System for military vehicle recognition in video streams using expert decision-support rules

The subject of the study is a military vehicle recognition system for video streams that combines a deep learning object detection model with an expert-rule module. The relevance of the work is determined by the need to improve the accuracy, robustness, and interpretability of video analytics systems under dynamic conditions, where high-quality training data are limited and objects are observed under varying lighting, viewing angles, backgrounds, and noise. The aim of the article is to develop and experimentally validate a system capable of recognizing classes of military vehicles in a video stream and generating recommendations based on detection results. To achieve this aim, the following tasks were set: to analyze data preparation approaches, form an extended training dataset using exponential augmentation, train a YOLO-like detector, implement frame-by-frame video analysis, perform spatial interpretation of object coordinates, and integrate the results with expert-rule tables. The research methods include deep learning, object detection, geometric, photometric, and compositional augmentation, evaluation using precision, recall, and mAP metrics, as well as rule-based interpretation of the target class and angular parameters. The main results consist in the construction of a modular pipeline that includes data preparation, detector training, video-stream inference, calculation of spatial parameters, and generation of recommendations. An experiment on a 10-second video fragment showed that augmentation increased the share of correct recognition of the tank class from 24.67% to 30.00% and reduced the number of frames without detection from 70.17% to 52.33%, but increased misclassification. It is concluded that exponential augmentation improves the model’s sensitivity to objects in lower-quality video, but requires classification stabilization, tracking, and further training. The practical significance of the work lies in the use of the proposed approach for monitoring, video analytics, and decision-support systems, while its theoretical significance lies in combining statistical recognition with expert interpretation of the results.