Learning Spatio-Temporal Representations for Action Recognition: A Genetic Programming Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Effective feature extraction is crucial for accurate human action recognition from video.
• Traditional handcrafted features often lack robustness and adaptability.
• Automatic learning of spatio-temporal features is a key research direction.

Purpose of the Study:

• To develop an automated method for learning discriminative and robust spatio-temporal motion features for human action recognition.
• To utilize genetic programming (GP) for evolving optimal feature descriptors.
• To achieve data-adaptive feature extraction mimicking human visual processing.

Main Methods:

• Employed genetic programming (GP) to evolve motion feature descriptors from primitive 3D operators.
• Integrated multi-layer descriptors to enhance data adaptivity and mimic human visual cortex structure.
• Used average cross-validation classification error from a support-vector-machine classifier as the GP fitness function.
• Extracted scale and shift invariant features from color and optical flow sequences.

Main Results:

• The GP-evolved features demonstrated superior performance compared to existing methods.
• The method achieved significant improvements on four popular action recognition datasets (KTH, HMDB51, UCF YouTube, Hollywood2).
• The learned descriptors were effective across different datasets and video types.

Conclusions:

• The proposed GP-based approach offers an effective way to automatically learn spatio-temporal features for human action recognition.
• This method provides a robust and data-adaptive alternative to handcrafted features.
• The evolutionary strategy accelerates convergence to optimal feature descriptors.