草原放牧绵羊牧食行为检测方法研究

Recent years, grassland resources were degenerated seriously due to man-made overgrazing in the Inner Mongolia prairie. In order to sustainable utilize grassland resources, ecology researchers have studied extensively how to affect grassland ecosystem for livestock grazing on the grassland, of which the research of animal grazing behavior in grazing grassland has become a research hotspot in ecology. However, the traditional methods of direct observation were utilized to realize monitoring of grazing behavior until now, which is time-consuming and laborious, maybe affects the normal animal feeding behavior. Therefore, it is very significance to study the system of automatic detect sheep grazing behavior in grazing grassland by means of information measure. Embedded microprocessor is designed so as to acquire the information which includes GPS, accelerometer and other environmental parameters sensor. After fusing audio and video signals which were from micro-camera, time and space data of sheep grazing is matched and related sheep grazing behavior is detected. In this paper, we obtain the Wiener entropy sound characteristic model base by means of developing the analysis of variance algorithm, the BP neural network classifier is used to achieve the identification goal. After preprocessing video signal, images are matched by using the sparse representation algorithm, feature extraction is realized by developing scale invariant feature transform algorithm, and parameters are optimized via deep learning model, and finally grazing sheep diet selection and other feeding behaviors are detected by researching the support vector machine classifier, then finishing the system of detection and recognition that aims at sheep grazing behavior in grazing grassland.

近年来内蒙古草原由于人为的过度放牧导致草地资源严重退化，为了可持续利用草地资源，生态学研究人员就放牧家畜对草地生态系统的影响进行了广泛的研究，其中放牧家畜的牧食行为研究成为生态学的研究热点。然而直到今日主要依靠传统的直接观测法来实现牧食行为的监测，费时费力也可能影响家畜的正常牧食行为。因此，利用信息手段研究一种自动检测草原放牧绵羊牧食行为的系统具有十分重要的意义。本课题选用嵌入式微处理器采集GPS、加速度传感器及其它环境参数传感器信息，融合微型摄像机音频、视频信号，实现放牧绵羊时空数据匹配及相关牧食行为的检测。本文通过ANOVA方差分析算法得出维纳熵声音特征模型库，经BP神经网络分类器实现识别目的。视频信号经预处理后采用稀疏表示算法配准图像、小波SIFT算子提取特征、深度学习模型优化参数，最后由支持向量机分类器实现放牧绵羊食性选择等采食行为检测，完成草原放牧绵羊牧食行为检测识别系统。

针对放牧绵羊牧食行为监测传统上主要依靠人工直接观测，费时费力、不准确，且影响家畜的正常牧食行为等问题，本文采用现代测试技术设计开发了放牧绵羊牧食行为自动检测系统，并建立绵羊牧食行为与草场质量的相关性，为生态工作者进行草场评价、草原管理等提供依据。.本课题选用嵌入式微处理器采集GPS、加速度传感器及其它环境参数传感器信息，融合微型摄像机音频、视频信号，实现放牧绵羊时空数据匹配及相关牧食行为的检测。采用K-mean聚类算法较准确地实现放牧绵羊的卧息行为识别，再匹配GPS速度信号得到放牧绵羊运动行为和站立行为的识别结果，结果表明放牧绵羊的三种运动行为识别率均达到85%以上。对放牧绵羊采食牧草的视频进行预处理，提取放牧绵羊采食到的牧草叶片图像，即ROI(Region of interest)。通过提取牧草的颜色、形状、纹理特征，进行有效的特征匹配与降维处理，由BP神经网络训练学习后完成牧草的分类，实现放牧绵羊采食植物种类的识别，与人工模拟法对比验证，整体识别率将近80%。.通过本项目的研究，为放牧行为检测、草原植被识别提供有价值的技术参考，为我国未来智慧畜牧业发展提供一定的技术支撑。