Gun identification from gunshot audios for secure public places using transformer learning

Area of Science:

• Artificial Intelligence
• Signal Processing
• Security Systems

Background:

• Mass shootings and terrorism pose significant societal threats, necessitating advanced security measures.
• Existing weapon detection systems often rely on visual analysis, which can be limited in certain environments.
• There is a need for real-time, cost-effective automated systems to enhance public safety.

Purpose of the Study:

• To develop and evaluate an automated system for detecting and classifying firearm types (rifle, handgun, none) based on gunshot audio.
• To compare the performance of convolution-based and transformer-based (self-attention) neural network architectures for audio gun detection.
• To assess the effectiveness of Mel-frequency-based audio features in distinguishing different gunshot sounds.

Main Methods:

• Utilized Mel-frequency cepstral coefficients (MFCCs) as audio features for gunshot analysis.
• Implemented and compared two deep learning architectures: Convolutional Neural Networks (CNNs) and Transformers (fully self-attention-based).
• Trained and evaluated the models on audio clips containing different types of gunshots and non-gunshot sounds.

Main Results:

• The transformer-based architecture demonstrated superior generalization capabilities on audio features compared to convolution-based methods.
• The proposed transformer model achieved a high classification accuracy of 93.87% for gunshot type detection.
• Achieved low training and validation loss values of 0.2509 and 0.1991, respectively, indicating effective model training.

Conclusions:

• Audio-based gun detection using transformer architecture is a viable and effective method for enhancing security.
• The developed model can function as a standalone weapon detection system or complement existing visual systems.
• This approach offers a promising solution for real-time threat detection in public spaces, improving overall safety.