Football Field Keypoint Detection Via Deep Lea Rning

In the era of artificial intelligence,humans try to give machine intelligence,so that machines can" see "the world,recognize objects,describe scenes',and react like humans.Deep learning came into being,as the most powerful weapon of artificial intelligence,which has caused extensive research and attention.Although deep learning has achieved historic progress in the related research of pictures,there is still much room for optimization in the understanding of video.In the research of football video comprehension,tasks like players tracking and recognition,automatic collection of control rate and other tactical data,and automatic football commentary combined with NLP don't have mature solutions currently.This thesis focuses on the key-point detection of football stadiums.In the research of football video comprehension,tasks like players tracking and recognition,automatic collection of control rate and other tactical data,and automatic football commentary combined with NLP don't have mature solutions currently.The key position information obtained by the key-point detection algorithm can provide relative position information for the player identification and tracking task in the football video comprehension project.It also can effectively reduce the error caused by the angle change of video and provide the necessary position information for the automatic football commentary and tactical generation task.The main work and research innovations of the paper are summarized as follows:(1)According to the characteristics of the video dataset of football match,this thesis established the corresponding evaluation metric,and proposed a detection-followed-by-regression cascaded CNN.In this thesis,the proposed CNN with two levels used the classic idea of face key-point detection and human key-point detection:the idea of cascade regression is used to detect the key position from coarse to fine;The key-point heatmaps obtained by the first-level network fuse with the original image as supplementary information;the second-level regression network uses the Heatmap as the Ground Truth,providing pixel-by-pixel supervision to regress key-point position and predict whether the point is visible or not.(2)This thesis leveraged adversarial leaniing to design loss function and solve the regression-to-the-mean problem[26]when regressing Heatmaps.We used the second-level network as a generator,then designed a discriminator and defined the loss function to judge Heatmap's reliability.Using the adversarial learning training method,the second-level network can predict a reliable Heatmap,and then get the key-point coordinates.(3)This thesis firstly designed experiment to verify the feasibility of the two-level network,and then optimized the network structure.The fir-st-level network used a VGG-FCN-based neural network and a ResNet-based residual network.The second level network uses a 7-layer simple neural network and a deep network based on the Hourglass structure.At the same time,two strategies for feature fusion were developed for experimental tuning.Finally,the optimal network structure based on ResNet and hourglass is obtained.The average error of the key point position reduces to 14px,and the accuracy of predicting whether the key point is visible is 95.6%,which meets the requirements for use.