| Intelligent animal husbandry has been widely studied and practiced as a typical application program of intelligent agriculture.Multi-target individual recognition has become an important part of the modernization and wisdom of animal husbandry.How to accurately identify different individuals in the same kind of targets in the shortest possible time is one of the key problems faced in the current visual recognition to achieve real-time accurate identification of similar multi-target individuals.In this thesis,by exploring the limitations and the problems to be solved in the multi-target individual recognition methods studied by previous authors,a method based on the invariance of physical spatial distance is proposed for the recognition of multi-target individuals.The specific research work is as follows:I.A feature vector generation method based on spatial distance invariance is proposed.This study uses a localization model based on monocular vision,extracts the world coordinates of target feature points through the localization model,calculates the line segment distance between feature points using Euclidean distance,and uses the set of line segments as the feature vector of the target.II.a two-stage determination method for target recognition is proposed.The first stage is mainly to perform matching of feature vectors,determine the error threshold by analyzing the distance error distribution of the feature vector elements,i.e.,line segments,determine the identical elements in two feature vectors and count them;further propose the similarity calculation method of the feature vectors,determine the similarity between the vectors by calculating the ratio of the number of identical elements to the total number of elements of the two feature vectors.The second stage is to determine the similarity error threshold by analyzing the similarity distribution between the feature vectors,especially the similarity between the feature vectors of different moments or stages of the same target,and determining whether the targets are the same.Experiments show that the accuracy of the method in this thesis is 99.742% in calculating world coordinates;the average accuracy of line segment length measurement is98.687%;the recognition rate is 97.096% in two-dimensional target experiments;the recognition rate is 95.2% in three-dimensional target experiments;and the recognition rate is 92.326% in bull recognition applications.The method studied in this thesis can reduce the number of feature points and reduce the feature dimensionality.This method has important research significance and application value for realizing real-time accurate identification of individual identity of clustered cattle and sheep in smart agriculture and animal husbandry,as well as real-time accurate identification of individual identity in similar targets in other fields. |
