基于立体视觉技术的生猪体重估测研究

发布时间：2018-01-07 10:17

本文关键词：基于立体视觉技术的生猪体重估测研究　出处：《中国农业大学》2016年博士论文　论文类型：学位论文

【摘要】：本研究针对我国设施养猪场现有养殖工艺问题,拟解决生猪生长过程中实时无接触自动估测猪只体重的难题,用于指导生产。对估测过程中动态获取猪只标准姿态图像,实时检测猪只体尺,建立猪只体重估测算法等关键技术进行探索,提出新的猪只体重估测模型,优化基于深度图像的猪体轮廓提取算法,开发图像自动筛选算法,研发适用于小圈群养的移动式猪体重无接触自动估测系统,并在实验猪舍和典型猪场进行验证。主要结论如下：(1)研究猪体重估测模型。优选适合机器视觉检测和人工验证的5个体尺,选用79组数据建立基于体尺主成分的幂回归模型；使用实验站测量97组数据验证,估测体重和实测值的相关系数为0.998,相对误差在4%以内,平均相对误差为2.02%。24组猪场试验数据表明,验证模型对猪个体估测体重平均误差为2.08kg,平均相对误差为(2.26±1.78)%,模型具有较高的估测精度。(2)基于LabVIEW图形化软件开发平台,构建立体视觉三维检测系统。该系统在2m物距范围内的三维检测误差小于1%,其中X轴、Y轴检测平均相对误差皆为0.65%,Z轴检测平均相对误差为0.34%；在所有区域中,视野中央检测精度最高。研发猪体背部三维检测与点云重构技术,与高精度激光三维扫描数据配准,处理猪体背部共计255587个点集,三维点云的平均误差为-3.29±4.51 mm,验证了系统的三维检测精度。(3)开发基于深度图像的猪体轮廓提取算法,解决灰度轮廓提取算法无法适应猪场恶劣光环境的问题。使用相对参数筛选体尺测点,能够较稳定地提取猪体尺检测关键点；增加手工提取关键测点的功能。测量猪场32组体尺数据,与人工测量体尺相比,猪体尺检测平均误差小于2cmm,相对误差在2%左右,体尺检测的精度较高。(4)开发生猪体尺体重无接触检测软件,构建基于对称度和延长系数的正常猪只姿态图像筛选算法,对异常图像的识别率为79.31%,对正常图像的识别率为73.06%,软件对猪只自由状态下的图像,体尺检测成功率为24.93%。研发移动式猪体重检测平台,能够检测多个猪圈的猪只体重,适用于我国现有的小圈群养的生产工艺,增加系统检测猪只容量,降低系统使用成本。在实验站开展长时间估测试验,体长、体宽、体高、臀宽、臀高的检测精度分别为1.44%、5.81%、4.94%、2.00%和1.64%。系统对4头猪共计268组体重估测数据的平均相对误差为(2.52±2.14)%。实现猪体尺体重无接触全自动估测。(5)利用物联网技术,开发生猪生产过程实时监控与管理系统。具备基于C/S架构的生态及生产过程数据采集功能,并结合专家决策技术,通过基于B/S架构的网页及手机客户端形式对授权用户发布相关信息。主要监管功能包括环境参数、生产过程信息及猪只行为音视频监控。
[Abstract]:In this study, our pig breeding facilities field existing process problems, to solve the growth process of pig in real-time non-contact automatic estimation problem of body weight of the animals, to guide the production of pigs. Dynamic access to the standard attitude estimation process image, real-time detection of pig body size, a pig weight estimation algorithm of key technology, put forward a new pig weight estimation model, optimization of extraction algorithm for contour depth image of pigs based on automatic screening algorithm for image development, research and development group of mobile circle pigs weight non-contact automatic estimation system, and verified in the experiment and the typical pig farms. The main conclusions are as follows: (1) study of pig weight estimation model suitable for. Machine vision detection and manual verification of 5 individuals using 79 pairs of feet, body size data to establish a regression model based on principal component of the power station; 97 groups were measured using experimental data Verify, estimation of correlation coefficient of body weight and the measured value was 0.998, the relative error is less than 4%, the average relative error is 2.02%.24 group of pig test data show that the model of individual pig weight estimation of the average error is 2.08kg, the average relative error is (2.26 + 1.78)%, the model has high estimation accuracy (2). LabVIEW software development platform based on the construction of stereo vision three-dimensional detection system. The system of 3D detection error in the 2m object in the range of less than 1%, the X axis, Y axis measuring the average relative error is 0.65%, the Z axis detection of average relative error was 0.34%; in all areas, the accuracy of the central visual field testing. Research of porcine somatic back three-dimensional detection and point cloud reconstruction, registration and high precision of 3D laser scan data processing, a total of 255587 pigs back point set, the average error of 3D point cloud is -3.29 + 4.51 mm, verify the system of 3D The accuracy of the detection. (3) the development of depth image contour extraction algorithm based on gray pig, solve the contour extraction algorithm can not adapt to the harsh light farm environment. Using the relative parameters of the screening body size measurement points, can be stably extracted from pig body size detection of key points; increase the manual extraction of key points of measurement of 32 swine function. Group size data, compared with the manual measurement of body size, pig body size detection with an average error of less than 2cmm, the relative error was about 2%, higher precision of body size detection. (4) a pig body size and body weight of non contact detection software, constructing normal pigs attitude image algorithm symmetry and extend the coefficient based screening, identification the abnormal image recognition rate of 79.31% of the normal image rate was 73.06%. The software of image on pigs under free condition, the success rate of body size detection developed for 24.93%. mobile BW detection platform, can detect a pigsty The body weight of the animals, suitable for the production process of our existing small herds, increase the system capacity reduction system uses detection of pigs, the cost. In the experiment station to carry out long time estimation test, body length, body width, body height, hip width, hip high detection accuracy were 1.44%, 5.81%, 4.94%, and 2% the 1.64%. system of 4 pigs in a total of 268 sets of estimating weight average relative error data for (2.52 + 2.14)%. The pig body size and body weight of non contact automatic estimation. (5) the use of networking technology, real-time monitoring and management system of the development of pig production process. With data acquisition and production process of ecological function based on C/S architecture and combined with expert decision technology, through the B/S structure of the website and mobile phone client form based on authorized users to publish relevant information. The main functions include the regulation of environmental parameters, the production process information and pig behavior audio and video monitoring.

【学位授予单位】：中国农业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：S828;TP391.41

【相似文献】