Answers are grounded in this paper and cite their sources.
Animal Activity Recognition (AAR) is a key component of Precision Livestock Farming (PLF), enabling continuous monitoring of animal behaviour, health, and welfare. Advances in machine learning (ML) and sensor technologies have significantly improved AAR accuracy; however, most systems depend on cloud-based architectures, which are impractical in rural settings due to limited connectivity and latency constraints. Edge Artificial Intelligence (Edge-AI) offers a promising alternative by enabling local on-device inference that improves responsiveness, reliability, and autonomy. This systematic review analyses 118 peer-reviewed studies published between 2018 and 2025, examining four critical components of the AAR pipeline: communication technologies, data acquisition methodologies, ML and deep learning (DL) model development, and Edge-AI implementation. We summarise approaches to data collection across livestock species, sensor placements, sampling frequencies, labelling strategies, environments, total animals, and total samples. Furthermore, we categorise and evaluate the ML algorithms used in AAR, discussing feature engineering, windowing strategies, and validation techniques. Our findings reveal that only a limited number of studies have explored Edge-AI in real-world deployments, underscoring challenges related to model compression, resource-constrained inference, and energy efficiency. To address these gaps, we synthesise deployment strategies that include TinyML frameworks and hardware-aware model optimisation. Compared with previous surveys, this review uniquely integrates the entire AAR development cycle, from data collection and model training to real-world deployment, providing a comprehensive reference for developing scalable, on-device livestock monitoring systems.
No takes yet. Share an insight, caveat, or question.
Patrick et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: