Recognition of Human Activities Using Depth Maps and the Viewpoint Feature Histogram Descriptor

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Human activity recognition (HAR) is crucial for applications like surveillance and healthcare.
• Existing methods often struggle with complex human movements and varying viewpoints.

Purpose of the Study:

• To propose and evaluate a novel approach for human activity classification using 3D point clouds.
• To compare the effectiveness of k-NN and BiLSTM classifiers for this task.
• To explore methods for improving classification efficiency and accuracy.

Main Methods:

• Depth maps are converted into 3D point clouds.
• Viewpoint Feature Histogram (VFH) descriptors are computed using the Point Cloud Library.
• Human activities are classified using k-NN with Dynamic Time Warping and bidirectional long short-term memory (BiLSTM) networks.
• Techniques like transfer learning and fuzzy integral are employed to enhance BiLSTM performance.

Main Results:

• The proposed method achieves classification results comparable to or better than the current state-of-the-art on two benchmark datasets.
• Efficiency improvements for k-NN classification were achieved through a two-tier model.
• BiLSTM performance was enhanced via transfer learning and network fusion.

Conclusions:

• 3D point clouds derived from depth maps are effective for human activity classification.
• Both k-NN and BiLSTM classifiers show promise, with potential for further optimization.
• The study demonstrates a robust and efficient approach to HAR.