Acoustic Event Detection for Crime Monitoring using BI- GRU and Other Deep Learning Methods

Criminal activities such as robbery, assault, and homicide pose significant risks, especially to individuals working alone in remote areas at night, with women being particularly vulnerable. Conventional image-based artificial intelligence detection systems often have limitations in accuracy due to poor lighting and obstructions. This study presents a solution using sound feature extraction to detect potential criminal activities in real-time. Machine learning techniques were applied to audio data from the Urban8K dataset, which includes ten sound classes, along with additional classes such as crackling fire and glass breaking from the ESC50 dataset. Audio files were converted to mono and downsampled to sixteen thousand hertz for preprocessing, and Mel spectrogram features were extracted to differentiate normal and alert sounds. Experimental results indicate that the Bi- Directional Gated Recurrent Unit model outperformed other models, achieving ninety-six point seven seven percent accuracy, ninety-six point six zero percent precision, ninety-three point five zero percent recall, and a ninety-four point eight eight percent F one score. This approach demonstrates the potential of sound-based analysis in improving the accuracy and reliability of real-time event detection systems, contributing to enhanced public safety.