Automated monitoring of health and welfare in groups of pigs using evidential reasoning and video-analytics

此次留学的总体目标是利用人工智能(AI)构建生猪健康预警系统。为此，我们将使用人工智能来分析猪圈的视频流。人工智能系统将监测围栏内猪之间的社会互动以及它们的个人行为。人工智能系统将随着时间的推移建立一张围栏内群体行为的图片，并识别与正常模式的偏差。我们想要探索的假设是，群体行为的细微变化可以预示未来的健康和福利问题。我们的目标是开发一种预警系统来提醒农民，以便他们能够在出现健康和福利问题的第一个迹象时迅速进行干预，以防止问题变得更加严重。在商业规模上，人类观察群体动态中可能表明早期疾病或异常行为的细微变化是不切实际的，因为这种观察是耗时的，而且现代养猪系统同时处理数百或数千个个体。相反，我们将使用人工智能来解决这个问题。一个摄像机将监控围栏，我们的人工智能计算机视觉软件将跟踪每只猪的动作，识别吃喝等活动，并识别猪之间的互动。为了将所有这些信息结合在一起，我们将使用证据推理网络(ERN)来完整地描述群体级别的行为。这些网络提供了一种原则性的方式，将不确定证据的多个来源，如多头猪的活动，结合在一起，给出了一幅连贯的世界图景。通过结合与多个个体猪的行为相关的证据，以及它们之间的相互作用，我们的目标是在群体层面上实现对行为模式的理解。检测群体行为的细微变化将成为我们的早期预警系统产生警报的基础。以下活动将由学生进行：1)建立和扩展现有的计算机视觉工具，用于监控动物的行为。其目标将是开发一种能够长期运行的实时跟踪和活动识别系统。该系统将产生每头猪的位置及其活动的长期记录，如社交互动、游戏以及喂养和饮水事件。目前的方法只能应用于有限的时间段，因为它们需要昂贵的离线非实时处理2)给定长期猪行为信息的数据集，我们将应用ERN来结合跟踪器提供的轨迹信息，以及基于深度学习的事件和活动检测方法，如动物姿势、摄食和饮水行为以及社会互动。我们将设计证据网络来推理和组合多个低水平证据来源，以触发与群体级别行为变化相关的警报。3)为了检测群体级别的交互/事件，同时减少对手动标记数据集的依赖，我们将探索自动挖掘长期行为数据集的方法。为了做到这一点，我们将开发一个正常猪行为的统计模型，并开发方法来检测当前行为是否偏离正常。然后，群体行为的变化会触发警报，发出福利事件的警告。4)我们最终将把我们的所有组件结合在一起，开发一个完整的系统来理解猪的群体行为。该系统将能够持续实时监测动物福利和健康，以便触发警报，并在需要时进行早期干预。除了明显的健康和福利益处外，该系统还有可能开辟一个新的研究领域，使数据挖掘能够以前所未有的方式应用于新获得的大型动物行为数据集。

The overall aim of this studentship is to use artificial intelligence (AI) to build an early warning system for pig health. To do this we will use AI to analyse a video stream of a pig pen. The AI system will monitor the social interactions between pigs within a pen as well as their individual behaviours. The AI system will build up a picture of the group behaviour within the pen over time and identify deviations from the normal pattern. The hypothesis we want to explore is that subtle changes in group behaviour can indicate future health and welfare problems. We aim to develop an early warning system to alert farmers so that they can quickly intervene at the first sign of health and welfare problems to prevent them from becoming more serious.On a commercial scale, human observation of subtle changes in group dynamics that may indicate early-stage disease or aberrant behaviour is impractical as such observations are time-consuming and modern pig systems deal with hundreds or thousands of individuals simultaneously. Instead, we will use AI to address this problem. A video camera will monitor the pen, and our AI computer-vision software will track the movements of each pig, recognise activities such as eating and drinking and recognise interactions between pigs. To combine all this information to give a complete picture of the group-level behaviour we will use Evidential Reasoning Networks (ERN). These networks provide a principled way to combine multiple sources of uncertain evidence, such as the activities of multiple individual pigs, to give a coherent picture of the world. By combining evidence related to the behaviour of multiple individual pigs, together with their interactions, we aim to achieve an understanding of the pattern of behaviours at the group level. Detecting subtle changes at the group level behaviour will form the basis for the alerts generated by our early warning system.The following activities will be performed by the student:1) Build on and extend existing computer-vison tools for monitoring animal behaviour. The aim will be to develop a real-time tracker and activity recognition system capable of long-term operation. The system will produce a long-term record of the location of each pig and its activities such as social interactions, play, and feeding and drinking events. Current approaches can only be applied for limited time periods as they require costly offline non-real-time processing.2) Given a dataset of long-term pig behaviour information, we will apply ERN to combine trajectory information, provided by the tracker, together with deep-learning-based event and activity detection methods such as animal posture, feeding and drinking behaviours and social interactions. We will design evidential networks to reason about and combine multiple sources of low-level evidence in order to trigger alerts related to group-level behavioural changes.3) To detect group-level interactions/events, while reducing reliance on manually labelled datasets, we will explore ways to automatically data-mine long-term behavioural datasets. To do this we will develop a statistical model for normal pig behaviour and develop methods to detect whether the current behaviour is deviating from normal. Changes in group behaviour will then trigger alerts giving warnings of welfare events.4) We will finally combine all our components together to develop a complete system for understanding the group behaviour of pigs. The system will be capable of continuously monitoring animal welfare and health in real time in order to trigger alerts and enable early intervention if needed. In addition to the obvious health and welfare benefits, this system will have the potential to open a new field of research, enabling data mining to be applied to newly available large datasets of animal behaviour in ways that have not been possible before.