Predicting livestock behaviour using accelerometers: A systematic review of processing techniques for ruminant behaviour prediction from raw accelerometer data

Precision Technologies are emerging in the context of livestock farming to improve management practices and the health and welfare of livestock through monitoring individual animal behaviour. Continuously collecting information about livestock behaviour is a promising way to address several of these target areas. Wearable accelerometer sensors are currently the most promising system to capture livestock behaviour. Accelerometer data should be analysed properly to obtain reliable information on livestock behaviour. Many studies are emerging on this subject, but none to date has highlighted which techniques to recommend or avoid. In this paper, we systematically review the literature on the prediction of livestock behaviour from raw accelerometer data, with a specific focus on livestock ruminants. Our review is based on 66 surveyed articles, providing reliable evidence of a 3-step methodology common to all studies, namely (1) Data Collection, (2) Data Pre-Processing and (3) Model Development, with different techniques used at each of the 3 steps. The aim of this review is thus to (i) summarise the predictive performance of models and point out the main limitations of the 3-step methodology, (ii) make recommendations on a methodological blueprint for future studies and (iii) propose lines to explore in order to address the limitations outlined. This review shows that the 3-step methodology ensures that several major ruminant behaviours can be reliably predicted, such as grazing/eating, ruminating, moving, lying or standing. However, the areas faces two main limitations: (i) Most models are less accurate on rarely observed or transitional behaviours, behaviours may be important for assessing health, welfare and environmental issues and (ii) many models exhibit poor generalisation, that can compromise their commercial use. To overcome these limitations we recommend maximising variability in the data collected, selecting pre-processing methods that are appropriate to target behaviours being studied, and using classifiers that avoid over-fitting to improve generalisability. This review presents the current situation involving the use of sensors as valuable tools in the field of behaviour recording and contributes to the improvement of existing tools for automatically monitoring ruminant behaviour in order to address some of the issues faced by livestock farming.