Data Collection, Modeling, and Classification for Gunshot and Gunshot-like Audio Events: A Case Study

Area of Science:

• Audio signal processing
• Machine learning for event detection
• Public safety technology

Background:

• Distinguishing dangerous audio events like gunshots from non-dangerous ones, such as plastic bag pops, is critical for public safety and resource allocation.
• Current audio event detection systems, including deep learning models, often confuse plastic bag explosions with real gunshots.

Purpose of the Study:

• To investigate the acoustic properties of plastic bag explosions and gunshots.
• To develop and evaluate a deep learning model capable of accurately differentiating between gunshot sounds and plastic bag pop sounds.
• To compare the effectiveness of Mel-frequency cepstral coefficients (MFCC) and Mel-spectrograms for audio feature extraction.

Main Methods:

• Recorded a dataset of plastic bag-popping sounds under various conditions (bag size, distance, environment).
• Trained a convolutional neural network (CNN) classification model using a combined dataset of gunshot sounds and recorded plastic bag pop sounds.
• Evaluated model performance using MFCC and Mel-spectrogram features for distinguishing between the two sound types.

Main Results:

• A deep learning model trained solely on urban sound datasets with gunshots failed to distinguish plastic bag pops.
• The same CNN model, when trained with added plastic bag pop sounds, demonstrated high performance in differentiating between the two audio events.
• Both MFCC and Mel-spectrograms proved effective for feature extraction in the classification task.

Conclusions:

• Incorporating specific non-threatening sound data, like plastic bag pops, into training significantly enhances the accuracy of deep learning audio event detection models.
• The developed CNN model shows promise for real-world applications in distinguishing life-threatening from non-life-threatening audio events, potentially reducing false alarms and optimizing public safety responses.