Weapon Activity Recognition

Kennedy HeckerAAAI 2025 · DOI

Introduction

Modern military training systems, such as the Multiple Integrated Laser Engagement System (MILES), effectively track target accuracy but lack the capability to monitor how a soldier operates their weapon. This gap in data prevents comprehensive assessments of weapons safety, procedural correctness, and overall readiness. Currently, monitoring these behaviors relies on direct human observation, which is not scalable for large-scale squad training. This paper proposes an automated Weapon Activity Recognition system using machine learning to identify specific mechanical interactions with firearms, ensuring adherence to safety protocols and doctrine.

Methodology

The proposed system utilizes Trigger Detection Modules (TDMs)—lightweight devices equipped with accelerometers—attached to M249 and M240 machine guns. To ensure robust data collection, sensors were placed on the pistol grip, the weapon rail, and the front breast pocket of an assistant gunner.

The core of the recognition system is a Long Short-Term Memory (LSTM) neural network. LSTMs are particularly suited for this task due to their ability to capture temporal dependencies in sequential data. This allows the model to differentiate complex, multi-step actions like "clearing a jam" from background noise or locomotive activities.

The researchers expanded the standard Wireless Sensor Data Mining (WISDM) dataset with custom tactical data. This included standard soldier activities (walking, rucking, riding in an HMMWV) and weapon-specific tasks (firing, clearing, reloading). Additionally, proximity data was collected to test the model's ability to distinguish a user's gunfire from ambient noise produced by nearby weapons.

Results and Discussion

The shift from general human activity recognition to human-tool interaction allows the system to treat the weapon's mechanical state as an extension of the soldier's activity. Preliminary results indicate high precision in distinguishing weapon-specific tasks from general movement.

Confusion matrix showing high precision for activities like 'Clearing'.

As shown in Figure 1, the "Clearing" activity was identified with 100% accuracy (2,163 instances) without confusion with locomotive activities like walking or rucking. This validates the hypothesis that unique mechanical motions are easily distinguishable from general soldier movements. Furthermore, the model maintained high accuracy for standard activities (e.g., 4,242 correct for walking), demonstrating that the inclusion of weapon-specific classes does not degrade baseline performance.

Conclusion and Future Work

While initial results are promising, future research will focus on testing in erratic operational environments. The team aims to refine the model to distinguish between more subtle weapon actions, such as different types of malfunction clearances. Additionally, efforts will be made to develop a user interface that provides commanders with actionable feedback rather than raw classification data, enhancing the practical utility of the system in the field.

Summary updated

BibTeX
@inproceedings{hecker2025weapon,
  author = {Kennedy Hecker},
  title = {{Weapon Activity Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35330},
}
Figure