Recognizing piglet play, in all its forms, using machine learning on data from accelerometers and computer vision

动物行为学家对幼年农场动物的玩耍很感兴趣，他们关心行为是如何进化和发展的，可以用来评估和改善动物的福利。然而，研究和实际应用需要有效和自动地测量这种行为。加速计非常适合用来检测以用力和突然移动为特征的比赛。计算机视觉是另一个很有前途的工具，它可以跟踪视频中的动物，并可以识别特定的行为。我们的目标是开发和比较加速计和基于视频的工具，使用机器学习训练来自动检测仔猪的不同类型的游戏，并以与人类观察者相同的方式对它们进行分类。我们的第一个目标是设计定制的加速计，用于佩戴在仔猪的腿和耳朵上。我们的第二个目标是训练和验证机器学习算法来检测游戏。我们将在一个游戏场地收集成对的小猪的视频和加速度计数据。计算机视觉算法将确定动物的位置和运动。数据将与人类观察者进行的行为分类同步，并训练机器学习模型，以使用来自耳朵加速计、腿部加速计或计算机视觉跟踪动物的数据对游戏类型(运动、物体、社交)和其他行为(例如攻击性)进行分类。我们的第三个目标是测试这些方法在魁北克和马尼托巴省典型农场条件下的表现。他们应该正确地识别个体之间和仔猪群体之间的差异，并表现得像人类观察者一样好，在很短的时间内以相同的方式对行为进行分类。这项研究将产生有价值的科学和社会影响。自动化工具将极大地扩展我们研究游戏的性质及其对动物福利的意义的能力。开发的工具将公开传播，容易被其他研究人员采用。我们将通过直接向猪肉行业展示结果来鼓励实际应用。该项目将使研究一种令人着迷的动物行为成为可能，并在日益可持续的食品生产系统中应用于改善农场动物的生活。

Play in young farm animals is of interest to animal behaviour scientists concerned with how behaviour evolves and develops, and can be used to assess and improve animal welfare. However, research and practical applications require that this behaviour be measured efficiently and automatically. Accelerometers are ideally suited to detect play, which is characterized by forceful and sudden movements. Computer vision is another promising tool that enables tracking of animals in video, and can recognize certain behaviours. We aim to develop and compare accelerometer and video-based tools, trained using machine learning to automatically detect different types of play in piglets, classifying them in the same way as human observers.Our first objective is to design custom accelerometers to be worn on piglet legs and ears. Our second objective is to train and validate machine learning algorithms to detect play. We will collect video and accelerometer data from pairs of piglets in a play arena. Computer vision algorithms will determine animal positions and movements. Data will be synchronized to behavioural categorizations made by human observers, and machine learning models trained to classify types of play (locomotor, object, social) and other behaviours (e.g. aggression) using data from ear accelerometers, leg accelerometers, or animal tracking by computer vision. Our third objective is to test how these methods perform in typical farm conditions, in Québec and Manitoba. They should correctly identify differences between individuals and between groups of piglets, and perform as well as human observers, categorizing behaviour the same way in a fraction of the time.This research will have valuable scientific and societal impacts. Automated tools will drastically expand our capacity to study the nature of play and its significance for animal welfare. Tools developed will be openly disseminated, readily adaptable by other researchers. We will encourage practical applications by presenting results directly to the pork industry. The project will enable research on a fascinating type of animal behaviour, and applications to improve the lives of farm animals in increasingly sustainable food production systems.