# Detection and Continuous Tracking of Breeding Pigs with Ear Tag Loss: A Dual-View Synergistic Method

## TL;DR

This paper introduces a dual-view system to detect and track breeding pigs that have lost their ear tags, improving farm management accuracy.

## Contribution

The novel contribution is a lightweight, dual-view synergistic framework for real-time detection and tracking of ear tag-lost pigs.

## Key findings

- The pruned detector achieved 94.03% mean average precision for bounding box detection and 90.16% for instance segmentation.
- The proposed tracker improved success rate, normalized precision, and precision by 4.39%, 3.22%, and 4.77% respectively.
- The framework is feasible for edge device deployment and shows promise for other livestock species.

## Abstract

Ear tags are widely used as the primary method for individual identification in breeding pigs. However, the loss of ear tags can disrupt production records, compromise health management, and create confusion in pedigree information, ultimately undermining genetic selection and precise management efforts. Simply detecting lost ear tags is not enough, as breeding pigs in large pens often look very similar, move frequently, and have ears that are frequently obscured from the staff’s sight, making it challenging for farm staff to promptly and accurately identify the affected animals. To address these challenges, we present an integrated detection–mapping–tracking framework that combines a localized top-down perspective with a panoramic oblique view. Our experimental results demonstrate that, while remaining lightweight, the proposed framework can automatically detect and continuously track breeding pigs that have lost their ear tags, even in complex farm environments. This enables staff to quickly identify loss events, accurately locate individual animals, and efficiently reapply new ear tags. As a non-invasive solution, our approach offers strong technical support for the automated monitoring and management of ear tag status in precision livestock farming and holds promise for application to cattle, sheep, and other livestock species.

The lossof ear tags in breeding pigs can lead to the loss or confusion of individual identity information. Timely and accurate detection, along with continuous tracking of breeding pigs that have lost their ear tags, is crucial for improving the precision of farm management. However, considering the real-time requirements for the detection of ear tag-lost breeding pigs, coupled with tracking challenges such as similar appearances, clustered occlusion, and rapid movements of breeding pigs, this paper proposed a dual-view synergistic method for detecting ear tag-lost breeding pigs and tracking individuals. First, a lightweight ear tag loss detector was developed by combining the Cascade-TagLossDetector with a channel pruning algorithm. Second, a synergistic architecture was designed that integrates a localized top-down view with a panoramic oblique view, where the detection results of ear tag-lost breeding pigs from the localized top-down view were mapped to the panoramic oblique view for precise localization. Finally, an enhanced tracker incorporating Motion Attention was proposed to continuously track the localized ear tag-lost breeding pigs. Experimental results indicated that, during the ear tag loss detection stage for breeding pigs, the pruned detector achieved a mean average precision of 94.03% for bounding box detection and 90.16% for instance segmentation, with a parameter count of 28.04 million and a detection speed of 37.71 fps. Compared to the unpruned model, the parameter count was reduced by 20.93 million, and the detection speed increased by 12.38 fps while maintaining detection accuracy. In the tracking stage, the success rate, normalized precision, and precision of the proposed tracker reached 86.91%, 92.68%, and 89.74%, respectively, representing improvements of 4.39, 3.22, and 4.77 percentage points, respectively, compared to the baseline model. These results validated the advantages of the proposed method in terms of detection timeliness, tracking continuity, and feasibility of deployment on edge devices, providing significant reference value for managing livestock identity in breeding farms.