Aquaculture Crayfish Behavior Recognition: 4-Stage Pipeline and Edge Benchmarks

Usage Notice

This is an assistive aquaculture monitoring and early warning system, not a sealed industrial automation safety cutoff device. Integration with dissolved oxygen sensors and automated feeding systems is advised.

What this solution does

System Architecture & Topology

Two things are required: the edge AI camera monitoring the tank, and the dashboard receiving events. Everything runs locally on the camera over standard IoT protocols without centralized GPU servers.

In high-density recirculating aquaculture systems (RAS) or breeding tanks, Australian freshwater crayfish (Cherax quadricarinatus) frequently suffer limb loss or fatal cannibalism during molting periods due to aggressive attacks. An edge AI smart camera mounted above the tank tracks each crayfish in real time. When two individuals approach and engage in combat, the camera confirms aggression within an 8-frame temporal window and publishes an MQTT alert to the farm dashboard, prompting automated feeding or staff intervention to protect soft-shell individuals.

• Open Source Repository: github.com/suharvest/crayfish-fight
• Sizing and Deployment: Reference Design Deployment Guide

• 4-Stage Funnel Pipeline

  Detection → Proximity Trigger → Single-frame Filter → 8-frame Temporal Mosaic, pruning >90% of redundant calculations to preserve edge compute headroom.

• Resolving 2D Overlap

  8-frame interaction crops are synthesized into a spatiotemporal mosaic, eliminating false contact triggers caused by vertical depth overlaps in top-down views.

• Multi-Frame Sex Voting

  Claw morphology is sampled across the first 10 track frames with multi-frame voting, achieving 82.0% Top-1 accuracy to automate tank sex-ratio audits.

• Edge Standalone & IP67

  On-camera NPU/TPU execution with single-cable PoE support; preserves local recording and event buffering even when external networks drop.

Live Device Demonstration

Top-down video telemetry: each crayfish in the frame receives a ByteTrack ID and sex classification tag (e.g. #3 male 91%). When individuals #2 and #5 initiate aggressive engagement, the contact zone highlights in flashing red with a FIGHT (temporal confirmed) overlay.

---

Hardware Sizing

Only an all-in-one IP67 smart camera is mounted over the tanks and connected to the aquaculture LAN:

	Hardware Model	Specs & Compute	When to Choose
	reCamera Pro (2GB) 4K Sensor + 3 TOPS NPU	Rockchip RV1126B · 3 TOPS NPU 2GB RAM · Linux 5.10	Recommended Tier. Broad RAS tanks and breeding pools requiring full 4-stage funnel and temporal mosaic classification
	reCamera 2002 (8GB) 1/2.8" Sensor + 1 TOPS TPU	SOPHGO SG2002 · 1 TOPS TPU 512MB RAM · Native C++	Cost-Optimized Tier. Dense small-tank rollouts focusing on detection and initial proximity triage

---

Deployment and Setup

1. Mounting and Optical Placement

• Mounting Height & Angle: Mount 0.8 ~ 1.5 m directly above the tank at a vertical top-down angle (80°~90°);
• Water Depth & Clarity: Suitable for depths 20~80 cm with visibility ≥ 25 cm; avoid placing aerators directly under the camera lens;
• Diffused Lighting: Uniform LED strip lighting (≥ 100 Lux) is recommended to prevent surface ripples from generating specular reflections.

2. Software Installation: Recommended Official Toolchain

This solution is integrated with the official Seeed automation framework. The primary recommended method uses solution-cli (or AI Solution Assistant) for automated precheck and deployment, with GUI options also supported:

View Field Deployment Guide 🖱️

Recommended: Automated Deployment with solution-cli (CLI Best Practice)

solution-cli is Seeed's unified deployment and sizing toolkit for industrial edge AI solutions (see Deployment Guide). Deploy with three steps without manually managing containers or tarballs:

# 1. Install or run solution-cli (skip if already available)
npm install -g @seeed-studio/solution-cli
# Or run on the fly via npx: npx @seeed-studio/solution-cli doctor

# 2. Diagnose local environment and scan network for reCamera hardware
solution-cli doctor
solution-cli device scan

# 3. Query solution details and deploy to target camera
solution-cli hub get crayfish_behavior_recognition --locale en
solution-cli deploy --solution crayfish_behavior_recognition --target recamera --host 192.168.42.1

Tip: If you are using AI agents like Claude Code, Codex, or Antigravity, prompt directly: "Please deploy reference design crayfish_behavior_recognition to the reCamera mounted over tank RAS-01 using solution-cli". The agent invokes the built-in skill to run pre-flight checks and mount the pipeline automatically.

