Table 2.

Feature extraction specifications by sensing Modality.

Sensing modality	CNN/Processing architecture	Extracted features	Feature dimension	Design application
Visual Cameras	ResNet-50 backbone	Crowd density maps, movement vectors, activity heatmaps	2048-dim embedding	Capacity management, flow optimization
Visual Cameras	YOLO v5 object detection	Person count, spatial distribution, dwell time	512-dim per region	Space utilization analysis
Acoustic Sensors	MFCC extraction (13 coefficients)	Frequency signatures (50–8000 Hz), sound pressure levels	39-dim (MFCC + Δ + ΔΔ)	Activity level detection
Acoustic Sensors	Spectral analysis	Social interaction indices, ambient noise	64-dim spectral features	Comfort assessment
Environmental	Statistical aggregation	Temperature (± 0.1 °C), humidity (± 2% RH), PM2.5, illuminance	8-dim parameter vector	Environmental quality optimization
Motion Detectors	Binary pattern analysis	Trajectory data, pathway usage frequency	128-dim spatial grid	Circulation planning

This layer utilizes convolutional neural networks for visual feature extraction, spectral analysis algorithms for acoustic processing, and statistical methods for environmental parameter characterization. The mathematical foundation for feature extraction can be expressed through the transformation function:.