Activity detection using Sequential Statistical Boundary Detection (SSBD)

The spiralling increase of video data has rendered the automated localization and recognition of activities an essential step for video content understanding. In this work, we introduce novel algorithms for detecting human activities in the spatial domain via a binary activity detection mask, the Motion Boundary Activity Area (MBAA), and in the time domain by a new approach, Statistical Sequential Boundary Detection (SSBD). MBAAs are estimated by analyzing the motion vectors using the Kurtosis metric, while dense trajectories are extracted and described using a low level HOGHOF descriptor and high level Fisher representation scheme, modeling a Support Vector Data Description (SVDD) hypersphere. SSBD is then realized by applying Sequential Change Detection with the Cumulative Sum (CUSUM) algorithm on the distances between Fisher data descriptors and the corresponding reference SVDD hyperspheres for rapid detection of changes in the activity pattern. Activities in the resulting video subsequences are then classified using an multi-class SVM model, leading to state of the art results. Our experiments with benchmark and real world data demonstrate that our technique is successful in reducing the computational cost and also in improving activity detection rates.