A Multi-Task Learning Framework for Head Pose Estimation under Target Motion.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Head pose estimation (HPE) from low-resolution surveillance data is crucial but challenging due to motion-induced appearance distortions.
• Existing monocular and multi-view HPE methods struggle with accuracy when subjects move, experiencing perspective and scale changes.

Purpose of the Study:

• To propose FEGA-MTL, a novel Multi-Task Learning (MTL) framework for robust head pose classification of freely moving individuals in multi-camera surveillance.
• To enhance HPE accuracy by addressing appearance variations caused by subject motion and camera perspectives.

Main Methods:

• FEGA-MTL partitions surveillance scenes into a spatial grid, clustering partitions into regions with similar facial appearance.
• It simultaneously learns region-specific head pose classifiers, guided by graphs modeling grid partition and head pose class similarities.
• The framework utilizes a person tracker to invoke the appropriate region-specific classifier at test time.

Main Results:

• FEGA-MTL demonstrates significant performance improvements over existing single-task and multi-task learning methods in multi-view scenarios.
• The framework effectively handles appearance distortions from camera perspective and scale changes during subject motion.
• Experiments confirm the superiority of the proposed FEGA-MTL approach for head pose estimation in challenging surveillance environments.

Conclusions:

• FEGA-MTL offers a robust and accurate solution for head pose estimation in dynamic multi-view surveillance.
• The region-specific classification approach effectively mitigates challenges posed by subject movement and varying camera views.
• The framework's adaptability to weakly supervised settings further broadens its applicability in real-world surveillance systems.