Learning to Recognize Human Activities Using Soft Labels.

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Human activity recognition (HAR) is crucial for robot-care applications.
• Traditional HAR systems require complete and accurate activity labels for training.
• Labeling data can be time-consuming, prone to uncertainty, and involve multiple annotators with inconsistent inputs.

Purpose of the Study:

• To develop a novel learning method for HAR that accommodates incomplete and uncertain labels.
• To introduce the concept of soft labels, allowing for weighted and multiple annotations per data segment.
• To enhance HAR system robustness in real-world scenarios with imperfect data.

Main Methods:

• Formulated the HAR task as a sequential labeling problem.
• Introduced soft labels to represent uncertainty and incompleteness in annotations.
• Embedded latent variables to capture sub-level semantic information.
• Proposed a max-margin framework for model parameter learning incorporating soft labels.

Main Results:

• Evaluated the proposed method on two challenging datasets.
• Simulated label uncertainty by altering labels in transition segments.
• Demonstrated significant performance improvements over state-of-the-art approaches.
• Showcased the effectiveness of soft labels in handling ambiguous and missing annotations.

Conclusions:

• The proposed soft label approach effectively addresses the limitations of complete and accurate labeling in HAR.
• This method offers a more flexible and robust solution for training HAR systems in robot-care.
• The findings suggest a promising direction for improving HAR accuracy and applicability in complex environments.