DENSIFY CAMPUS CONGESTION PREDICTION

SIMULATION REPLAY (DEMO) 90-Day Telemetry · Model Validated
Simulated Time 13:00
Total Active Campus Crowd Ground Truth
3,840 / 8,000 Capacity
Peak Density Zone Primary
Block A 77.9%
Campus Energy Draw Simulated Proxy
286.4 kWh
Bottleneck Alerts Primary
0 Severe Zones
Transit Shuttle Headway Queuing Model
4.8 min avg wait
Campus Architectural Layout & Real-Time Density Heatmap SIMULATED TELEMETRY REPLAY
Interactive campus architectural layout and crowd density heatmap
Normal (<45%)
Moderate (45-70%)
High (70-85%)
Severe Bottleneck (>85%)

Edge Computer Vision & Multi-Object Pedestrian Tracking Pipeline

Optical detection workflow: High-speed frame extraction, YOLOv8-Nano object detection, persistent DeepSORT centroid tracking, and virtual tripwire ROI crossing logic that converts raw optical pixel streams into validated zone headcount and directional ingress/egress rates. Usable with synthetic CCTV corridor, live camera, or video file.

CCTV Camera Feed & Optical Detection HUD
Ingress In (+1)
42
Egress Out (-1)
18
Net Live Occupancy
24
Active Track IDs
7
Vision AI Pipeline: Pixels to Validated Zone Counts

High-Speed Stream Ingestion

STAGE 01 // INGESTION

Optical RTSP video feed, local USB camera, or simulated corridor capture buffer sampled at 24.2 FPS with frame buffer normalization.

YOLOv8 Deep Neural Object Detection

STAGE 02 // INFERENCE

Single-stage convolutional inference targeting person class bounding boxes [xmin, ymin, xmax, ymax] with confidence score threshold ≥ 0.80 (80%).

DeepSORT Centroid & Trajectory Tracking

STAGE 03 // TRACKING

Kalman filter motion state estimation and cosine appearance distance association maintains persistent identities (e.g. Track #104) across occlusions.

Virtual Tripwire ROI Directional Logic

STAGE 04 // SPATIAL ROI

Calibrated geometric crossing vector registers direction: downward crossing increments Ingress (+1), upward crossing increments Egress (-1).

Digital Twin Headcount Integration

STAGE 05 // TWIN SYNC

Net accumulated headcount (Ingress − Egress) streams synchronously into LIB_01, driving spatial heatmaps and ML forecasts.

Verifiable ML Forecast Methodology & Provenance

Historical Dataset: Trained on 90 consecutive academic days (N = 15,120 hourly records) across 7 facilities.
Forecast Horizon: Rolling 24 consecutive hours starting from selected simulated time (13:00 Today → 12:00 Next Day).
90% Confidence Bounds: Computed analytically as ŷ ± 1.645 × RMSE (±4.9% error margin) derived from the test-set holdout Root Mean Squared Error.

24-Hour Machine Learning Occupancy Forecast

Ensemble Random Forest regression curve with 90% confidence bounds, test-holdout actuals, and automated bottleneck detection

24-Hour Machine Learning Occupancy Forecast Chart

Facility Capacity vs Headcount Demand

Design limit vs Nominal Baseline vs Simulated What-If load

Facility Capacity vs Headcount Demand Chart

Campus Energy & HVAC Load Profile

Predicted hourly power draw (kWh) correlated with human thermal output and ambient weather

Campus Energy and HVAC Power Draw Chart
Scenario Parameter Controls
Final Examination Period: Inactive
Peak Congestion
77.9%
Baseline Nominal
Severe Bottlenecks
0
0 Severe Zones
Extra Energy Draw
+0 kWh
+0.0 kg CO2e
Max Transit Delay
4.8 min
Normal Headway

AI Administrative Dispatch Directives & Mitigation Actions

Facility Nominal Cap Baseline Occupancy Simulated Occupancy Crowd Delta Transit / Queue Status Severity

Facilities Telemetry & Derived Physical Proxy Metrics

Headcount & Occupancy: Primary ground-truth physical metric measured across all facilities.
Energy, CO2, and Wi-Fi Probes: Secondary physical proxies modeled from occupant metabolic emissions (100W sensible thermal heat, 0.005 L/s CO2 per person in ASHRAE 62.1 spaces) and 1.24x device-to-person probe ratios.

Dataset Provenance & Temporal Holdout Split Methodology

Dataset Size: 15,120 records (90 calendar days × 24 hourly intervals × 7 campus facility zones).
Chronological Split: Days 1–72 (80%, N = 12,096) strictly for training; Days 73–90 (20%, N = 3,024) held out as strictly future unseen test data.
Zero Cross-Period Data Leakage: Random cross-validation is barred for time-series telemetry because random shuffling leaks same-day future hours into training. The chronological holdout guarantees evaluation on unseen future temporal horizons.

Machine Learning Model Architecture & Performance Benchmarks

Evaluation metrics on 15,120 historical operational records across 7 campus facilities (80/20 chronological train-test split)

Model Algorithm Task RMSE (Root Mean Sq Err) MAE (Mean Abs Err) R² Score (Variance Explained) Selection Status

Random Forest Feature Importance

Gini impurity reduction across operational parameters

Random Forest Feature Importance Chart

Model Validation: Actual vs Predicted Occupancy

Comparison on unseen chronological test holdout samples

Model Validation: Actual vs Predicted Occupancy Chart

Congestion Risk Classifier: Confusion Matrix

Classification Accuracy: 93.62% | Weighted F1-Score: 0.9323 across 4 operational severity states (Chronological Test Holdout)

True \ Predicted Low (<45%) Moderate (45-70%) High (70-85%) Severe Bottleneck (>85%)