Human Shape-Motion Analysis In Athletics Videos for Coarse To Fine Action/Activity Recognition Using Transferable Belief Model

Abstract

We present an automatic human shape-motion analysis method based on a fusion architecture for human
action and activity recognition in athletic videos. Robust shape and motion features are extracted from
human detection and tracking. The features are combined within the Transferable Belief Model (TBM)
framework for two levels of recognition. The TBM-based modelling of the fusion process allows to take
into account imprecision, uncertainty and conflict inherent to the features. First, in a coarse step, actions are
roughly recognized. Then, in a fine step, an action sequence recognition method is used to discriminate activities.
Belief on actions are made smooth by a Temporal Credal Filter and action sequences, i.e. activities,
are recognized using a state machine, called belief scheduler, based on TBM. The belief scheduler is also
exploited for feedback information extraction in order to improve tracking results. The system is tested on
real videos of athletics meetings to recognize four types of actions (running, jumping, falling and standing)
and four types of activities (high jump, pole vault, triple jump and long jump). Results on actions, activities
and feedback demonstrate the relevance of the proposed features and as well the efficiency of the proposed
recognition approach based on TBM.
Key Words: Video Analysis, Human Tracking, Action and Activity Recognition, Transferable Belief Model.