People Tracking and Counting

In surveillance systems, counting people passing through monitored areas constitutes a significant research challenge. This process typically relies on background subtraction operations. However, background differencing methods face two primary challenges: dynamic background model estimation and shadow removal. To address these issues, a bidirectional counting algorithm has been proposed. For developing a more robust counting system, Gaussian Mixture Models (GMM) are implemented to describe background scenes using probability density functions composed of multiple Gaussian distributions. Nevertheless, this algorithm cannot provide classification methods for detecting shadows within moving foreground objects. To enhance performance, background models are integrated with color models (typically using HSV or YCbCr color spaces for better illumination invariance). This combination delivers improved motion object detection by eliminating shadows from foreground elements through chromaticity analysis. A multi-class feature tracking algorithm (employing features like SURF, ORB, or color histograms) is applied to handle occlusion problems in multi-object tracking scenarios. To improve counting accuracy in individual schemes, we propose bidirectional counting approaches requiring multi-level backward tracking procedures involving trajectory analysis and re-identification techniques. This counting system architecture demonstrates potential for higher accuracy and improved performance even in crowded scenes and dynamically changing environments.