# GCNTrack: A Pig-Tracking Method Based on Skeleton Feature Similarity

## TL;DR

GCNTrack is a new method for tracking pigs using skeleton features, improving accuracy even when pigs leave and re-enter camera view.

## Contribution

GCNTrack introduces a skeleton feature similarity-based tracking method with dual-tracking strategy for improved pig tracking in incomplete camera fields of view.

## Key findings

- GCNTrack achieved 84.98% MOTA and 82.22% IDF1 score for short-duration videos.
- Tracking precision was 74% for medium-duration videos with frequent pig entries.
- The method showed average 6.28 identity switches per pig in long-duration tracking.

## Abstract

Pig tracking has gradually become a requirement in the modern pork industry, facilitating the implementation of automated and intelligent management. Common cameras cannot monitor the entire pig housing in the production environment, which results in identity matching errors caused by pigs leaving and entering the camera’s detection area. This article proposed a pig-tracking method combined with re-identification. It provides technical support for automated and intelligent monitoring by enabling pig tracking in the modern pork industry.

Pig tracking contributes to the assessment of pig behaviour and health. However, pig tracking on real farms is very difficult. Owing to incomplete camera field of view (FOV), pigs frequently entering and exiting the camera FOV affect the tracking accuracy. To improve pig-tracking efficiency, we propose a pig-tracking method that is based on skeleton feature similarity, which we named GcnTrack. We used YOLOv7-Pose to extract pig skeleton key points and design a dual-tracking strategy. This strategy combines IOU matching and skeleton keypoint-based graph convolutional reidentification (Re-ID) algorithms to track pigs continuously, even when pigs return from outside the FOV. Three identical FOV sets of data that separately included long, medium, and short duration videos were used to test the model and verify its performance. The GcnTrack method achieved a Multiple Object Tracking Accuracy (MOTA) of 84.98% and an identification F1 Score (IDF1) of 82.22% for the first set of videos (short duration, 87 s to 220 s). The tracking precision was 74% for the second set of videos (medium duration, average 302 s). The pigs entered the scene 15.29 times on average, with an average of 6.28 identity switches (IDSs) per pig during the tracking experiments on the third batch set of videos (long duration, 14 min). In conclusion, our method contributes an accurate and reliable pig-tracking solution applied to scenarios with incomplete camera FOV.