Abstract
There are major security concerns in the transportation sector of developing countries, and this is due to the lack of adequate infrastructure put in place to combat this issue. This paper presents a controlled evaluation of a rule-based system implemented on a Raspberry Pi 5 platform for analysing audio signals and detecting violence in real-time. For this exercise, four acoustic features were evaluated. These were Root Mean Square (RMS) energy, Zero Crossing Rate (ZCR), High-Frequency Ratio, and Spectral Centroid. The evaluation was conducted using 40 pre-recorded audio samples gotten online. These audio samples comprised of gunshot sounds, human (male and female) screaming, and sounds of normal daily activities. The results gotten from the exercise showed a detection accuracy of 92.5%, precision of 100%, recall (sensitivity) of 85%, specificity of 100%, and an F1-score of 91.9%. The screaming samples achieved 90% detection, while the gunshot samples achieved 80%, with missed detections resulting due to the short-duration of the acoustic signatures they possessed. It was also observed that the first time of running the system (cold start) had a significant delay of about 3-8 seconds each time it was booted. Subsequently, the reruns had a reduced delay of about 1-2 seconds as it had already familiarized itself with the activities. From the findings gotten, it has been confirmed that rule-based feature fusion delivers very competitive detection performance without the extra computational demands of different machine learning approaches. This makes it a practical and affordable solution for violence detection in resource-constrained public environments.

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Copyright (c) 2026 Scholarly Journal of Science and Technology Research and Development