农业是我国国民经济的基础，农业自动化程度直接影响着生产效率和农业产量。因此，为提高农业生产率，实现生产与管理的自动化和智能化，机器视觉作为一种高效手段已逐渐应用于各项农业生产中，其不疲劳性也正逐渐显现着优势。与种植业并列为农业支柱的养殖业，正在从传统的粗放型向现代化集约型转变。特别是在某些大型规模化养殖场中，很多环节已实现自动化，如蛋的收集，饲料供给和环境监控等。但仍有一项劳动密集型工作尚未实现自动化，即病死禽的及时探测和清除。不及时的清除可能会造成禽群疾病大面积感染，造成重大经济损失。本文的预期采用视觉方式监测禽类的生长健康状况，实现禽群健康自动监测和预警。 本文在禽舍环境控制器项目基础上，设计一种辅助控制系统的禽只健康图像检测系统，并简单的介绍了图像检测硬件平台。通过摄像机固定位置采集现场画面传输给上位机，上位机将视频进行算法处理将某区域禽群健康状况最终显示在操作中心的界面上，工作人员可减少往返控制中心和养殖场的频率，免于长期处于恶劣环境中逐一排查。 基于以上背景，本文在实验室环境下模拟小型养殖场研究一种禽类健康状况的图像检测方法，为后期本检测系统投入使用提供理论依据。主要研究工作如下：首先对活动性作为判断是否健康的依据进行观察实验并得出肯定结论。视觉方式检测首先要根据采集的视频图像选取适宜特征进行目标提取，采用目标追踪的方法得到目标运动轨迹，对轨迹结果处理分析并对健康与非健康进行分类。运动目标检测采用高斯混合背景建模方法，根据颜色阈值对目标进一步提取，形态学方法进行噪声去除；目标跟踪采用对目标区域进行角点检测，并进行角点光流跟踪。跟踪实验证明单目标及多目标分离运动的情况，跟踪效果良好；而多目标相互遮挡较多时会跟踪出错，后期需要进行改进。对跟踪结果，提出宽比、高比、面积比三种影响特征的概念作为BP网络分类决策输入特征，经过试验表明，在初始值等参数、网络结构确定情况下，检测正确率能达到80%以上。

Agriculture is the foundation of the national economy in our country, whose automation degree directly affects the productivity and production. Therefore, to improve the agricultural productivity, realize the automation and intelligentization of production and management, machine vision, as a kind of efficient method, has been gradually applied in the agricultural production, and meanwhile its indefatigability shows gradually. Parallel to planting as the agricultural pillar, aquaculture is converting from traditional extensive to modern intensive. Especially in some large-scale farms, many production processes have realized automation, such as egg collection, feed, water supply and environmental monitoring, etc. But still there is a labor-intensive work that has not yet realized automation, that is the ill or dead birds timely detection and removal. If not, healthy chicks widespread infection will bring heavy losses to the whole farm. Visual monitoring the poultry health growth condition is expected to achieve in this paper. Based on the project of birdhouse environment controller, we design an image-assisted monitor system of birdhouse environment controller, and simple introduce the image detection hardware platform. Collect the scene video through the fixed camera and transmit to PC wirelessly; PC will process the collected video via algorithm to realize the poultry health condition displayed on the user interface finally, making the operation center intuitively catch the current regional healthy and unhealthy bird information, reducing frequency of staffs commute between control center and farms, and free from them long-time checking chicks one by one in bad environment. In view of the background above, we simulate a small farm in the laboratory to research poultry health image detection method, and provide theory foundation for the system later put into use. The main work is as follows: firstly make experimental observation and judgment to the evidence that activity condition differentiates healthy or not. Visual detection firstly should select appropriate extracted features according to video images, target tracking method is adopted to get the target trajectory, and processing with the target trajectory result to the classification of healthy and unhealthy. To moving target detection, Gaussian mixture background modeling method is used, the color threshold method for further target detection, mathematical morphology method for noise removal. To moving target tracking corner detection is used in target area, and then the detected corner points will be used to flow tracking. Tracking experiment proved that the tracking effect of single objective and separatist multi-objective movement is good; but shield multi-objectives may track inaccurately, and the algorithm needs improving later. In this paper, for tracking results processing, the conceptions of height ratio, width ratio, and area ratio are proposed to be the input features of BP Neural Network classification decision. The simulation shows that with the initial value, network structure parameters determined, the detection accuracy can reach more than 80%.