An Improved Interest Point Detection Approach Based on Combined Harris-SIFT Principles

Practical applications such as image registration, 3-D scene reconstruction, object detection, motion tracking, image classification and robot localization rely on the extraction of stable and well localized interest points. This paper proposes an improved interest point detection approach based on combined Harris-SIFT principles. Instead of direct application of Harris corner detector on the original image, the idea is to apply this conventional detector on difference-of-Gaussians (DoGs) of input image, extracted using SIFT principles as baseline. The applied scale-space configuration consists of single octave with varying levels of Gaussian blur. SIFT descriptor is utilized for robust feature description followed by feature matching using efficient nearest neighbor approach. The proposed approach minimizes the probability of poorly localized detections along the edges and also significantly improves the count of stable, well localized interest points. This subsequently increases the feature match rate which was found to be approximately double to that of the conventional Harris and popular SIFT approach, as per the obtained experimental results. Comparison with other state-of-the-art interest point detection approaches is also presented to highlight the improvement aspect.