Leveraging a Neuroevolutionary Approach for Classifying Violent Behavior in Video.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Automated surveillance analysis is crucial for security but limited by human resources and judgment.
• Machine learning, particularly neural networks, offers automated solutions for video analysis.
• Designing effective neural network architectures and hyperparameters is a significant challenge.

Purpose of the Study:

• To propose a novel model that automatically generates neural networks for behavior classification in videos using a genetic algorithm.
• To develop and evaluate both shallow and deep neural network architectures for this task.
• To optimize the process of neural network design for efficient computational resource utilization.

Main Methods:

• A genetic algorithm was employed to evolve neural network architectures for behavior classification.
• Two types of networks were developed: shallow networks (dense layers) using direct encoding and deep networks (3D convolutional layers) using indirect encoding.
• Input data varied based on network type: pose evolution for shallow networks and video sequences for deep networks.

Main Results:

• The proposed model successfully generated neural networks capable of behavior classification in videos.
• Relevant results were achieved on the Kranok-NV dataset.
• The evolved networks were evaluated using standard classification metrics, demonstrating effectiveness.

Conclusions:

• The genetic algorithm approach provides an effective method for automating the design of neural networks for video behavior classification.
• The study demonstrates the feasibility of generating both shallow and deep network architectures tailored to specific input data types.
• This automated approach has the potential to improve the efficiency and accuracy of surveillance systems.