The Lameness Recognition Of Pigs Based On The Space-time Interest Point

Intelligent video monitoring is widely used in precision farming areas, such as lameness detection, breathing detection, behavior recognition and so on. Animal behavior perception based on the technology of video surveillance has become a precise research hot spot in the field of animal husbandry.In this paper,based on machine vision, we proposed recognition method of the pig lameness based on space-time interest points aiming at the problem of lame recognition of pig in videos.The method in the actual motion video of the pig obtained the good performance of lame recognition. This study has far-reaching significance in respect of pig lame detection,prevention of spreading and real-time processing.Firstly,aimed at the problem of the key parts of the movement detection of pigs in videos in this paper,we used two kinds of detection algorithms of space-time interest points based on Gabor filter and space and time constraint respectively.The algorithm based on Gabor filter can detect the interest points obtained from the video which have a significant intensity change of image in aspect of time,corresponding to the pigs’ moving parts in the videos, which can detailedly describe the behavior of pigs. Although the detection algorithm can get rich interest points, however, due to the complexity of the environment, the light changes and other reasons,the detection results of the algorithm contained many background points, thus affecting the accuracy of the identification method.In order to solve this problem, this paper went a step further to use the detection algorithm based on space and time constraint.Through analyzing the spatial and temporal distribution of interest points in videos, the algorithm used space and time constraint inhibiting two kinds of background points in the videos.The first one is the point with no obvious structural information, such as the point of homogeneous regions in videos. the other is the point which has no significant change in the time axis with obvious structural information,such as the point at he fence in the background.Secondly, aiming at the problem of feature quantization and recognition, this paper used the Gauss mixture model and the soft quantization method.Traditional word packet model uses the Kmeans clustering algorithm to obtain the visual dictionary, and uses hard vote to get aquantitative characteristics, which ignores the relationship between local descriptors, thus easily leading to larger quantization error.In this paper,we got multiple Gaussian components by training the Gauss mixture model. Each Gaussian component was used as a visual word to construct a visual dictionary. Then we quantified the video based on soft voting which takes the probability analysis of descriptors into account thus reducing the quantization error very well.Finally, we did cross experiment with two detection algorithms of space-time interest point and two quantitative methods. In this paper,we detected the interest points and quantified the features with two kinds of methods respectively. Among them, when we used the detection algorithm based on time and space constraint and adopted the feature quantization method based on Gauss mixture model and soft voting, the lame recognition rate reached 96%.The experimental results show that the proposed method can distinguish between normal and lame videos and improve the recognition rate. The study has far-reaching significance on the timely lame detection and fast processing of the pig.