Research On The Key Technology Of Curling Video Motion Analysis

In the upcoming 2022 Beijing Winter Olympic Games,a variety of winter sports events are in full swing.In our country,the slogan of "Science and Technology Winter Olympic Games" is put forward.And scientific and technological means are proposed to assist athletes for training.This thesis is aimed at the training of curling-maker athletes.The classic matches in the previous videos are helpful for fixing the action of athletes by using repeated training.In order to automatically obtain key matches from the video of a curling match,key frame detection is needed,and then the detection of the curling object is needed to reproduce the positions of the curling in the key frames.This thesis focuses on two tasks: key frame detection and object detection.The main works include as following:Firstly,in order to obtain the changes of the two sides,a key frame detection algorithm by combining shot boundary detection and image classification is proposed for curling video.It can automatically obtain the curling location changes in the barracks area.The match video usually shows the movement process of curling from the home plate area to the barracks area,so that the key frame containing the barracks area is usually at the shot boundary.In addition,the barracks area contains the large ring,which has obvious visual characteristics.Based on the above observations,the framework is proposed.Firstly,candidate key frames are obtained by shot boundary detection.Furthermore,by constructing a classification neural network,the candidate key frames are classified,the redundant key frames are removed,and the real key frames are extracted.The experimental results show that the proposed algorithm achieves the optimal performance and efficiency.Secondly,for the problems of inaccurate object detection due to ignoring global information in the anchor-free object detection algorithm and unbalance between positive and negative samples in the key point object detection algorithm,an object detection algorithm based on the ratio information is proposed.By predicting the length,width and ratio information of the target box,the global characteristics of the target box are used to predict the more accurate prediction boxes.At the same time,compared with the key point based object detection algorithm which only regards the corner or center of the object as positive samples,our algorithm regards all the points in the boxes as positive samples,which alleviates the imbalance of positive and negative samples to a certain extent and makes the network easier to converge.Experimental results demonstrate that the proposed algorithm achieves good performance on curling data set and MS-COCO data set.Finally,in the existing object detection algorithms,the center region of the object box is directly assigned with the highest confidence score.When the center region is occluded,it almost contains no object features,resulting in the generation of high-score but low-quality prediction box,which possibly reduces the performance of the algorithm.To solve this problem,an object detection algorithm based on adaptive feature weight assignment is proposed.By introducing the attention mechanism,the network is designed to automatically learn the confidence of different regions in the target box,so that the final predicted high confidence region is concentrated in the region with higher recognition,rather than the fixed central region,so as to improve the accuracy of object detection.Experimental results demonstrate that the performance of the proposed algorithm is further improved on curling data set and MS-COCO data set.