AI CCTV Analytics System
Multi-Camera Re-Identification Pipeline
Real-time multi-camera occupancy tracking fixing identity loss and duplicate counts with 90% re-ID accuracy across retail zones.
The Engineering Challenge
High-traffic retail and QSR stores needed reliable occupancy tracking and customer journey heatmaps across 8+ non-overlapping camera feeds.
Core ProblemExisting tracking suffered from massive duplicate counting whenever visitors traversed blind spots between cameras or experienced occlusion.
- Process 8+ 1080p video feeds at 30 FPS in real time on edge hardware
- Re-identify individuals across different lighting and camera angles
- Zero cloud bandwidth streaming — edge-first processing
System Architecture
End-to-end data pipelines and component isolation boundaries.
YOLOv11 object detector coupled with BoT-SORT / ByteTrack, accelerated via TensorRT FP16 and NVDEC hardware video decoding, feeding an ONNX re-identification feature extractor.
Key Engineering Decisions
Corrected ONNX ReID Input Dimensions & Normalization
Diagnosed and resolved subtle aspect ratio distortion in image preprocessing that previously degraded feature embeddings by 35%.
NVDEC + TensorRT FP16 Acceleration
Moved decoding to dedicated hardware NVDEC engines and converted weights to TensorRT FP16, increasing edge throughput by 2.8x.
Dual Re-ID Confidence Thresholding
Implemented spatiotemporal velocity gating to discard physically impossible transitions across camera boundaries.
Production Verification & Results
- Achieved 90% multi-camera re-identification accuracy across active retail zones
- Completely eliminated duplicate counts caused by camera handoff blind spots
- Sustained 30 FPS inference per camera stream on single edge workstation
Production Stack & Tooling
Next Iterations & Roadmap
- Graph Neural Network (GNN) trajectory association across wider camera blind spots
- Edge-to-cloud telemetry sync for multi-store executive analytics
Want to build something similar?
I specialize in taking complex agentic workflows, duplex voice architectures, and high-throughput computer vision models into production.