摘要
为高效提取生猪的行走快慢特征,以微软公司Kinect作为图像采集设备,采集生猪运动深度图像序列。在对各帧深度图像进行骨架提取、剪枝的基础上,采用基于路径相似性骨架图匹配法确定生猪前后肢骨架端点,进一步以骨架端点所属骨架枝子集像素值特征判定端点远近侧属性。以生猪前后肢远、近侧端点的帧间相对坐标变化建立了生猪运动模型,提出了通过帧间坐标变化点集拟合正弦曲线计算生猪行走完整步的方法。最后,通过计算序列完整步与序列采集时间长度比值提取生猪步频特征。通过对采集的28个生猪运动深度图像序列及其镜像序列共56个图像序列进行的试验,表明该文提出方法的正确率达到82.1%。该项研究对于开展生猪异常步态分析,进一步建立生猪多源特征融合的计算机视觉异常监测系统,提高生猪异常行为预警可靠性具有重要意义。
To further research the extracting method of the pig gait features, the paper firstly focus on the extraction of the pig gait frequency. A gait frequency extraction method was proposed based on analyzing the skeleton endpoints of depth image. Firstly, a series of processes, including skeleton extracting and pruning, were taken to the frames of depth image sequences. Secondly, a path similarity skeleton graph matching method was introduced to distinguish the fore-leg endpoints and the hind-leg endpoints from the skeleton graph. Then considering the characteristics of the depth image, a rule to distinguish the far-side endpoint and the near-side endpoint was constructed by calculating the average value of neighbor skeleton points of the endpoint. After ascertaining the skeleton endpoints of four legs, a variable was defined to represent the relative position of the far-side endpoint and the near-side endpoint, along the horizontal direction between frames in a sequence. Furthermore, the fitting sine curves were used to represent the variations of the fore-leg endpoints and the hind-leg endpoints separately. At last, the reciprocal of the fitting sine curve frequency can be calculated and the INTPART of double reciprocal was regarded as the fore-leg steps（FS） or the hind-leg steps（HS）. The complete step（CS） was defined as the minimum of FS and HS. The finally gait frequency can be calculated by using the CS value to divide the duration of the sequence. To verify the proposal method, 28 depth image sequences of pig moving were acquired by using the KINECT depth camera, at the Rongxin pig farming of Danyang city in Jiangsu province, China. Another 28 sequences were achieved by mirror transforming along the horizontal direction to the native sequences. Experiments were taken for all the 56 sequences by using the proposal method. Experimental results show that the success rate of the method proposed in this paper is 82.1%, up to 92% for the situation when the pig moves continuously and the moving directions is perpendicular or nearly perpendicular to the axis of the depth camera only. Incorrect results often appear when the pig stays for a long time between steps or by non-cross steps, it needs to further adapt the proposal method. For the situation of rough variation of the pig body occurring in the sequence, the proposal method is not suited because the matching of skeleton points can not be achieved. That is the insufficient point of the proposal method. The proposal method would help to carry the further research of the abnormal gait of pig and construct the abnormal monitoring system by fusion of multi-source vision features.
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2014年第10期131-137,共7页
Transactions of the Chinese Society of Agricultural Engineering
基金
国家自然科学基金资助项目(31172243)
教育部博士点基金资助项目(20103227110007)
江苏高校优势学科建设工程资助项目(苏政办发〔2011〕6号)
关键词
步态分析
图像处理
模型
骨架端点
深度图像
生猪异常监测
gait analysis
image processing
models
skeleton endpoints
depth image
pig abnormal monitoring
