A Study On The Scene-recognition Based Advertisement Embedding Technology In Sports Videos

Sports event nowadays has become a vital source of income for event host and broadcaster.To maximize the profit,they'll try to advertise as much as possible during the event.However,the actual amount of advertisement is often limited by playing field and the experience of audiences,etc.Advertisement embedding system could lift such limitation.It can place an advertisement at any size,anywhere on demand,without affecting the watching experience of audiences.This technology can also be used for showcasing extra information,such as the score of the match,player's profile,etc.In such an application,it's important to fuse the real scene together with virtual information,and the effectiveness of the fusion determines the system's performance directly.In this thesis,we place scene recognition at the core of our advertisement embedding system.It provides the essential information for embedding by analyzing the transformation between the playing field in the video and the real playing field with feature detection and tracking.There are several companies had already developed such a system for sports video and put their solution out in market overseas.Recently,along with the sheer increasing amount of online video platforms and live broadcasting services on the Internet,domestic facilities and companies also started to devote resources to the field of research and achieved some result.Due to the diversity of sports videos,it's hard to design a unified framework that is capable of processing every kind of sports video.In our work,we'll use tennis video as the primary example.There are three main topics discussed in this thesis:The first topic is the feature point extraction of the playing field in the video frame.There are a lot of noises in a video frame regarding to playing field,including players and audiences.We need to extract feature from the field despite the interference from noises.In our work,we start from the most widely used court feature,field lines.We observe the feature of field lines in tennis court first,and then generate a mask for field lines in the frame using Lightness channel in HSL model.Finally,we use a variant of Hough Transformation to extract those lines from the mask.The second topic is the matching between the playing field in a video frame and a standard playing field.In our work,we propose a new court matching algorithm based on matching field lines.Then we calculate perspective transformation between those two fields according to the matching result.This transformation provides essential information for the actual embedding.The third topic is field tracking and stability analysis.Because both field detection and matching are very expensive to compute,we exploit the temporal continuity in video in order to achieve higher efficiency.Also,the embedded advertisement might suffer from “jittering” due to the error that aroused from feature extraction,field tracking,and perspective transform calculation.To improve the stability,we implemented moving average filter,a data smoothing technique based on statistic method,in our system.