Weakly Supervised Violence Detection in Surveillance Video.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Systems

Background:

• Automatic violence detection is crucial for public safety, but manual surveillance monitoring is prone to human error.
• Existing methods often lack spatial localization, focusing only on short clip classification.
• The challenge lies in efficiently and accurately identifying violent events in extensive video data.

Purpose of the Study:

• To develop a weakly supervised method for simultaneous spatial and temporal violence detection in surveillance videos.
• To overcome the limitations of previous studies by incorporating spatial localization.
• To utilize only video-level labels for training, simplifying data annotation.

Main Methods:

• A temporally extended Fast-RCNN architecture was employed for action tube generation.
• Leveraged pre-trained person detectors, dynamic images, and tracking algorithms for proposal generation.
• Extracted spatiotemporal features using deep neural networks and applied multiple-instance learning for classification.

Main Results:

• Achieved high accuracy on public datasets: 97.3% on Hockey Fight, 92.88% on RLVSD, and 88.7% on RWF-2000.
• Demonstrated comparable performance to state-of-the-art methods.
• Successfully performed both spatial localization and temporal detection of violent actions.

Conclusions:

• The proposed weakly supervised method effectively detects spatially and temporally violent actions in surveillance footage.
• This approach offers a robust solution for enhancing security through automated video analysis.
• The method's reliance on video-level labels makes it practical for real-world applications.