Pose Adaptive Motion Feature Pooling for Human Action Analysis

摘要

Ineffective spatial-temporal motion feature pooling has been a fundamental bottleneck for human action recognition/detection for decades. Previous pooling schemes such as global, spatial-temporal pyramid, or human and object centric pooling fail to capture discriminative motion patterns because informative movements only occur in specific regions of the human body, that depend on the type of action being performed. Global (holistic) motion feature pooling methods therefore often result in an action representation with limited discriminative capability. To address this fundamental limitation, we propose an adaptive motion feature pooling scheme that utilizes human poses as side information. Such poses can be detected for instance in assisted living and indoor smart surveillance scenarios. Taking both video sub-volumes for pooling and human pose types as hidden variables, we formulate the motion feature pooling problem as a latent structural learning problem where the relationship between the discriminative pooling video sub-volumes and the pose types is learned. The resulting pose adaptive motion feature pooling scheme is extensively tested on assisted living and smart surveillance datasets and on general action recognition benchmarks. Improved action recognition and detection performances are demonstrated.