Football Game Video Analysis Method with Deep Learning.
Researchers
Journal
Modalities
Models
Abstract
Football is a beloved sport, and its wide audience makes football video one of the most analytically valuable types of video. Researchers have achieved certain research results in football video content analysis. How to locate interesting event clips from a complete long video is an urgent issue to be addressed in football game video analysis. The granularity of sports event detection results with traditional machine learning is relatively coarse, and the types of events that can be detected are limited. In recent years, deep learning has made good progress in the research of video single-person events and action detection, but there are few achievements in the detection of sports video events. In response to this problem, this work uses a deep learning method to build an event detection model to detect events contained in football videos. The whole model is divided into two stages, in which the first stage is utilized to generate candidate event fragments. It divides the football video to be detected into a sequence of frames of a certain length and scans using a sliding window. Multiple frame sequences within a sliding window form a segment, and each segment is a prediction unit. The frame sequence features within the segment are obtained through a three-dimensional convolutional neural network, which is used as the input of each time point of the bidirectional recurrent neural network and further integrated to generate the event prediction of the segment. The second stage is to further process the above results to remove all segments predicted as nonevents. The thresholds are set according to the detection effect of various events to filter out event fragments with higher probability values, obtain the start and end positions of the events through merging, classify and mark them, and finally output complete event fragments. This work has carried out comprehensive and systematic experiments to verify correctness of the proposed method.Copyright © 2022 Nian Liu et al.