Recognizing Shapes in Video Sequences Using Multi-class Boosting

We model the spatio-temporal variations of the shape of objects in a video sequence using a unique SVD-like decomposition. The decomposition is used to compute shape features, which form an approximation of the original shape sequence. The features are used to train separate classifiers using multi-class boosting strategy. We demonstrate the effectiveness of the proposed approach for shape recognition using the China Lake outdoor surveillance dataset; and compare the results using mean shapes as baseline. We illustrate the usefulness of the proposed shape features for detecting shapes of interest using the SIG group activity dataset.