Research On Karting Competition Behavior Analysis Based On Multi-Target Tracking

At present,karting,as an emerging sport in China,has attracted more and more people’s attention and love.Detecting,tracking and analyzing karts,so as to obtain some exciting sports information in the process of karting competition,and extract the wonderful pictures in the game video,is a hot application of the combination of technology and sports.The main research contents of this paper are as follows:(1)Construct 8 kart detection datasets of different scenes,with a total of 2400(8 × 300)images,and manually mark the position of the kart in the image.Each kart data has both the category label and the upper left corner coordinates and the coordinates of the lower right corner.(2)In view of the problems that the traditional target detection method has poor detection effect on karts,false detection and high missed detection rate,this paper introduces the YOLOv3 algorithm to achieve efficient detection of karts,trains karting detection models in different scenarios and evaluates the models.The average detection accuracy of each model is between 91.38%and 99.76%.Finally,compared with the traditional method,the results show that the detection effect of the method in this paper is better than that of the traditional target detection method.(3)Since it is necessary to obtain the motion information of multiple karts in the same picture to recognize the collision,overtaking and other wonderful pictures in the process of karting competition,this paper adopts the multi-target tracking algorithm to complete the karting tracking task.First,the Kalman filter is used to predict the detection frame parameters of each detected kart and the motion state of the kart is estimated,and the similarity between the image features and motion features of the kart between frames is calculated.The similarity matrix is used as a measurement parameter to obtain the association result between the frame and the frame target.The average tracking accuracy of this method for kart multi-targets is 89.61%.(4)After completing the tracking of the kart,the trajectory change of the kart in the video screen is obtained,and then the scene calibration of the kart track is performed to realize the mutual conversion between the pixel coordinates and the coordinates of the midpoint in the real world,and estimate the real world of the kart.The coordinate changes in the system can be obtained to obtain the complete motion trajectory and speed time series change data.Finally,the LSTM network is used to learn the time series characteristics of the kart trajectory,and the analysis results of the kart overtaking,collision,drift and other behaviors are obtained.(5)The design of the behavior analysis system for kart racing:using python as the basic language,the behavior analysis system is constructed based on the Qt Designer development interface.It can gradually display the results of kart detection,tracking,scene calibration,and behavior analysis.Aiming at the characteristics of high speed and small size of karts,this paper proposes to use the YOLOv3 network with balanced speed and accuracy to detect karts,and use Kalman filtering and Hungarian matching to achieve multi-target tracking of karts with high efficiency and low resource consumption.Aiming at the changeable characteristics of karting trajectories,the method of fusing motion features and image features is used to calculate the similarity between targets.The work of this paper has achieved good experimental results on the established data set,which has certain practical significance for improving the entertainment of kart competitions,and provides technical support for the extraction of wonderful videos of kart competitions.