ENet: Event Based Highlight Generation Network for Broadcast Sports Videos

Abdullah Aman Khan,Yunbo Rao,Jie Shao
DOI: https://doi.org/10.1007/s00530-022-00978-8
IF: 3.9
2022-01-01
Multimedia Systems
Abstract:Handcrafting sports video summaries based on highlights and important events from broadcast sports videos is a laborious and time-taking task. Amateur content creators and professional bodies around the world spend hundreds of man-hours to keep the audience up to date with the latest happenings by means of such highlights. In this paper, we present a deep learning-based method capable of automatically generating highlights from a broadcast sports video based on important events and user preferences. Our proposed method classifies the broadcast sports video scene to generate a summary based on highlights or important events. As various sports have different rules, playfield scenarios, and high inter-class similarities, it is quite challenging to devise a generalized method capable of handling different categories of sports. To overcome such problems and to enhance the highlight generation performance, the proposed method internally segregates the sports category and then utilizes various convolution neural network based feature extraction branches to recognize the important events. Additionally, a branch selector mechanism is introduced to select the relevant convolution neural network branch, which predicts the important sports event/activity. We performed extensive experiments using different deep learning architectures. In terms of important event recognition, the results of the experiments show the superiority of our proposed method.
What problem does this paper attempt to address?