农业工程学报
農業工程學報
농업공정학보
2013年
10期
168-174
,共7页
马丽%纪滨%刘宏申%朱伟兴%李伟%张涛
馬麗%紀濱%劉宏申%硃偉興%李偉%張濤
마려%기빈%류굉신%주위흥%리위%장도
养殖%计算机视觉%图像处理%单只猪%完整轮廓%侧视图
養殖%計算機視覺%圖像處理%單隻豬%完整輪廓%側視圖
양식%계산궤시각%도상처리%단지저%완정륜곽%측시도
cultivation%computer vision%image processing%single pig%full contour%profile
由于完整轮廓猪只的侧视图具有便于行为分析的价值,因此,研究从猪舍监控视频中自动分割出单只猪理想侧视图的视频段对猪的行为分析是有意义的.为了识别每帧图像猪轮廓图的侧视图属性,该文通过图像处理获取猪只轮廓图后,提出联立猪只外接矩形高宽比和低频傅里叶系数构建猪只侧视图的特征向量,并根据样本训练集得到理想侧视图和非理想侧视图特征向量均值和方差,利用马氏距离判别法从测试视频中识别未知帧图像的类别,结果表明有91.7%猪只轮廓图的侧视图属性能正确识别,表明本方法是有效的.本研究可为后继单只疑似病猪行为分析提供条件.
由于完整輪廓豬隻的側視圖具有便于行為分析的價值,因此,研究從豬捨鑑控視頻中自動分割齣單隻豬理想側視圖的視頻段對豬的行為分析是有意義的.為瞭識彆每幀圖像豬輪廓圖的側視圖屬性,該文通過圖像處理穫取豬隻輪廓圖後,提齣聯立豬隻外接矩形高寬比和低頻傅裏葉繫數構建豬隻側視圖的特徵嚮量,併根據樣本訓練集得到理想側視圖和非理想側視圖特徵嚮量均值和方差,利用馬氏距離判彆法從測試視頻中識彆未知幀圖像的類彆,結果錶明有91.7%豬隻輪廓圖的側視圖屬性能正確識彆,錶明本方法是有效的.本研究可為後繼單隻疑似病豬行為分析提供條件.
유우완정륜곽저지적측시도구유편우행위분석적개치,인차,연구종저사감공시빈중자동분할출단지저이상측시도적시빈단대저적행위분석시유의의적.위료식별매정도상저륜곽도적측시도속성,해문통과도상처리획취저지륜곽도후,제출련립저지외접구형고관비화저빈부리협계수구건저지측시도적특정향량,병근거양본훈련집득도이상측시도화비이상측시도특정향량균치화방차,이용마씨거리판별법종측시시빈중식별미지정도상적유별,결과표명유91.7%저지륜곽도적측시도속성능정학식별,표명본방법시유효적.본연구가위후계단지의사병저행위분석제공조건.
The use and wide application of video monitoring and control systems in pig pens is necessary for the automation and analysis of intelligence to improve the development tendencies of the pig-raising industry. The complete profiling of a pig is convenient to behavior analysis and to judge if it is sick. A pig’s wandering causes the lack of its full profile being captured on video. For example, only part of the body or no body at all in the image when intelligently monitoring a single pig in a pigpen will result in an abnormal profile. Not all profiles, or camera angles, of the pig are efficient for behavior analysis in the case that the pig’s profile is not fully revealed, such as the pig directly facing the camera. Only scenes with the fully exposed profile of the pig are convenient for observing and analyzing the symptoms of the pig. Therefore, it is essential to automatically segment the video that is monitoring the pigpen. A novel method is proposed for setting the frame attributes based on the level of exposure to a pig's profile in a standing posture. Two types of attributes are presented for each frame in the video recording, which represent the applicable and non-applicable profile of a pig. A pig’s profile feature vector is calculated for differentiating the attributes after obtaining a profile of a single pig by image processing. The vector is composed of both the aspect ratio of the pig’s external shape and a group of low frequency Fourier coefficients based on the pig’s contour. The aspect ratio varies with the angle between the axis of the pig body and the horizontal line on the left side of the ground. The result of the test shows that the aspect ratio index is ideal and plays a significant role only in both the acute angle areas, i.e., [?30?, 30?] and [?150?, 150?]. Eight low frequency Fourier coefficients are also verified enough to describe the characteristic shape of the pig body in reconstructing the profile image of a single pig by the Fourier inversion method. Means and variances of both the feature vectors of the applicable and non-applicable profile contours are obtained from the sample training set. The category of the unknown frame is identified by the Mahalanobis distance discrimination from the testing video. The results showed that 91.7%of frames in the pig’s profile could be properly recognized. Therefore, this method of producing the frame attributes based on a single pig’s contour is reliable. This study may be helpful for subsequently analyzing behavior of a single pig in suspected cases.