Counting people with low-level features and Bayesian regression.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Estimating crowd size in videos with diverse pedestrian movement is challenging.
• Traditional methods often rely on object detection or tracking, which can be computationally intensive and fail in dense or occluded scenarios.

Purpose of the Study:

• To develop a robust and efficient crowd counting method for inhomogeneous crowds.
• To estimate pedestrian counts without explicit object segmentation or tracking.

Main Methods:

• Segmentation of crowds into homogeneous motion components using a mixture of dynamic-texture motion model.
• Extraction of holistic low-level features from segmented regions.
• Learning a feature-to-count mapping using Bayesian regression, specifically exploring Gaussian process regression and a Bayesian Poisson regression model with a derived closed-form approximation.

Main Results:

• The proposed regression-based methods achieve accurate crowd size estimation across various crowd densities.
• The methods outperform state-of-the-art pedestrian detectors in accuracy.
• Demonstrated efficiency and robustness over 2 hours of video data with diverse scenarios and outliers.

Conclusions:

• The proposed approach offers a reliable alternative for crowd counting, especially in complex, dynamic environments.
• Bayesian regression models, particularly the approximated Poisson regression, provide accurate and computationally efficient crowd size estimates.