Research On Highlight Event Detection Method In Basketball Video

With the fast development of multimedia techniques and internet techniques, an explosive proliferation of sports video is made available on broadcast and Internet. How to find the video that we are interested in is becoming an evident need in an era of “information explosion”, especially for basketball sports enthusiasts, How to just browse the highlights in the basketball match instead of watching the whole video game has become an urgent problem to be solved, and highlights event detection has become a hotspot in research of sports video analysis. Highlights event detection event detection method according to its research content can be divided into two main directions: methods based on feature optimization and methods based on model improvement. Methods based on feature optimization aims to improve the wonderful event detection results by constructing new features which are able to express the video content more properly or combining the existing features. This method often needs to define specific features according to the specific video content, so it is difficult to find a generic feature for all sports video. Methods based on model improvement aims to improve the wonderful event detection results by adjusting the machine learning method structure to express the video content more properly, the universality of this method is high but the results are not satisfactory by far. This thesis is mainly researching on the basketball match video from perspectives of both feature definition and machine learning method selection, and proposes a new highlights event detection method for basketball video. The main research work of this thesis are as follows:(1)Feature extraction. Firstly, the basketball video are divided into two parts of the video image data and video and audio data according to its content: the audio data are directly using for audio feature analysis and extraction; the video image data are further divided into three parts of video frame, video shot and video event/video scene according to its structure, and then the image feature are extracted from video frame layer and middle level semantic feature are extracted from the video shot layer. In the end, feature extraction experiments are conducted to prove the effectiveness of feature definition of this thesis.(2)Basketball shooting and foul event detection. The support feature of each highlight is firstly found out using the concept lattice clustering technology according to the audio-video features and middle level semantic features defined in this thesis, and then the support feature are weighted to construct the affective arousal feature. Then, the audio shot are processed to obtain the whistle lens features using whistle lens detection method defined in this thesis. Whit the affective arousal feature and the whistle lens features are combined as input, an effective HCRF(Hidden Conditional Random Field) is constructed to realize highlight detection of basketball shooting and foul. In the end, Experimental results show the effectiveness of this thesis.Finally, summarize and analyze the results of this study, as well as future research are discussed.