A Wavelet-Based Approach to Image Feature Stability Assessment

In this paper, we present a novel method for assessing image-feature stability. The method hinges on applying the discrete wavelet transform to the image features under study throughout a number of video frames in an image sequence. For purposes of stability assessment, we recover the imagefeature vectors for each video frame and then track them trough a series of consecutive frames in the image sequence. We apply the discrete wavelet transform to the time series constructed from the pairwise Euclidean distances for each of the image features under study and use the wavelet transform coefficients to assess their stability. We then recover the stable features by clustering together those time series which exhibit largely constant low-pass wavelet coefficients. We present results of the stability analysis for Harris corners, Maximally Stable Extremal Regions, and Scale Invariant Feature Transform regions extracted from two real-world video sequences. We also elaborate on the applications of our method to indexing, retrieval, and compression of stable image feature vectors.