Automated Referee Whistle Sound Detection for Extraction of Highlights from Sports Video

This paper proposes a simple and automated referee whistle sound detection (RWSD) for sports highlights extraction and video summarization. The proposed method is based on preprocessor, linear phase bandpass finite impulse response (FIR) filter shorttime energy estimator and decision logic. At the processing stage the discrete audio sequence is divided into non-overlapping blocks and then amplitude normalization is performed. Then, a bandpass filter is designed to accentuate referee whistle sound and suppress other audio events. Then, the filtered signal is fed to short-time energy (STE) estimator which includes amplitude squarer and linear filter to obtain a positive signal. In this work, we use decision rules based on the amplitude-dependent threshold and time-dependent threshold for detecting of referee whistle sound regions. The performance of the proposed design is tested using a large scale audio database including American football, soccer, and basket ball. The total duration of the test audio signal is approximately 12 hours and 11 minutes. The proposed method results in time-instants of boundaries of whistle sounds and then time instants are used to automatically extract the sports highlights from the unscripted video. Then, audio perception of the extracted sound segments is performed to indentify the false positive (FP) and false negative (FN). The proposed method has a detection failure rate of 19.4% (42 FP and 26 FN) and detects 324 whistle sounds successfully. The sensitivity and reliability of the proposed design are 92.5% and 80.5%, respectively. The design is implemented in MATLAB 7.0 version environment with the following system specifications: Intel (R) Pentium (R) Dual Quad CPU @ 2.40 GHz and 3 GB of RAM. The computation time is approximately 0.3-second for processing of 1-second block. General Terms Multimedia, Audio Classification, Content-based audio and video retrieval and summarization