Detection And Measurement Of Curling Movement Based On Intelligent Vision

In today's global competition,the application of artificial intelligence in the sports industry has gradually begun to rise to the will of the country,which is of vital strategic significance for establishing China's status as a sports power.This topic is based on this situation,with the current research direction of the sports industry ‘curling' as the research background,combined with artificial intelligence technology to bring digitization and information to the traditional curling arena,and built an intelligent curling auxiliary training system.The system can realize the intelligent understanding and analysis of the curling process through the tasks of identifying and detecting curling targets,motion measurement and trajectory prediction.This subject first constructed a multi-camera synchronous acquisition subsystem based on the curling structure and track characteristics.Through this subsystem,multiple curling motion images were collected as research samples,and a first relatively complete curling dataset in China was constructed by manual annotation.The dataset can not only provide training samples for the current algorithm,but also provide sample support for the posture detection and game strategy research.With the support of data samples,this subject mainly studied intelligent curling auxiliary training in two aspects: target detection and motion measurement.In the object detection part,this paper analyzes the current advanced target detection neural network in computer vision and selects two main optimization schemes for network compression and regression strategy improvement based on the characteristics of the curling task.Experimental results prove that the optimized network can reduce a large amount of calculations while ensuring accuracy,reduce the requirements on the computing power of the device,and thus can be better applied in embedded devices and realize the task of real-time detection.The curling motion measurement part is mainly divided into two parts: 3D coordinate calculation and trajectory matching.In the calculation of the 3D position coordinates of curling,we compared the EPn P algorithm,binocular vision method and interpolation method.Through comparison experiments,we obtained that the interpolation method error in static scenes is smaller,and binocular vision is more accurate.We can choose corresponding algorithms for different scene tasks,reflecting the flexibility of the overall algorithm;In the trajectory matching,we innovatively apply the KM algorithm in the coordinate point matching,the curling trajectory is obtained by using the matching pair,which fully utilizes the detection information and the continuity of time and space and the effect in practical application is also very ideal.Then we combined the detection and measurement algorithm to build an intelligent auxiliary training system.The system showed good functions in experimental tests and was able to provide preliminary but accurate auxiliary information for sports training in real time.