Academic project report / June 10, 2024
Multi-Camera Re-identification and Tracking
An edge-oriented person detection and re-identification prototype that shares embeddings instead of streaming every camera feed to a central server.
Abstract
This academic project combines YOLOv8 Nano person detection with a ResNet-50 re-identification model on Raspberry Pi devices. Each camera processes video locally and exposes embeddings over HTTP, allowing a central monitor to associate observations using cosine similarity without receiving continuous raw video. The report evaluates quantization, accuracy, runtime, deployment constraints, and the privacy consequences of making cross-camera tracking more accessible.
Problem
Multi-camera person tracking usually sends every video feed to a central system. That design is straightforward, but bandwidth and centralized compute requirements grow with every camera. This project explored a distributed alternative: perform detection and re-identification on each edge device, then exchange compact model outputs rather than continuous raw video.
System
Each Raspberry Pi runs YOLOv8 Nano to detect people and a ResNet-50 model trained on Market-1501 to generate re-identification embeddings. Camera frames, crops, bounding boxes, and embeddings are made available over HTTP for a monitoring interface. Cross-camera matches use cosine similarity between embeddings.
Both models were quantized from FP32 to INT8 for constrained hardware. The report found that the ResNet-50 model retained 79% mean average precision after quantization, while measured runtime fell by approximately 49% for object detection and 37% for re-identification.
Collaboration
This project and report were completed by Eric Ye and Ryan Rumana for Stanford’s EE292D course. The work was also invited for presentation at Stanford’s ACT4AERO workshop.
Limitations and ethics
The prototype does not claim production-grade persistent identity tracking. The report identifies future work including stronger temporal association, Kalman filtering, camera-specific calibration, more efficient model families, and relative camera localization.
It also treats privacy as a first-class concern. Moving inference to the edge may reduce raw-video transfer, but accessible person re-identification can still enable pervasive surveillance. Any deployment needs explicit governance, retention limits, access controls, and a clear justification for tracking people across spaces.