Machine Learning Based Audio-Only football Goal Detection Under Clean and Degraded Acoustic Conditions

Abstract

Automatic football goal detection is impotent for real time events monitoring sports analysis system and intelligent broad casting system. And this detection works based on videos of live streaming, audio of commentary and crowd cheering, and many other technologies. As existing work relies on the visual information mainly or combined audio video modalities. However, during live streaming the video stream may become unavailable, delayed or corrupted, while audio remains accessible. This creates a need for alternative information source for reliable goal detection using audio only. This research works on football goal detection using only audio, particularly under imperfect conditions of audio. This research proposed CNN + Mel-spectrogram based model for football goal detection using only audio, with particular focus on detection under degraded acoustic conditions. The proposed approach uses audios of match, convert them into Mel- spectrogram, which further use as input to Convolutional neural network (CNN) to classify goal or no goal event. This study also investigates the acoustic patterns related to change in audio quality to support reliable goal detection. For examining the real-world streaming conditions, the system evaluated under both clean and artificially degraded audios by including different levels of noise. The no goal situation where strong crowd reaction arises is also evaluated as false goal detection. Performance is evaluated using accuracy, precision, recall and F1- score. The aim of research is to investigate the use of only audio for football goal detection and to determine the feasibility of CNN based models for acoustic representation while visual information is temporarily unavailable and audio quality also compromised.