Deep Learning Based Soldier Face Detection and Counting Method for Military Tactical Operations and Artificial Intelligence Powered Weapons

Abstract

In battlefield and tactical operations, automatic detection and tracking of soldiers or military personnel is essential and plays vital roles in soldier presence detection, soldier density calculation, soldier movement tracking, and direction of soldier arrival determination for autonomous robot navigation and war tank direction control, soldier facial expression and emotion analysis, and facial-based soldier authentication. Existing radio frequency (RF) or wireless radios-based soldier localization and tracking approaches suffer from frequent communication failures and heterogeneous channel interferences, and also cannot be used to locate a specific soldier in a particular troop or soldier’s injury conditions. In this paper, we present a YOLOv9 architecture-based soldier face detection and counting method using visual images that can be effectively used for soldier presence detection, tracking, and soldier density calculation. The YOLOv9 face detection (YOLOv9-Face-Detect) model is trained by using a wide variety of face databases having only civilian faces. The trained YOLOv9-Face-Detect model is tested using 250 soldier images having diverse image qualities and soldier density with soldiers wearing different types of personal protective equipment (PPEs), which are taken from public websites. The effectiveness of the YOLOv9-based soldier face detection and counting method was evaluated in terms of the estimated and reference face counts. The method had a mean absolute error of 0.87 and a root mean square error of 2.06 persons per image for the soldier image database, with a model size of 49.2 MB and inference time of 17.66 ms. Evaluation results on the different image qualities demonstrate that the YOLOv9 network-based image analysis can be an effective solution for soldier presence sensing and monitoring applications.