Abstract :
[en] Proper management and maintenance of infantry firearms are critical to operational readiness, safety, and cost control within modern military contexts. Weapon fleets degrade differently depending on their usage. Traditional methods of tracking weapon usage through manual shot counting are inaccurate and inefficient. Electronic shot counters enable preventive and predictive maintenance by providing armourers with precise, quantitative measures of weapon usage.
Traditional electronic shot-counting solutions require expert knowledge and extensive study for each firearm. In response to market demands for increased customisation and shorter lead times, recent machine-learning-based solutions have been proposed. However, these solutions are limited by the difficulty of acquiring sufficiently large, fully-labelled datasets, restricting their generalisation capabilities.
To address this limitation, we propose EDGAR (Embedded Detection of Gunshots by AI in Real-time), a novel technique capable of working directly with data labelled only by the total number of events in a time series. This approach significantly reduces labelling efforts, enabling the creation and use of datasets several orders of magnitude larger than those typically available. Furthermore, we show how the technique can be extended to effectively support discrimination between live and blank ammunition, as well as detection of suppressor usage, with minimal additional computational overhead. We demonstrate that these classification tasks can be executed in under 100 ms on highly constrained embedded microcontrollers, thus enabling real-time shot detection.
Extensive experiments conducted across a range of firearms, including FN Minimi, FN MAG, FN M2HB-QCB, and M134 Minigun, demonstrate that EDGAR significantly outperforms unsupervised algorithms and achieves comparable or superior performance to human-generated state-of-the-art algorithms, particularly for discrimination tasks. Practical field deployments validate the robustness and real-time capabilities of the proposed method.
Finally, leveraging the large datasets made accessible through our approach, we empirically investigate the impact of dataset size on model performance. We identify critical thresholds required for effective model generalisation and provide practical guidelines for efficient dataset acquisition and model training.
These results enable a new generation of electronic shot counters and targeted maintenance strategies, thereby reducing maintenance costs, preventing incidents, and increasing the operational availability of weapon fleets.
