基于计算机视觉的群养猪身份和进食行为识别方法研究

In order to prevent the epidemic and spread of swine disease，Timely Detection of pig abnormal behavior is very important. The group-housed pigs will be taken as study objects in this project. Considering the variability of pig position in pigpen, the invariance principle of translation, scale and rotation as well as visual saliency strategy will be studied deeply to select feature adaptively. We focus on the methods of pig’s shape, color and texture feature extraction, data fusion and feature optimization under the deformed state of pigs, then build a feature vector space and the recognition model for pig’s identity based on computer vision. We will analyze the spatial and temporal relationship between pig’s body position and the specific markers, and study the movement of ‘feature points’ on contour profile of a pig and its frequency characteristics. Hidden Markov Model，neural network Deep Learning and Dynamic data acquisition and analysis will be used to build the recognition model in drinking water and feeding behavior of pigs and estimate drinking water consumption and feed intake of pigs based on computer vision. The new recognition method of drinking water and feeding behavior of pigs will be proposed based on computer vision. The database in drinking water and feed intake of pigs will be set up. Once the abnormal behavior is detected, the keeper will be notified to deal with it. The continuous non-intrusive behavior recognition and health monitoring of pigs will be implemented to improve the level of intelligence and precision of pig industry. This research can provide a theoretical basis for video recognition of animal identity and behavior, and provide a new method for developing pigs' behavior and health monitoring equipment with independent intellectual property rights.

为了防止猪病的流行与蔓延，及时发现猪的异常行为非常重要。本项目以群养猪为研究对象，考虑猪在猪舍中位置的可变性，研究具有平移、尺度、旋转不变性原理和视觉显著性策略自适应筛选特征，提出猪在形变状态下的外形、颜色和纹理等特征的提取、融合和优化方法，构建基于计算机视觉的猪的身份特征向量和识别模型；分析猪的相应体位与特定标志物之间的空间和时间关系，研究猪体外形轮廓上特征点的运动轨迹和频域特性，采用隐马尔科夫理论、神经网络深度学习和动态数据采集与分析等方法构建猪的饮水和采食行为识别模型以及饮水量和采食量估算模型，提出基于计算机视觉的猪的饮水和采食行为识别和定量分析新方法；建立猪的饮水和采食量档案，发现异常及时处理，实现连续无侵扰的猪的行为识别与健康状态监控，提高养猪业的智能化精细化水平。本研究可为动物身份和行为的视频识别提供理论基础，为研发具有自主知识产权的猪的行为与健康监控装备提供新方法。

生猪个体的饮水、采食异常直接影响到猪的健康状态。如果能及时发现这些异常行为，切断病猪传染源，就可将损失降到最低程度。本项目以群养猪为研究对象，对猪个体的身份和饮水及采食行为进行识别研究。在视频图像采集系统的构建中，分别从保证研究精度及质量和探索使用低成本设备便于推广应用的角度出发，分别研究构建了高、中低档次的三种视频图像采集系统。在群养猪的身份识别方面，针对猪的非刚体特点，研究猪的形状、颜色和纹理等多维特征提取方法，提出了基于多尺度局部差分方向数模式的猪身份识别、基于韦伯纹理局部描述子的猪身份识别和基于改进的光流法和深度学习的猪的身份识别等方法，识别准确率可达95.71%。在群养猪的饮水行为识别方面，通过计算猪只头部与饮水区域的交集面积，获取空间特征，根据连续帧图像计算猪只的运动幅度，获取饮水行为时间特征，研究并提出了一种基于空间特征与时间特征相结合的饮水行为识别方法，识别率约为95%。为了进一步提高识别精度，研究并提出一种基于ResNet50和LSTM的猪饮水和玩耍饮水器分类算法。此算法的分类精度约为92.5%。此研究为探索破解难以区分猪饮水和玩耍饮水器的难题提供了新思路。在群养猪的采食行为识别方面，采用改进的粒子群-支持向量机法（IPSO-SVM）提取关键帧，利用光流法提取各方向光流箭头数目以及幅值等光流特征，再利用支持向量机分类器对猪的采食行为进行识别，准确率达93.33%。为了实现猪个体的采食行为识别和进食时间计量，采用CNN架构XceptionV3和LSTM网络提取视频片段中进食区域的时空特征，提出了一种基于Xception和LSTM的猪进食行为识别算法。所提出算法能够以98.4%的正确率识别猪进食行为，且能够识别98.5%的猪个体进食时间。该项研究对提高我国规模养猪场的健康养殖及智能化水平具有重要的学术意义和应用价值。