SOGAETING: AI-Based Cow Mounting Behavior Detection

YOLOv10-based intelligent vision system for detecting cow mounting behavior (estrus period) in real time, helping reduce livestock breeding losses and improving farm management efficiency.

---

1. Project Overview

This project proposes an AI-based solution to prevent economic losses caused by undetected mounting behavior in cows during their estrus period.
Since over 65% of mounting occurs at night, the system provides real-time detection and alert notifications to farmers.

• AI Model: YOLOv10 (parameter-tuned high-precision version)
• Objective: Real-time and accurate detection of mounting behavior
• Focus: Robust detection under low light, occlusion, and group interactions

---

2. Key Features

• Real-Time Detection
  Detects mounting behavior from farm CCTV streams in real time using YOLOv10.

• Automated Behavior Recognition
  Recognizes estrus-related mounting behaviors using deep learning–based object detection
  combined with multi-frame validation logic.

• Database Integration
  Stores both CCTV footage and detection results for record tracking and retraining.

---

3. Model Architecture & Optimization

YOLOv10 (Specially Tuned for Mounting Detection)

• Real-time detection optimized for farm CCTV conditions
• Parameter tuning targeted at cow mounting motion and frame continuity

Training Parameters:

Epochs	Batch	ImgSz	Mosaic	Mixup	Copy-Paste	HSV-H	HSV-V
50	32	640	0.5	0.1	0.5	0.05	0.6

Class Labeling:

• Mounting Behavior (Positive) vs Normal Behavior (Negative)
• Label annotations verified using recorded farm videos

---

3.1 Post-Processing Pipeline

After raw YOLOv10 detection, multi-stage filtering was applied to ensure temporal consistency:

Stage	Condition	Description
1. Confidence Filter	≥ 0.75	Only high-confidence detections kept
2. Height Difference (Δy)	≥ 10 px	Detect vertical pixel displacement indicating mounting
3. Temporal Consistency	> 5 consecutive frames	Ensures sustained action before alert
4. Frame Gap	≤ 30 frames	Groups nearby detections as one event

Filter Flow:

• Raw Detection → Confidence ≥ 0.75 → Height Δ ≥ 10px → Consecutive Frames > 5 → Frame Gap ≤ 30 → ✅ Mounting Confirmed

---

4. Dataset

• Company-Provided Data
  • Labeled mounting behavior dataset
  • Real-farm CCTV videos (day & night)

---

5. Experimental Process

5.1 YOLOv10 Parameter Tuning

Parameter	Value	Purpose
Epochs	50	Optimal convergence
Batch Size	32	Stable GPU memory usage
Image Size	640	Speed–accuracy balance
Mosaic / Mixup / Copy-Paste	0.5 / 0.1 / 0.5	Data diversity for cow herd motion
HSV-Hue / Value	0.05 / 0.6	Nighttime adaptation enhancement

Validation Control:

• Early stopping + overfitting monitoring
• CPU/GPU split validation for consistency

---

5.2 Additional Filtering Conditions

Condition	Logic
Confidence ≥ 0.75	Filter low-confidence detections
Height Δ ≥ 10 px	Capture significant upward motion
Frames > 5	Confirm sustained contact behavior
Frame gap ≤ 30	Treat detections within window as one event

---

6. Final Results

Performance Metrics (Test Dataset)

Metric	Result	Target	Status
Recall (Detection Rate)	93.7%	≥ 90%	✅ Achieved
FPR (False Positive Rate)	16.6%	≤ 20%	✅ Achieved

Model achieved 93.7% recall and reduced false alarms to 16.6% using post-validation filters.
Most missed events occurred when multiple cows overlapped under poor illumination.

---

7. Environment

• Framework: YOLOv10 (Ultralytics)
• Libraries: PyTorch, OpenCV, NumPy, Pandas
• Hardware: NVIDIA RTX GPU (CUDA 11.x)
• Development Platforms: VSCode, Docker

---

8. System Integration

Architecture Overview: CCTV → Custom-trained YOLOv10 → Post-filtering → Alert Module → Database Storage (MySQL) → Dashboard Display (Web)

• Inference: YOLOv10 best.pt (user-trained checkpoint)
• Filtering: Custom confidence/height/frame validation
• Storage: MySQL (timestamps + detection results)

9. Run detection on sample video