and inland scenarios. The proposed method is based on K-means clustering followed by seed based region growing using Fast Marching Method. For detecting the horizon line, two clusters are used in water-sky region like in marine environment images whereas three or more clusters are used in water-land-sky region like in in-land rivers/lakes images. In most cases, the upper part of the frame belongs to sky region whereas lower part belongs to water region. After K means clustering, based on the selection of seed point in lower part of the frame, the water region is segmented using fast marching method from non water regions and hence the horizon line is detected. This proposed method performance is compared with edge detection followed by Hough transform for different datasets. Experimental results show that the proposed method detects efficient line without compromising the processing time. Index Terms -Unmanned surface vehicle (USV); Horizon line detection; Marine environment scenes, in-land rivers and lakes, K-means clustering.