Aquaculture Crayfish Behavior & Fighting Recognition
Detect fighting and cannibalism risks in aquaculture tanks with on-device AI.
Autonomous 4-stage funnel pipeline running entirely on reCamera edge devices — detects aquaculture crayfish fighting, molting cannibalism risks, and sex ratios with zero video leaving the camera.
Reusable Site Templates
Fixed overhead tank cameras, high stocking densities, and accidents occurring in seconds.
High-Density Fattening & Molt Protection
RAS tanks and holding ponds. Alerts when soft-shell molting individuals are attacked.
Broodstock Sex Ratio & Density Monitoring
Tracks sex balance and interaction intensity in breeding tanks.
Aquaculture Research & Behavior Logging
Automated 24/7 digital event logger recording aggression frequencies and durations without manual watching.
Not a fit: Heavy reflection dead zones, extremely turbid water. Dead individuals raise no fighting events.
How it works
The edge AI camera monitors the tank, confirms aggression locally via a 4-stage funnel pipeline, and alerts the farm dashboard.
Two things: the edge AI camera that produces the picture and runs detection, and whatever consumes the alert. Everything runs locally on the camera; video never leaves the device. What crosses the network is a few hundred bytes of JSON.
Camera & Placement
Mounted 0.8m–1.5m overhead or angled at 45°, streaming 1080p @ 10–15 FPS. Two hardware tiers:
- reCamera Pro (1/1.8" Large Sensor): Superior sensitivity and clarity for wide tanks, multi-target tracking.
- reCamera 2002 (1/2.8" Standard Tier): Compact, cost-optimized for high-density standard tank setups with PoE.
Edge Detection Engine
The pipeline runs entirely inside the camera with zero cloud streaming:
| Model | Sensor & Compute | Engine | FPS | Best For |
|---|---|---|---|---|
| reCamera Pro | 1/1.8" · RV1126B · 3 TOPS | Python Kit Full 4-Stage | 10–15 FPS | Wide tanks, 8-frame mosaic temporal validation |
| reCamera 2002 | 1/2.8" · SG2002 · 1 TOPS | Native C++ cvimodel | 12–15 FPS | Standard tanks, massive scale, fight triage |
Alert & Telemetry Exits
- MQTT Broker: Topic
crayfish/behavior/eventsends structured state JSON. - WebSocket Protocol: Port 8124 serves triggered interaction ROI image crops.
- RTSP Stream: Port 8554 with hardware OSD bounding boxes.
How well it works
Real on-device visual outputs and structured events captured directly by the camera.
Measured results
Engineering benchmark on real aquaculture dataset.
| Metric | Typical | Setup |
|---|---|---|
| Fighting recall | 95.0% | reCamera Pro (8-frame temporal) |
| 2D overlap false positive rejection | 90.5% | reCamera Pro |
| Trigger to alert latency | 66 ms | On-device MQTT |
| Defocus blur detection rate | 88.0% | Augmented with 15% defocus blur |
| Sex classification accuracy | 82.0% | Top-1 via male claw patch |
What you get out of it
Once deployed, these topics are the whole interface.
| Topic / Port | Payload | Retained |
|---|---|---|
crayfish/behavior/event | State JSON: state (start/change/end), fight_detected, confidence, pair_ids, duration_sec | No |
crayfish/detection | Bounding boxes: crayfish_count, sex_ratio, plus bbox per track | No |
| WebSocket 8124 | Binary snapshot stream of triggered interaction ROI | — |
RTSP 8554 /live0 | Live video with OSD bounding boxes | — |
Porting It to Your Own System
The underlying architecture forms a reusable pattern: underwater/dense target detection → geometric proximity trigger → ROI crop → temporal verification.
| Layer | What porting costs you |
|---|---|
| Camera placement rules and diffuse lighting requirements | reuse as-is |
| Edge runtimes (RV1126B RKNN / SG2002 CVI TPU) | reuse as-is |
| Geometric proximity triggers & debounced state machine | reuse as-is |
| MQTT / WebSocket / RTSP topic interfaces | reuse as-is |
| Object detection weights | retrain for new species |
| Temporal aggression thresholds | calibrate against target species |
How to deploy
Check the setup and its product list, then follow the deployment guide.
Hardware Sizing & Deployment Setup
Select site conditions to generate recommended bundles and guides, or let an AI assistant estimate hardware and deploy for you.
Deployment Options
reCamera Pro (1/1.8" Sensor · High-Res Wide Tank & Multi-Target)
1/1.8" large-sensor 4K/1080p capture with 3 TOPS NPU. Covers wide breeding tanks with clear details, running full 4-stage funnel with 8-frame mosaic temporal confirmation and sex voting.
1 steps · ~10min
AI Camera
recamera_pro
FAQ
Yes. Detection, tracking, proximity trigger, and temporal fight verification all run inside the reCamera onboard accelerator. No internet access is required.
No. Video remains strictly local on the camera. Only lightweight JSON alerts (a few hundred bytes) and triggered snapshot crops traverse the network.
The models are trained with 15% defocus blur augmentation, and the 4th stage 8-frame temporal mosaic specifically rejects 2D perspective alignment errors.
Choose reCamera Pro (1/1.8" Large Sensor) for wide tanks, multi-target tracking, and full temporal verification; choose reCamera 2002 (1/2.8" Standard Tier) for high-density standard tanks and cost-sensitive PoE scaling.
Didn't find your answer?
Contact Us
We Are Glad to Be Your Hardware Partner !