Abstract Poultry behavior is an important indicator for diagnosing poultry diseases. Accurate pose estimation is the basis of poultry behavior analysis, and it provides poultry disease warning methods. On large-scale poultry farms, it is usually a farmer or veterinarian who watches the pose of the broiler chicken to determine whether they are sick. When the posture of the bird is abnormal, the breeders can address the problem promptly. Accurate tracking of birds can better estimate their posture. In this paper, pose estimation based on a deep neural network (DNN) is applied to analyze the broiler chicken’s behavior for the first time. First, the pose skeleton is constructed through the feature points of the broiler chicken, and then, it is used to track specific body parts. Furthermore, the naive bayesian model (NBM) was used to classify and identify the poses of broiler chickens. Preliminary tests revealed that we could identify chickens in standing, walking, running, eating, resting, and preening states by comparing the postures of classified broiler chickens. The test precision of behavior recognition is 0.7511 (standing), 0.5135 (walking), 0.6270 (running), 0.9361 (eating), 0.9623 (resting), and 0.9258 (preening). Our research provides a noninvasive method for broiler chicken behavior analysis, which can be used for future behavior analysis in broiler chicken farming.

中文翻译：

---

基于深度神经网络的肉鸡姿态估计与行为分类

摘要 家禽行为是诊断家禽疾病的重要指标。准确的姿态估计是家禽行为分析的基础，它提供了家禽疾病预警方法。在大型家禽养殖场，通常是养殖户或兽医通过观察肉鸡的姿势来判断它们是否生病。当鸟的姿势出现异常时，饲养员可以及时解决问题。准确跟踪鸟类可以更好地估计它们的姿势。本文首次应用基于深度神经网络（DNN）的姿态估计来分析肉鸡的行为。首先通过肉鸡的特征点构建姿势骨架，然后用于跟踪特定的身体部位。此外，朴素贝叶斯模型（NBM）用于分类和识别肉鸡的姿势。初步测试表明，通过比较分类肉鸡的姿势，我们可以识别鸡的站立、行走、跑步、进食、休息和梳理状态。行为识别的测试精度为0.7511（站立）、0.5135（行走）、0.6270（跑步）、0.9361（进食）、0.9623（休息）和0.9258（整理）。我们的研究为肉鸡行为分析​​提供了一种无创的方法，可用于未来肉鸡养殖的行为分析。行为识别的测试精度为0.7511（站立）、0.5135（行走）、0.6270（跑步）、0.9361（进食）、0.9623（休息）和0.9258（整理）。我们的研究为肉鸡行为分析​​提供了一种无创的方法，可用于未来肉鸡养殖的行为分析。行为识别的测试精度为0.7511（站立）、0.5135（行走）、0.6270（跑步）、0.9361（进食）、0.9623（休息）和0.9258（整理）。我们的研究为肉鸡行为分析​​提供了一种无创的方法，可用于未来肉鸡养殖的行为分析。