Tracking Sports Players with Context- Conditioned Motion Models Jingchen Liu, Peter Carr, Robert T. Collins and Yanxi Liu CVPR 2013
Demo
Bayesian Tracking Formulation Associate detections/observations to trajectories
Kinematic Motion Models Continuity of motion alone may be insufficient to resolve identity
Challenges for Tracking Sport Players Weak appearance features Player movements are highly correlated Current game situation influences how each individual will move Independent per-player motion models are tractable
Context-Conditioned Motion Models Motion models conditioned on the current situation Context implicitly encodes multi-player interaction
Hierarchical Data Association
Describe the probability of continuing as Context features: – Absolute position
Hierarchical Data Association Describe the probability of continuing as Context features: – Absolute position – Relative position
Hierarchical Data Association Describe the probability of continuing as Context features: – Absolute position – Relative position – Absolute motion
Hierarchical Data Association Describe the probability of continuing as Context features: – Absolute position – Relative position – Absolute motion – Relative motion
Context-Conditioned Motion Models Describe the probability of continuing as Radom decision forest of 500 trees
Performance