Crowd Counting with Semantic Scene Segmentation in Helicopter Footage.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Recent crowd counting neural networks achieve high accuracy but lack semantic understanding.
• Existing methods struggle with nuanced roles (e.g., pedestrian, student) and contextual information.
• Foolproof autonomous systems require robust semantic understanding beyond simple counts.

Purpose of the Study:

• To improve pedestrian counting accuracy in helicopter footage by incorporating semantic context.
• To introduce a new dataset specifically for pedestrian counting from aerial videos.
• To demonstrate the simultaneous achievement of crowd counting and semantic segmentation using a unified neural network.

Main Methods:

• Utilized semantic segmentation to extract contextual information from surrounding entities.
• Developed a new dataset from helicopter videos for pedestrian counting tasks.
• Employed hard parameter sharing within a single neural network for joint crowd counting and semantic segmentation.

Main Results:

• Demonstrated that incorporating semantic segmentation significantly increases pedestrian counting accuracy.
• Showcased that crowd counting and semantic segmentation can be performed simultaneously with comparable or improved accuracy.
• Validated the generic applicability of the proposed method to various crowd density estimation techniques.

Conclusions:

• Integrating semantic segmentation is a viable method for enhancing crowd counting accuracy in aerial imagery.
• A single neural network can effectively perform both crowd counting and semantic segmentation tasks.
• The developed approach offers a robust solution for improving the semantic understanding of crowds in autonomous systems.