TL;DR
This paper presents a novel 3D point cloud-based system for non-contact pig weight estimation, addressing the inefficiencies and stress associated with traditional manual weighing. By combining advanced point cloud denoising, feature extraction from the pig's back, and a convolutional neural network (CNN), the method achieves accurate weight estimation with simplified equipment and reduced processing complexity.
What Problem Does This Paper Address?
Accurate and timely monitoring of pig weight is crucial for effective farm management, impacting decisions on feeding, health, and market readiness. Traditional manual weighing methods are labor-intensive, time-consuming, and can induce stress in pigs, potentially affecting their well-being and growth. Existing non-contact methods, particularly those relying on 2D imaging, often struggle with accurately capturing body volume and are susceptible to variations in lighting and camera parameters. While 3D imaging offers more potential, existing approaches can involve complex hardware setups and computationally intensive processing, making them less practical for large-scale farms. This research aims to bridge this gap by developing an accurate, efficient, and cost-effective 3D system for pig weight estimation.
Methods
The proposed method employs a multi-stage approach:
- 3D Point Cloud Denoising: To handle noise inherent in 3D sensor data (e.g., from occlusions or equipment limitations), the researchers integrate two denoising techniques:
- Statistical Filtering: Used to remove large-scale outlier points by analyzing the distance of points to their neighbors.
- DBSCAN Clustering: A density-based algorithm employed to remove smaller, discrete noise points. This is a key contribution, offering robustness in segmenting pig point clouds from background noise.
- Pig Back Segmentation and Feature Extraction:
- Plane Projection: The point cloud data is projected onto a plane to focus on the pig's back.
- Convex Hull Detection: This technique is used to refine the dataset by removing extraneous head and tail point cloud data, ensuring that only the relevant back features are considered.
- Voxel Down-sampling: To improve real-time efficiency and reduce computational load, the point cloud data is down-sampled using a voxel grid approach, preserving essential features.
- Back Feature Extraction: Six key parameters are extracted from the pig's back: envelope volume, envelope area, projection area, shoulder width, belly width, and hip width. These features are chosen for their correlation with body weight.
- Weight Estimation using CNN: The extracted back features are fed into a Convolutional Neural Network (CNN). The CNN is designed to learn complex relationships between these features and the actual pig weight, enabling accurate estimation. The paper details the standard CNN architecture, including convolutional, down-sampling, fully connected, and output layers.
Datasets
The study utilized data collected from a real production environment at Wens Pig Farm in China.
- Experimental Subjects: 198 pigs (Large White × Landrace breed) with body weights ranging from 190 kg to 300 kg.
- Data Acquisition: A 3D depth Intel RealSense D435i camera was used to capture color and depth images. Point cloud data was acquired using ROS parsing.
- Dataset Size: A total of 10,000 sets of pig 3D point cloud data were acquired. For training and validation, 140 pigs were used as training samples, and the remaining 58 pigs served as test samples.
Key Insights and Ideas
- Hybrid Denoising for Robust Segmentation: The combination of statistical filtering and DBSCAN proves effective in separating pig point clouds from complex backgrounds, a common challenge in real-world environments.
- Focus on the Pig's Back: By specifically extracting features from the pig's back using convex hull detection and plane projection, the method isolates crucial morphological indicators of weight, reducing the impact of variable head and tail poses.
- Voxel Down-sampling for Efficiency: This technique is crucial for making 3D point cloud processing practical for real-time applications, significantly reducing data volume without sacrificing critical feature information.
- CNN for Complex Feature Learning: The use of a CNN allows for the sophisticated learning of non-linear relationships between extracted body dimensions and weight, outperforming simpler regression models.
Key Results
The proposed model achieved promising results:
- Mean Absolute Error (MAE): 12.45 kg
- Mean Absolute Percentage Error (MAPE): 5.36%
- Root Mean Square Error (RMSE): 12.91 kg
These metrics indicate a high degree of accuracy, meeting production demands. The paper also highlights that the relative estimation error for the majority of pigs fell within 6%. The CNN-based model demonstrated superior performance compared to Radial Basis Function (RBF) Neural Networks, with lower MAE, MAPE, and RMSE values.
Limitations and Future Work
- Pig Posture Influence: The authors acknowledge that pig postures can influence the reliability and accuracy of back feature extraction, leading to deviations between actual and extracted parameters.
- Future Work: They propose developing different models to estimate pig weight under various postures and considering feature differences across these postures.
- Data Noise and Incompleteness: Uncontrolled pig movements can lead to noisy point cloud data, including non-uniform sampling, voids, and omissions.
- Future Work: The researchers plan to integrate multiple consecutive frames of point cloud data to densify and fill in sparse or missing portions of the point cloud.
- Generalization: Further research is needed to broaden the scope of target subjects and enhance the weight estimation model's generalization capabilities.
- Integrated System: The authors envision using the weight estimation output to assess growth status, provide tailored feeding recommendations, and generate alert notifications, suggesting the development of a comprehensive weight management system.
Summary updated