Cow Lameness Behavior Recognition Based On Improved Sparse Overcomplete Dictionary Technology

The main goal of behavior recognition is to extract effective motion features from the input video of several behaviors,and analyze these features to quickly and accurately identify the type of the behavior.Behavioral factors are crucial in assessing livestock health and commercial interests.Therefore,animal behavior recognition technology has been widely used in large-scale animal husbandry.However,in recent years,scholars have found that in the process of animal behavior recognition,the extracted feature points representing animal behavior show sort of redundancy,which has a certain impact on the recognition results.In addition,the computational time and convergence speed of different dictionary learning will change in behavior recognition.Moreover,online recognition technology relatively falls behind in the current recognition technology.In view of the above problems,this thesis will carry out relevant research.The improved sparse and over-complete dictionary algorithm is proposed to identify the lameness behavior of dairy cows.The main work is as follows:(1)Firstly,in order to find the optimal search direction,the conjugate gradient pursuitKSVD(CGP-KSVD)algorithm is used to describe and represent the behavior of dairy cows at the language level.It is applied to the initial stage of sparse coding construction,so that the sparse super-complete dictionary learning algorithm can quickly find the optimal direction,reduce the computational load due to large-scale storage and calculation of Hessen matrix,thus speeding up the convergence speed of the original sparse super-complete dictionary learning algorithm.(2)Secondly,this thesis proposes the technology of fusing spatio-temporal interest points and dense trajectory graph to extract spatio-temporal interest points twice,which is applied to the video bottom feature extraction and representation stage.The proposed method makes the extracted spatio-temporal interest points contain abundant detailed feature information,while reducing redundant features and computing load.(3)In this thesis,a recognition library of cow lameness behavior recognition is constructed,and the framework of recognition method and a comparison method are proposed to verify the validity of the method.The comparison between different perspectives and classical recognition methods is carried out,the validity,generalization,real-time on-line detection and quantitative analysis of the proposed algorithm framework are verified.The experimental results show that the proposed algorithm has high recognition accuracy and generalization ability under the same environment,and the execution time of the algorithm is significantly improved.To sum up,this thesis studies the process of identifying lameness behavior of dairy cows.The application of technology in the whole recognition process is analyzed in detail,and the improved methods are compared and analyzed.Experiments on different scales and perspectives,the related work of this thesis provides a reference for the follow-up study,and relevant technologies can be a necessary complement to non-contact sensor monitoring and other technologies.