Novel Mutual Information Analysis of Attentive Motion Entropy Algorithm for Sports Video Summarization

This study presents a novel summarization method, which utilizes attentive motion analysis, mutual information, and segmental spectro-temporal subtraction, for generating sports video abstracts. The proposed attentive motion entropy and mutual information are both based on an attentive model. To capture and detect significant segments among a video, this work uses color contrast, intensity contrast, and orientation contrast of frames to calculate saliency maps. Regional histograms of oriented gradients based on human shapes are also adopted at the preliminary stage. In the next step, a new algorithm based on mutual information is proposed to improve the smoothness problem when the system selects the boundaries of motion segments. Meanwhile, differential salient motions and oriented gradients are merged to mutual information analysis, subsequently generating an attentive curve. Furthermore, to remove non-motion boundaries, a smoothing technique based on segmental spectro-temporal subtraction is also used for selecting favorable event boundaries. The experiment results show that our proposed algorithm can detect highlights effectively and generate smooth playable clips. Compared with existing systems, the precision and recall rates of our system outperform their results by 8.6 and 11.1 %, respectively. Besides, smoothness is enhanced by 0.7 on average, which also verified feasibility of our system.