Feature matching for multi-beam sonar image sequence using KD-Tree and KNN search

Abstract

Feature matching for image sequences generated by multi-beam sonar is a critical step in widespread applications like image mosaic, image registration, motion estimation and object tracking. In many cases, feature matching is accomplished by nearest neighbour arithmetic on extracted features, but the global search adopted brings heavy computational burden. Furthermore, sonar imaging characteristics such as low resolution, low SNR, inhomogeneity, point of view changes and other artefacts sometimes lead to poor sonar image quality. This paper presents an approach to the feature extraction, K-Dimension Tree (KD-Tree) construction and subsequent matching of the features in multi-beam sonar images. Initially, Scale Invariant Feature Transform (SIFT) methods are used to extract features. A KD-Tree based on feature location is then constructed. By K Nearest Neighbour (KNN) search, every SIFT feature is matched with K candidates between a pair of consecutive frames. Finally, the Random Sample Consensus (RANSAC) arithmetic is used to eliminate wrong matches. The performances of the proposed approach are assessed with measured data that exhibited reliable results with limited computational burden for the feature-matching task.