The Research On Player Detection Method In Soccer Game Video

Video object detection is an important research topic in the field of computer vision,and detection of players in soccer game video is an important application of video object detection in sports video activity analysis.And it is the fundamental base of video analysis of soccer match,such as highlight detection,caption extraction,player tracking and tactic analysis and so on.Although the great progress has been made on detection algorithm research recently,there still are many challenges,such as motion blur caused by the movement of players and shots,illumination challenging,heavy occlusion during match and the scale changes of players in different shots.In order to solve these problems,a variety of detection algorithms are proposed.Recently,convolutional neural network detection model based on deep learning outperforms and achieves the state-of-the-art.It is an end-to-end network which integrates muti-tasks:feature extraction,region proposal and detection.And sharing weights among these tasks makes the network more capable of learning on the whole data set.Therefore,starting from the analysis of the characteristics of the convolutional neural network itself and the causes of the player's detection errors,this paper propose some optimization of the model structure and greatly improved the accuracy of player detection.The main work and innovation of this article are as follows:1.Propose a player detection model based on reverse connection convolutional neural network(RC-CNN).It is known that the features extracted from convolutional neural networks are hierarchical,we propose a method that can convey semantic information back to shallow convolutional layers for improve the accuracy of smaller player.And we design the comparative experiments to explore the reasonable position of the model,and also verify the effectiveness of the reverse connection model to improve the detection accuracy of smallscale players.Then,the convolution prediction is performed on the multi-scale feature maps obtained from multiple reverse connection models,and the category confidence and position offset of each bounding boxes are predicted.Finally,the complete suppression in the traditional non-maximal suppression algorithm is improved,and the attenuation suppression is set by judging the overlapping area of the bounding box with the maximum confidence.At last,the network model uses the modified non-maximum suppression algorithm to perform post-processing of the detection results,reducing the situation of missed detection and false detection.Experiments are conducted on two publicly available data sets:Soccer Player and KITTI.The results show that the proposed RC-CNN detection model has achieved good results.2.Explore the application of the proposed player detection model in the virtual presenter location in soccer games.Firstly,by analyzing the characteristics of the images that need to be explained and the characteristics of the images that are suitable for interpreting by the virtual commentator,in order to meet the actual needs,the player detection model proposed in this paper will be used to find the key players by resetting the confidence threshold.Secondly,in combination with the key areas that need to be explained,image frames suitable for interpreting into virtual commentators and having key narration values are determined.Thirdly,based on the scene description requirements and the visual sensory characteristics,the final position and scale of the virtual presenter in the image are determined,thus completing the localization of the virtual presenter.Through testing the video game segment downloaded and segmented from the FIFA World Cup website,the experimental results prove the validity of the proposed algorithm model in the localization task of the virtual presenter.