A Deep Regression Approach for Human Activity Recognition Under Partial Occlusion.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human Activity Recognition (HAR) is crucial for video analysis but struggles with occluded body parts.
• Existing HAR research often uses ideal, occlusion-free datasets, limiting real-world applicability.
• Occlusion significantly degrades HAR performance by obscuring key body movements.

Purpose of the Study:

• To develop a robust HAR approach capable of handling partial body part occlusion.
• To address the underestimation of occlusion challenges in current HAR research.
• To improve HAR accuracy in realistic, unconstrained environments.

Main Methods:

• Proposed a novel deep Convolutional Recurrent Neural Network (CRNN) for HAR.
• Modeled human motion using 3D skeletal joints, assuming occluded parts remain occluded.
• Tackled occlusion by formulating HAR as a regression task to reconstruct missing motion data.

Main Results:

• Achieved significant performance increases compared to baseline methods on occluded data.
• Demonstrated the effectiveness of the CRNN in reconstructing occluded body part motion.
• Validated the approach on four publicly available human motion datasets.

Conclusions:

• The proposed regression-based CRNN effectively handles partial occlusion in HAR.
• This work is the first to address HAR under occlusion as a regression problem.
• The findings pave the way for more reliable HAR systems in real-world applications.