Fight Recognition in video using Hough Forests and 2D Convolutional Neural Network

Area of Science:

• Computer Vision
• Machine Learning
• Behavior Analysis

Background:

• Action recognition is crucial for computer vision, but detecting specific events like violence remains challenging.
• Existing methods for violence detection often struggle with computational efficiency, hindering practical application in surveillance.
• Deep learning approaches, like 3D Convolutional Neural Networks, process entire video sequences but may include redundant information.

Purpose of the Study:

• To develop a computationally efficient and accurate method for detecting aggressive behaviors in videos.
• To address the limitations of existing handcrafted and deep learning feature extraction techniques for violence detection.
• To propose a hybrid feature framework that leverages both handcrafted and learned features.

Main Methods:

• A novel hybrid framework combining "handcrafted" and "learned" features for action recognition.
• Utilizing spatiotemporal features that balance motion and appearance information.
• Evaluating the proposed method on three benchmark datasets for violence detection.

Main Results:

• The hybrid method achieves higher accuracy compared to previous feature learning techniques.
• The proposed approach demonstrates comparable computational efficiency to existing methods.
• The method outperforms state-of-the-art approaches on two out of three benchmark datasets.

Conclusions:

• Hybrid feature frameworks offer a promising direction for efficient and accurate violence detection.
• Focusing on crucial motion features, combined with learned representations, enhances performance.
• The developed method has significant potential for real-world video surveillance applications.