Pig Behavior Recognition Based On Spatio-Temporal Interest Points

Parameter analysis of animal behavior based on computer vision is of important research value and there is extensive application space in it, the traditional livestock and poultry breeding way have been profoundly influenced by it. Video monitoring on pig is still at the stage of artificial monitor in most of large-scale piggery, howerver, artificial monitor is weak in real-time ability and easy to miss-detected or error-detected because of tiredness. To solve this problem, this paper took the pigs in large-scale piggery as the research object, mainly researched on the algorithm of pig behavior description based on spatio-temporal interest points and bag of words model and the algorithm of behavior recognition.This paper was under support of a Planned Science and Technology Project of Guangdong Province named "Research on Warning Model of Abnormal Behaviour on Pig Eating and Excretion"(Grant No. 2012A020602043). According to the actual situation of the large-scale piggery, a pigs monitoring solution based on web camera was designed and a algorithm of pig behavior description based on local presentation was presented, mean while achieve the recognition of some kinds of the major pig behaviors.In terms of pig behavior description: frame difference, mixed Gaussian background modeling, optical flow were implemented to describe pig behaviors. They cannot exactly segment the piggery background and pig and their robustness are weak as the image noise and pigs were partly occluded. To solve these problems, local representation was implemented to describe pig behaivor. A research was done on pig's biology and behavioral science. After that, four pig's behaviors such as getting together to keep warm, eating,exploring and walk were chosen to be classified. A comparision experiment betwwen Harris and SUSAN on the interest pint detection on pig image was done. The experiment showed that Harris performed well than SUSAN on the interest point detection on pig image. A improved Harris called Harris3 D spatio-temporal interest point detectionalgorithm was implemented to detect the pixel point which has a great change on its value in pig behavior video. Accroding to another comparision experiment and the reality of pig video monitor, when the pyramid's total level is 3, Harris3 D performed well on detecting the spatio-temporal interest point of pig behavior. After analysing the distribution of spatio-temporal interest points which were detected when pigs were performing the getting together to keep warm, eating, exploring, walking, it is demonstrated that Harris3 D spatio-temporal interest point can effectively detected the moving part of pig and their regularities when pig were performing some behaviors. In order to get the statistics of orientation of gradient and optical flow in the local neighbourhood of spatio-temporal interest point, HOG/HOF descriptor was used to describe the pigs' local spatial-temporal feature, whose center is the detected Harris3 D spatio-temporal interest points, and the size of local volume is(?x, ?y, ?t).In terms of pig behavior modeling and classification: Firstly, the K-Means clustering algorithm is used to cluster the HOG/HOF descriptor to build a bag of words model, and the HOG/HOF descriptors are mapped into the bag of words to generate histogram vectors for modeling pig behaviors. Secondly, Finally, the histogram vectors generated before are taken as support vectors of SVM to classify four pig behavior.Finally, two different large-scale pig farms in Guangzhou Conghua, Tianhe were selected to capture video for five days. The experimental testing showed that:the proposed pig behavior recognition algorithm's rate of accuracy were 92.31% and it can effectively classify pig behaviors, mean while when the dictionary size is 100, the algorithm is state-of- the-art.