Alternative: GUI Installation (SenseCraft Web / App Center)

1. Open http://<camera_ip> in a browser;
2. Navigate to App Center, search for crayfish_behavior_recognition;
3. Click Install; the camera automatically configures detection and temporal pipelines;
4. In settings, configure the farm MQTT Broker address (e.g. 192.168.1.100:1883) and enable the daemon.

---

Algorithm Architecture: 4-Stage Funnel & Temporal Mosaic

Camera Sensor (1080p @ 10-15fps)
   │
   ▼ [Stage 1] Object Detection     YOLO11n (640×640) → Bounding Boxes + ByteTrack
   │
   ▼ [Stage 2] Proximity Trigger    Center distance < 1.1× body length or IoU > 0 for ≥ 3 frames;
   │                                Crop interaction zone with 20% margin
   │
   ▼ [Stage 3] Single-Frame Triage  YOLO11n-cls (224×224) binary filter weeds out normal swimming
   │
   ▼ [Stage 4] Temporal Confirmation YOLO11n-cls (448×448 mosaic) validates 8-frame spatiotemporal matrix;
   │                                Completely eliminates 2D vertical overlap false alarms
   │
   ▼ Event Exit                     Publishes to MQTT/WS only on state transition (start / change / end)

Parallel Branch: Multi-Frame Sex Classifier

Within the first 10 frames of a new track ID, crops of the claw and abdomen are classified with YOLO11n-cls (128×128) via multi-frame voting, outputting male / female / unknown with 82.0% Top-1 accuracy.

---

Benchmarks and Field Telemetry

Evaluated on over 2,000 annotated field video frames across aquaculture research facilities:

1. Model Precision & Recall

Model	Resolution	Task	Benchmark Metric	Runtime Target
det_v2_n	640×640×3	Crayfish Detection	mAP50: 0.950 (P: 0.914, R: 0.880)	RV1126B RKNN / SG2002 TPU
behavior_temporal	448×448×3	8-frame Fight Verification	Recall: 0.950 (Precision: 0.905)	RV1126B RKNN INT8
behavior_cls	224×224×3	Single-frame Filter	Recall: 0.775, Precision: 0.816	RV1126B RKNN / SG2002 TPU
sex_cls	128×128×3	Sex Classification	Top-1: 0.820 (10-frame voting)	RV1126B RKNN / SG2002 TPU

2. Edge Hardware Execution

• reCamera Pro: 10~15 FPS (Full 4-stage closed loop, ~18% CPU);
• reCamera 2002: 12~15 FPS (Detection + proximity alert, ~25% CPU).

---

Telemetry Interfaces and Protocols

1. Farm Monitoring Dashboard (MQTT)

• Broker: mqtt://<camera_ip>:1883
• Topic: crayfish/tank/<tank_id>/event
• Aggression Alert Payload:

{
  "timestamp": 1793339665200,
  "tank_id": "RAS_04",
  "event_type": "FIGHT_CONFIRMED",
  "severity": "CRITICAL",
  "actors": [
    {"track_id": 2, "sex": "male", "box": [210, 140, 320, 260]},
    {"track_id": 5, "sex": "unknown", "box": [260, 160, 370, 280]}
  ],
  "duration_sec": 4.5
}

2. Periodic Tank Inventory & Sex Ratio Audits

• Topic: crayfish/tank/<tank_id>/stats
• Publishes hourly inventory counts, sex ratios, and aggression frequencies to assist automated feeding schedules.