Adaptive temporal compression for reduction of computational complexity in human behavior recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video analytics and human behavior recognition are increasingly important in fields like virtual reality and surveillance.
• Three-dimensional convolution (3D CNN) is now standard for extracting spatio-temporal features in video analysis.
• 3D CNNs face challenges including high parameter counts, computational complexity, and GPU dependency, slowing down training.

Purpose of the Study:

• To address the computational challenges of 3D CNNs in human behavior recognition.
• To propose an efficient module for reducing computational load and improving training speed.
• To facilitate real-time human behavior recognition through optimized video analysis.

Main Methods:

• Development of an Adaptive Time Compression (ATC) module.
• Integration of ATC as an independent component into existing deep learning architectures.
• Implementation of data compression by eliminating redundant video frames.

Main Results:

• The ATC module significantly reduces GPU computing load and time complexity.
• Negligible loss in accuracy was observed despite data compression.
• The module enables faster and more efficient spatio-temporal feature extraction.

Conclusions:

• The Adaptive Time Compression module offers an effective solution to the computational demands of 3D CNNs.
• ATC facilitates real-time human behavior recognition by optimizing video processing.
• This approach enhances the practicality of deep learning models for video analytics applications.