Weak Supervised Behavior Recognition Algorithm Based On Domain Knowledge Graph

With the increasing maturity of artificial intelligence technology,the integration of artificial intelligence and technical and tactical analysis of competitive sports is an important way to improve the level of technical and tactical analysis.In this paper,according to the technical characteristics of karate,to improve the human behavior recognition algorithm,in order to improve the robustness and practicability of the algorithm in the karate technique and tactics analysis system.At present,the behavior recognition algorithm based on the spatio-temporal convolution model of human bone data has received more and more attention and achieved good results.However,there are still some deficiencies when applied to the analysis of karate techniques and tactics.To solve these problems,this paper carries out the following work:(1)The training and test videos used by many behavior recognition algorithms only contain a single action,while in the practical application of karate technique and tactics analysis,there are only long videos containing multiple actions,which leads to a serious decline in recognition accuracy.In this paper,the existing sequential action detection algorithm is used to divide the long video into one video segment containing only a single action,which provides a more effective input data source for the behavior recognition algorithm.(2)Most of the traditional behavior recognition algorithm models are trained based on the general data set in the open field,so it is difficult to generalize to specific fields.However,the data set of competitive sports such as karate is seriously missing,which adds challenges and difficulties to the recognition of sports behavior.In this paper,an end-to-end weakly supervised behavior recognition algorithm is designed,which integrates domain knowledge graph to express external prior information,and effectively solves the problem of serious absence of karate data set.(3)In karate this kind of competitive sport,the sports behavior with interference factors,frequent change direction,move more quickly,the characteristics of spatio-temporal figure convolution model is heavily dependent on the performance of figure of topology structure,the conventional topology structure is hard to get this kind of competitive sport in karate effective characteristic information.In this paper,delaunay triangulation algorithm is used to improve the topology diagram construction method,and the external prior information of knowledge graph is incorporated to enhance the robustness and practicability of the algorithm.In this paper,the shortcomings of the above traditional methods are studied in detail,and put forward their own innovative ideas.In view of the deficiency of(1),this paper uses the existing timing motion detection algorithm.The paper solves several other shortcomings of the main innovation points as follows:(1)This paper pre-built domain knowledge graph,and a critical key points in the process of algorithm dynamically build knowledge graph and designed a into the expression of a priori knowledge about the external domain knowledge graph of end-to-end weak supervision and behavior recognition model,can be in small sample karate training data set,and achieve higher recognition accuracy.(2)Aiming at the problem that topology structure feature expression ability is not strong in behavior recognition model,delaunet triangulation algorithm is used in this paper to realize a new topology structure,which can extract spatial structure information in skeleton data more effectively.(3)According to the technical characteristics of karate,the external prior information of the knowledge graph is integrated to make the model more fully express the characteristic information of karate technical movements,and enhance the robustness and practicability of the algorithm.