Measurements, Analysis, Classification, and Detection of Gunshot and Gunshot-like Sounds

Area of Science:

• Acoustic analysis
• Machine learning
• Signal processing

Background:

• Rising gun violence necessitates advanced detection methods.
• Distinguishing gunshot sounds from acoustic clutter is challenging.
• Machine learning offers potential for automated gunshot detection.

Purpose of the Study:

• To analyze feature importance for discriminating gunshot sounds.
• To reduce dimensionality of Mel-frequency cepstral coefficients (MFCCs) for classification.
• To compare feature importance methods and assess sound similarity.

Main Methods:

• Audio data acquisition and measurement of gunshot-like sounds.
• Feature extraction using Mel-frequency cepstral coefficients (MFCCs).
• Application of Random Forest (RF) and SHapley Additive exPlanations (SHAP) for feature importance and reduction.
• Uniform Manifold Approximation and Projection (UMAP) for feature space visualization.

Main Results:

• Identified key MFCC features crucial for discriminating gunshot sounds.
• Demonstrated effectiveness of RF and SHAP in feature selection and reduction.
• UMAP visualization confirmed the separation of gunshot sounds from other noises in the feature space.

Conclusions:

• The developed approach enhances the ability to discern gunshot sounds from background noise.
• Feature importance analysis is vital for building accurate gunshot detection systems.
• This method provides a viable tool for improving public safety through acoustic monitoring.