Automatic Individual Pig Detection and Tracking in Pig Farms

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Individual pig detection and tracking are crucial for video-based farm monitoring.
• Challenges include light variations, similar pig appearances, shape changes, and occlusions.

Purpose of the Study:

• To develop a robust, on-line method for detecting and tracking multiple pigs without manual identification.
• To ensure the method works under both daylight and infrared (nighttime) conditions.

Main Methods:

• Coupled a Convolutional Neural Network (CNN)-based detector with a correlation filter-based tracker.
• Utilized a novel hierarchical data association algorithm for robust tracking.
• Employed a one-stage prediction network with multi-scale features for optimal accuracy/speed trade-off.
• Defined a 'tag-box' for each pig and used key-point tracking with learned correlation filters.

Main Results:

• The method robustly detects and tracks multiple pigs in complex farm environments.
• Successfully addressed challenges like light fluctuation, similar appearances, shape deformations, and occlusions.
• Demonstrated effective correction of drifted tracks and integration of tracking fragments.

Conclusions:

• The proposed method offers a feasible solution for long-term individual pig tracking in complex settings.
• Shows significant commercial potential for advanced livestock monitoring systems.