Mixture of Heterogeneous Attribute Analyzers for Human Action Detection

Abstract

We propose a human action detection framework called “mixture of heterogeneous attribute analyzer”. This framework integrates heterogeneous attributes learned from static and dynamic, local and global video features, to boost the action detection performance. To this end, we first detect and track multiple people by SVM-HOG detector and tracklet generation. Multiple short human tracklets are then linked into long trajectories by spatio-temporal matching. Human key poses and local dense motion trajectories are then extracted within the tracked human bounding box sequences. Second, we propose a mining method to learn discriminative attributes from these three feature modalities: human bounding box trajectory, key pose and local dense motion trajectories. Finally, the learned discriminative attributes are integrated in a latent structural max-margin learning framework which also explores the spatio-temporal relationship between heterogeneous feature attributes. Experiments on the ChaLearn 2014 human action dataset demonstrate the superior detection performance of the proposed framework.