基于时空兴趣点的生猪行为识别

发布时间：2019-05-09 04:37

【摘要】：基于计算机视觉的动物行为参数分析具有重要研究价值和应用空间,其对传统的畜禽养殖方式产生深远的影响。如今大多数规模化养猪场中,仍然采用人工监控生猪养殖视频的方式,其监控实时性不高,容易因疲劳造成误检和漏检。为了解决这一问题,本文以规模化养猪场的生猪为研究对象,重点研究基于时空兴趣点和词袋模型的生猪行为描述与行为识别等算法。本文以导师主持的广东省科技计划项目“猪只采食与排泄行为智能识别及异常预警系统(2012A020602043)”为背景,针对规模化养猪场的实际情况,设计了一套基于网络摄像头的生猪监控方法与一种基于局部表示的生猪行为描述方法,实现了生猪的一些主要行为的识别。在生猪行为描述方面,通过实验分析发现帧差法、混合高斯背景建模法、光流法等基于整体表示的生猪行为描述方法存在无法精确地分割得到生猪轮廓且因为图像噪声和生猪被部分遮挡而鲁棒性偏低等问题,本文选择局部表示方法作为生猪行为描述的方法。通过生猪生物学与行为学研究和当前健康养殖的需要,本文选择扎堆取暖、采食、探究、慢走等生猪行为进行识别研究。通过实验比较Harris和SUSAN两种角点检测算法在生猪图像上的兴趣点检测效果,发现Harris算法比SUSAN算法检测生猪兴趣点的效果更好,本文采用Harris3D时空兴趣点检测算法检测生猪行为视频中变化剧烈的像素点,通过对比实验和结合生猪视频监控实际发现生猪视频金字塔总层数为3时,Harris3D时空兴趣点对生猪行为兴趣点检测的效果较好。通过对比分析生猪扎堆取暖、采食、探究、慢走行为的时空兴趣点分布情况,发现Harris3D时空兴趣点能较好地检测出生猪四种行为的运动部位及各自规律。为了统计时空兴趣点邻域内的梯度方向和光流方向统计量,采用以时空兴趣点为中心的(△x,△y,△t)时空体内的HOG/HOF描述子描述生猪局部时空特征。在生猪行为识别建模与分类方面,本文采用词袋模型对生猪四种行为进行建模,采用Matlab实现了对HOG/HOF描述子的K-Means聚类和将HOG/HOF映射到词袋中。经过K-Means聚类和HOG/HOF描述子映射后,得到生猪四种行为的平均直方图,分析发现生猪四种行为之间差异性较大,基于Harris3D时空兴趣点和词袋模型的生猪行为识别可行,且估计生猪行为识别准确率较高。将对生猪行为建模的词频直方图向量作为SVM的特征向量进行行为分类识别实验。最后分别选择广州市从化、天河两个不同的规模化养猪场进行了总共5天的视频采集。实验结果表明,本文提出的生猪行为识别算法准确率达到92.31%,能够很好地识别生猪行为,词典大小为100时行为识别算法的性能最高。
[Abstract]:The analysis of animal behavior parameters based on computer vision has important research value and application space, and it has far-reaching influence on traditional livestock and poultry breeding methods. Nowadays, most large-scale pig farms still adopt the method of manual monitoring of live pig breeding video, whose real-time monitoring is not high, and it is easy to misdetect and miss detection due to fatigue. In order to solve this problem, this paper takes pigs from large-scale pig farms as the research object, and focuses on the algorithms of pig behavior description and behavior recognition based on space-time interest points and word bag model. Based on the Guangdong Science and Technology Program Project "Intelligent Identification and abnormal early warning system (2012A020602043) for Pig feeding and excretion behavior", this paper aims at the actual situation of large-scale pig farms. A set of live pig monitoring method based on webcam and a pig behavior description method based on local representation are designed to realize the identification of some main behaviors of live pigs. In the description of pig behavior, the frame difference method, mixed with Gao Si background modeling method, was found through experimental analysis. There are some problems, such as optical flow method, in describing pig behavior based on global representation, such as the problem that the pig contour can not be accurately segmented and the robustness is low because of image noise and partial occlusion of live pig, and so on. In this paper, the local representation method is selected as the method of pig behavior description. Based on the study of pig biology and behavior and the needs of healthy breeding, this paper chooses to identify the behavior of live pigs such as heating, feeding, exploring, slow walking and so on. By comparing the effects of Harris and SUSAN corner detection algorithms on pig images, it is found that Harris algorithm is better than SUSAN algorithm in detecting interest points in live pigs. In this paper, Harris3D space-time interest point detection algorithm is used to detect pixel points in pig behavior video. Through comparative experiments and combined with pig video monitoring, it is found that the total number of pyramid layers of live pig video is 3. The Harris3D space-time interest point has a good effect on the detection of pig behavior interest point. By comparing and analyzing the distribution of temporal and spatial interest points of pig heating, feeding, exploring and slow walking behavior, it is found that Harris3D space-time interest points can well detect the movement sites and their respective laws of the four kinds of behavior of live pigs. In order to calculate the gradient direction and optical flow direction statistics in the neighborhood of space-time interest points, the local space-time characteristics of live pigs are described by using HOG/HOF predictors in (x, y, t) space-time with space-time interest points as the center. In the aspect of pig behavior recognition modeling and classification, this paper uses word bag model to model four kinds of pig behavior, uses Matlab to realize K-Means clustering of HOG/HOF predictors and maps HOG/HOF to word bag. After K-Means clustering and HOG/HOF descriptive mapping, the average histogram of the four behaviors of live pigs is obtained. it is found that there are great differences among the four behaviors of live pigs, and the recognition of pig behavior based on Harris3D space-time interest points and word bag model is feasible. It is estimated that the accuracy of pig behavior recognition is high. The word-frequency histogram vector of pig behavior modeling is used as the feature vector of SVM to carry on the behavior classification and recognition experiment. Finally, two different large-scale pig farms in Guangzhou, Conghua and Tianhe, were selected to collect video for a total of 5 days. The experimental results show that the accuracy of the proposed algorithm is 92.31%, and it can recognize the behavior of live pigs very well, and the performance of the algorithm is the highest when the dictionary size is 100.
【学位授予单位】：华南农业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;S828

【参考文献】