The development of intelligent systems is making a significant impact on various aspects of life, especially in the military. However, in Vietnam, research of intelligent systems that utilize deep learning (DL) in the military domain remains limited. In this work, we will propose an intelligent system for military applications that utilizes modern deep learning technologies. This system is designed to address the task of automatic recognition and scoring in military training, consisting of two main modules: the action recognition module and the scoring module. Our contribution consists in the implementation of a comprehensive pipeline for the action recognition module with a skeleton-based approach, proposing a method to evaluate training video quality by calculating the distance between joint movement vectors in reference videos and trainee videos; presenting a complete experiment that integrates both modules on the dataset of military training videos. The construction of a military training video dataset is also part of the work we have undertaken. Experimental results demonstrate that the system based on our proposed method is feasible and has great potential for practical applications. With an accuracy of up to 95% for the action recognition module, while the scoring module can analyze and assess the differences and execution quality of the actions between the reference videos performed by experts and the training videos performed by trainees, based on various distance metrics.