A twin-level feature synthesis and long-term coherence framework for multi-object animal tracking in outdoor farm environments

Tracking animal activity in outdoor farm environments is crucial for livestock management, yet it remains a challenging task due to dynamic changes in cow appearance, variable lighting and unpredictable animal movements. Traditional vision-based systems, while effective indoors, often fail outdoors as they rely on consistent visual cues, leading to unstable tracking and poor identity association. This paper introduces TwinTrack, a multi-object tracking framework designed to address these challenges. The proposed framework leverages a Twin-Level Contextual Feature Synthesizer (TLCFS) to extract both fine-grained visual details and high-level semantic features, ensuring robustness under diverse environmental conditions. Additionally, a Dynamic Long-Term Temporal Consistency Module (DLTC) improves tracking stability by mitigating the effects of dynamic behaviors and scene fluctuations. The application of TwinTrack to outdoor farm environments demonstrates its ability to monitor livestock effectively, with experimental results showing stable, long-term tracking performance.<p></p>