부산대학교 도서관

Classifying calves' behavior from accelerometer sen-sors attached to neck collars is a promising way to assess calf welfare in commercial farms. Developing global models that achieve high performance across multiple calves is challenging due to inter-individual diversity, which limits application on commercial farms. It can be assumed that there is a greater or lesser degree of similarity between calves linked to their age, weight, breed, mobility, health status, etc. Therefore, developing global models that use calf similarity in the classification seems promising. This paper aims to develop Personalized Weighted AdaBoost (PWA) models where the weights in the training set are tuned according to the similarity with the query calves. The approach adopted consists of (i) measuring the similarity between each pair of calves by testing three different strategies (label-based, pairwise-based, and cluster-based similarity) and (ii) tuning the weights so that the training calves with the highest similarity to the query calf get higher weights. This approach was implemented with an accelerometer dataset collected from 21 pre-weaned calves. The accelerometer data was manually annotated using videos to obtain the behaviors (lying, walking, eating, running, and drinking milk) used for this work. Performance was compared against a global AdaBoost model without the assignment of weights (baseline). PWA models improve the performance compared to the baseline, regardless of the strategy used to compute the similarity metric (label-based: 0.64 +/-0.132, pairwise: 0.59 +/- 0.084, cluster-based: 0.62 +/- 0.145; baseline: 0.56 +/- 0.138 [average balanced accuracy over 10 query calves +/- standard deviation]). Although this approach still suffers from a lack of data and an unbalanced dataset, it seems extremely promising for developing robust calf behavior classification models from accelerometer data, thus offering new opportunities to improve calf welfare in commercial farms.