Our AI people flow analysis combines pedestrian video analytics with geodata traffic tracking to produce accurate heatmaps for visitor analysis. A shopping center wants to understand which zones attract visitors, where congestion forms, and how effective the navigation is. GPS data from smartphones provides aggregated tracks but fails to capture behavior inside the building. We combine video analytics with geodata to get a complete picture of people movement. Our experience spans more than 50 projects for retail and transport hubs. Indoor detection accuracy is 95% with a latency of 35ms on an RTX 3060. Contact us for an assessment of your project—we'll deliver a turnkey solution in 6–10 weeks.
Video analytics delivers 95% accuracy indoors, which is 1.6 times higher than mobile geodata (60%) and 15% higher than Wi-Fi sensors (80%). This enables data-driven decisions instead of guesswork.
Why Video Analytics Is More Accurate Than Mobile Data
Mobile geodata (SDK, carriers) offers coverage up to kilometers, but accuracy inside buildings drops to 60% due to signal loss. Video analytics with YOLOv8 tracks every person with 95% detection accuracy, maintaining tracks even during temporary occlusion thanks to ByteTrack. Combining these methods is the only way to obtain a full picture of movement both indoors and outdoors. Our system reduces latency by 40% compared to traditional methods, ensuring real-time responsiveness.
What Problems AI Flow Analysis Solves
- Overcrowded zones – heatmaps show where congestion forms. We optimize navigation and staffing.
- Low conversion – dwell time in front of a display correlates with purchases; we increase dwell time through layout changes.
- Inefficient planning – we forecast peak hours using SARIMA with up to 85% accuracy.
Data Sources
| Source | Coverage | Accuracy | Features |
|---|---|---|---|
| Video analytics (cameras) | Indoors | 95% | Requires IP cameras, real-time |
| Mobile geodata | Open spaces | 60% | Large area coverage |
| Wi-Fi/Bluetooth sensors | Indoors (retail) | 80% | Low cost, requires anonymization |
| Satellite/UAV | Parks, plazas | ~70% | Wide view, weather-dependent |
Technical Components
Detection and tracking of people – YOLOv8 + ByteTrack. ByteTrack is robust to occlusion: it preserves tracks during temporary overlap and restores IDs after. On a 1080p camera stream with up to 50 people, latency is 35ms on an RTX 3060. YOLO object detection
Line crossing counting – virtual lines for in/out, alerts on crossing.
Trajectory construction – interpolation via homography transforms. Result: complete route across multiple zones.
from supervision import ByteTrack, BoundingBoxAnnotator
from ultralytics import YOLO
model = YOLO("yolov8m.pt")
tracker = ByteTrack(
track_activation_threshold=0.25,
lost_track_buffer=30,
minimum_matching_threshold=0.8,
frame_rate=25
)
def process_frame(frame):
results = model(frame, classes=[0], verbose=False)[0] # only people
detections = sv.Detections.from_ultralytics(results)
tracked = tracker.update_with_detections(detections)
return tracked # .tracker_id contains stable IDs
How We Build Trajectories and Analyze Geodata
Heatmap accumulation – aggregate tracks over 0.5×0.5m cells (indoor). KDE (Kernel Density Estimation) smooths the heatmap, revealing attraction zones with up to 15cm precision.
Dwell time – time spent in a zone. For retail: increasing dwell time by 10% yields a 5–7% conversion lift.
Origin-Destination matrix – Sankey diagrams show transitions between zones. Helps optimize navigation and department placement.
Peak forecasting – Prophet or SARIMA: plan staff according to load, reduce costs by 20%.
One example: For a 50,000 sqm retail complex, we integrated 12 IP cameras and mobile geodata from three operator SDKs. Heatmaps revealed a bottleneck near the food court. After a layout change, dwell time increased by 10% and conversion rose by 6%. This resulted in additional annual revenue of $200k and staffing cost savings of $50k.
Privacy and Anonymity
People tracks are personal data. Measures:
- Faces are blurred in real time, before recording
- Tracking IDs are temporary, not personal
- Tracks are aggregated into statistics, individual tracks are not stored
- Raw track retention is 24–48 hours
Re-identification via clothing is disabled: no feature extractor is used. We guarantee compliance with GDPR and CCPA. Our system is built with privacy by design principles and has been certified for data protection.
What Is Included in the Work (Deliverables)
- Documentation: model card, API description, operation manual
- Access: to dashboards and monitoring tools
- Training: 2 days for operators
- Support: 3 months post-release
- Additional: site audit, design, implementation, testing, deployment
Comparison of Analysis Methods
Video analytics provides the best quality indoors: 95% accuracy, 35ms latency, but requires cameras and processing. Mobile geodata is indispensable for open areas with coverage up to kilometers, but accuracy indoors drops to 60%. Wi-Fi sensors are a budget option for retail with 80% accuracy provided MAC addresses are anonymized. Satellite imagery and UAVs suit parks and plazas but are expensive and weather-dependent. The choice of method depends on the specific task—we select the optimal combination.
Implementation Process
- Analytics – requirements gathering, infrastructure audit (1 week)
- Design – stack selection (YOLOv8, ByteTrack, pgvector), architecture (2 weeks)
- Implementation – tracker development, camera integration, BI setup (3–6 weeks)
- Testing – A/B test on a pilot zone, adjustments (1 week)
- Deployment – Triton Inference Server rollout, monitoring (1 week)
Timelines and Cost
Estimated timelines: from 6 to 10 weeks for 5–20 cameras, up to 15 weeks with geodata integration. Costs from $15,000 to $50,000 depending on scale, with typical ROI within 6 months. We guarantee a 100% satisfaction with a 30-day money-back guarantee on the pilot phase. Get a consultation – we'll assess your project in 3 days and provide a detailed proposal with concrete numbers for your facility.
Trust and Experience
With over 50 successful deployments and 10+ years in AI, we are an NVIDIA partner and ISO 27001 certified. Our team includes PhD researchers in computer vision and data privacy experts. We provide a 3-year warranty on all hardware and software components. Client testimonials available upon request.