What if the camera feed drops?
The capture loop marks the stream unavailable and stops creating observations.
- Signal
- Frame timestamp age exceeds the expected interval.
- Response
- Recover the feed before treating missing telemetry as animal behavior.
Live
Computer Vision / AI2024
CowRec System
Moved livestock recognition and behavioral analysis to the edge so raw video can stay local while useful herd telemetry moves outward.
- Role
- Computer vision and backend engineer
- Environment
- Edge computer-vision system
- Ownership
- Detection pipeline · Tracking · Edge processing · Telemetry model · Backend API · Containerization
cowrec-edge5 system boundaries
Barn Cameras
Frame Capture
YOLOv8 Tracker
Animal Registry
Edgeprimary inference location
YOLOv8detection and tracking model
Localraw-video processing boundary
APIaggregated telemetry surface
01 / Interactive system
Move through the system layer by layer
cowrec-edgeEdge computer-vision system · Read-only
Live
System map
Boundaries and responsibilities
Public or untrustedAuthenticated boundaryInternal-only dependency
Barn Cameras selected.
02 / Context
The problem behind the system
Precision livestock work needs individual-animal signals, but sending continuous barn video to a central platform is expensive and privacy-heavy.
CowRec places inference near the cameras, then promotes only the tracking and behavioral information needed for farm decisions.
Read the full project overview
A computer vision system designed for precision livestock farming. It uses YOLOv8 and convolutional neural networks to process real-time video feeds, allowing for the identification, tracking, and behavioral analysis of individual animals.
The architecture runs in edge computing environments, collecting health and productivity data locally to enable data-driven decisions in dairy production.
03 / Decisions
The trade-offs that shaped the build
DecisionInference at the edge
Keep the expensive and sensitive stream near its source.
- Reason
- The useful output is behavior data, not centralized raw footage.
- Trade-off
- Deployment and model updates must reach distributed edge boxes.
- Revisit when
- When connectivity, cost, and governance justify centralized processing.
DecisionAggregate before export
Move derived signals, not frames.
- Reason
- Reduces bandwidth and limits the external data surface.
- Trade-off
- Remote teams cannot re-run inference on the original footage.
- Revisit when
- For explicitly consented diagnostic capture workflows.
DecisionLow confidence means review
Do not turn uncertainty into a hard result.
- Reason
- Barn conditions vary and identity mistakes compound over time.
- Trade-off
- Some observations remain unresolved.
- Revisit when
- As calibration data and model quality improve.
04 / Operations
The unhappy path is part of the design
What if confidence falls below threshold?
The system withholds a strong identity or behavior conclusion.
- Signal
- Low-confidence detections increase in the local runtime.
- Response
- Review camera position, lighting, calibration, and model fit.
05 / Implementation
How the decisions appear in the build
edge/detect.py
01# Real-time herd inference at the edge. Raw video never leaves the farm.02from ultralytics import YOLO03 04model = YOLO("weights/cowrec-v8.pt") # fine-tuned on the on-farm dataset05 06for frame in stream.read(): # RTSP feed from barn cameras07 result = model.track(frame, persist=True, conf=0.6, verbose=False)[0]08 09 for box, tid in zip(result.boxes.xyxy, result.boxes.id):10 cow = registry.resolve(int(tid)) # stable per-animal identity11 cow.observe(bbox=box, at=frame.ts)12 if cow.gait_anomaly(): # early lameness / health signal13 alerts.enqueue(cow.id, kind="gait", severity="review")14 15 # Only aggregated telemetry leaves the edge box - never raw frames.16 telemetry.flush(registry.snapshot())Representative / sanitized excerpt
Real-time per-animal tracking at the edge - only aggregates leave the box, never raw video.
PythonYOLOv8OpenCVFastAPIDocker
06 / Security
Threats, controls, and what remains
No control is presented as total risk elimination.
Raw farm video leaves the edge
- Control
- Local inference and aggregate-only telemetry design
- Residual risk
- Edge device compromise can still expose local frames.
Tracking identity drifts
- Control
- Persistent IDs plus confidence and unresolved-state handling
- Residual risk
- Occlusion can still require human review.
Network interruption stops analysis
- Control
- Inference and state processing remain local
- Residual risk
- Remote telemetry is delayed until connectivity returns.
07 / Evidence
Proof, source, and inspectable outcomes
08 / Results
What the project demonstrates
Raw video processing is kept at the edge
Per-animal observations become useful aggregated telemetry
Network loss does not automatically stop local inference
- The data boundary matches the useful product output.
- Confidence-aware behavior keeps uncertainty visible.
- Stronger long-term identity evaluation
- Field calibration workflow
- Edge deployment observability
- Computer vision systems
- Edge architecture
- Privacy-aware data reduction