AI-Telecommunications Synergy in Public Safety Systems Advancing Intelligent Law Enforcement

Abstract

The increased rate of gun-related events witnessed in the context of the public safety determines the need to have intelligent systems of real-time surveillance in Internet of things (IoT) infrastructures. The current acoustic detection systems have a tendency to fail when trying to classify finer details of the gunshot, operate in a restricted space, and classify acoustically similar types of gunshots. To address this, we propose a class-aware augmentation strategy that selectively modifies specific audio classes to enhance inter-class discriminability, followed by standardized feature extraction at 22,050 Hz. In this paper, we have introduced lightweight Transformer-based model to detect and recognize gunshot instances in real-time and with multiple classes via 128-band log-mel spectrograms. The system operates across edge and fog layers, leveraging Augmented Covering Arrays (ACAs) and a MOEA/D-based optimizer to balance latency, energy consumption, and processing load. To enhance contextual awareness and dynamic threat prioritization, we introduce four intelligence metrics: Crime Risk Score (CRS), Crime Temporal Pattern Index (CTPI), Emergency Response Delay Impact Score (ERDIS), and Threat-Aware Priority Index (TAPI). An AutoML method is applied to optimize hyperparameters of models and reduce the effect of mixed up non-gunshot acoustic phenomena. Experimental results on 13-class gunshot data showed classification accuracy of 99.67%, representing 17.17 percentage point improvement. The macro-averaged F1-score above 0.993. Five-fold cross validation yielded average accuracy of 99.10%. With Streamlit interface the accuracy of the system is 98.10% in real-time implementation which validates the applicability on the use of the IoT to drive public safety.