Extract the Relational Information of Static Features and Motion Features for Human Activities Recognition in Videos

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Static and motion features are crucial for human activity recognition.
• Existing methods often lack sufficient information for complex activities.
• The Bag-of-Word (BoW) model is a common representation but can be improved.

Purpose of the Study:

• To propose a novel method for human activity recognition by extracting relational information from static and motion features.
• To enhance the discriminative power of the Bag-of-Word model for complex activities.
• To improve the accuracy and robustness of human activity recognition systems.

Main Methods:

• Utilizing a divisive algorithm based on KL-divergence for codebook reconstruction.
• Constructing a bipartite graph to model relationships between different feature sets.
• Applying k-way partitioning to create a new, relationally-rich codebook.
• Representing videos with enhanced Bag-of-Word vectors incorporating relational information.

Main Results:

• The proposed method achieves promising results on several human activity recognition datasets.
• The enhanced Bag-of-Word representation captures stronger relational information.
• The new codebook construction leads to more discriminative video representations.

Conclusions:

• Extracting relational information from static and motion features significantly improves human activity recognition.
• The proposed bipartite graph and k-way partitioning method offers a robust approach to enhancing BoW models.
• This work provides a valuable contribution to the field of complex human activity recognition.