Automatic cattle identification system based on color point cloud using hybrid PointNet++ Siamese network.

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Manual cattle health checks are labor-intensive.
• Existing 2D vision systems struggle with environmental variations and require retraining for new cattle.
• Accurate cattle identification is crucial for effective health monitoring systems.

Purpose of the Study:

• To develop a novel, adaptable cattle identification system using color point clouds.
• To overcome limitations of existing 2D vision-based methods.
• To enable accurate identification of individual cattle without model retraining.

Main Methods:

• Utilized RGB-D cameras for capturing color point clouds.
• Implemented a hybrid detection method combining 2D depth image analysis and point cloud conversion.
• Employed a lightweight tracking approach with IoU-based matching and mask size analysis.
• Developed a PointNet++ Siamese Network with triplet loss for feature extraction and identification.

Main Results:

• Achieved an average identification accuracy of 99.55% over 13 days.
• Successfully identified individual cattle, including unknown individuals, without retraining the model.
• Demonstrated robust feature extraction from color point clouds.

Conclusions:

• The proposed color point cloud-based system offers a highly accurate and adaptable solution for cattle identification.
• This technology can be integrated into comprehensive cattle health monitoring systems.
• Eliminates the need for model retraining, improving efficiency in dynamic farm environments.