Quadruped Gait Detection in Low Quality Wildlife VideoSion Hannuna, Quadruped Gait Detection in Low Quality Wildlife Video. PhD thesis. University of Bristol. December 2007. PDF, 14898 Kbytes.
The work described within this thesis describes two novel approaches to detecting walking quadrupeds in unedited wildlife film footage. The first technique exploits trends in the angular variability of the 2-D power spectra of natural images to extract discriminatory features for classification using both k-nearest neighbour (KNN) and support vector machine (SVM) classifiers. In various experiments wildlife stills are classified correctly as `gait' vs. `non-gait' with a success rate of around 88%. The second, more complex, approach aims to extract phase information from the timed interplay of the coordinated movements comprising quadruped gait. This is an extremely challenging problem. Variable lighting, moving backgrounds and camouflaged animals make traditional foreground extraction techniques such as optical flow and background subtraction unstable. In the work described here, a sparse set of points is tracked over a short film clip and a RANSAC type algorithm is used to segment the foreground region. A novel technique, employing normalised convolution is then utilised to interpolate dense flow from the sparsely defined foreground. Dense flow is extracted for a number of clips demonstrating quadruped gait and other movements. Principal component analysis (PCA) is applied to this set of dense flows and eigenvectors not encapsulating periodic internal motion characteristics are disregarded. The projection coefficients for the remaining principal components are analysed as one dimensional time series. Projection coefficient variation reflects changes in the velocity and relative alignment of the components of the foreground object. These coefficients' relative phase differences are deduced using spectral analysis and degree of periodicity using dynamic time warping. These parameters are used to train KNN and SVM classifiers with the latter segmenting the training data with 97% success rate. By generating projection coefficients for unseen footage, the system has successfully located examples of quadruped gait previously missed by human observers.