Prediction Of Free Throws In Basketball Games Based On Video Human Pose Estimation

In basketball games,free throws are an extremely important way of scoring,and free throws can often change the direction of the game and even decide the winner or loser of the game.In recent years,the team’s training on free throw techniques has gradually become standardized,and professional shooting coaches are employed to guide players on free throws.However,traditional free throw analysis is based on vision or sensors to analyze athletes in training,which is time-consuming and labor-intensive.,and it is difficult to make accurate and stable predictions about the state of players’ free throws in actual games.This thesis studies the algorithm of target detection and human pose estimation.The video analysis technology of deep learning brings a more digital effect to the prediction of free throws.The main contents of this thesis are as follows:(1)Collect a large number of recorded videos of basketball games from different perspectives and different venues and preprocess them,label the free-throw players in the videos for subsequent target detection model training,label the videos according to the results of the free-throw,and label the video frames with the free-throw players’ free-throw Text labels are added to the sequences and the corresponding free throw results to complete the establishment of the dataset.This thesis adopts the deep learning target detection algorithm and improves it.In view of the problem that the loss function of the original algorithm converges too slowly when the prediction frame does not coincide with the target frame and does not consider the aspect ratio relationship between the target frame and the detection frame,this thesis adopts The CIo U loss function accelerates the convergence speed while considering the aspect ratio,improves the model performance,and introduces an attention mechanism to improve the recognition accuracy of the network.(2)The joint point information is obtained by recognizing the player through the method of human gesture recognition,and the adjacent frame judgment and repair are carried out for the identified wrong joint point.In view of the problem that the position of the limbs cannot be unified due to the different body shapes of the athletes,after obtaining the key point information,the absolute position information is converted into the angle information,which improves the stability of the recognition,improves the recognition effect in the actual complex scene,and reduces the uncertain interference.Factors such as changes in field illumination,athlete motion blur,occlusion,etc.,affect the results of limb recognition.(3)Based on the above method,a free throw hit probability prediction system is implemented.The system is tested,and the test results show that the system can effectively detect the target of the free-throw player and identify its body posture,so as to predict the probability of free-throw hits.