Object Acquizition of Stationary and Moving Targets
Myroslav Riabyi, Sergey Edward Lyshevski, Роман Сергійович Одарченко · 2024
This paper studies near real time obj ect acquisition by means of detection, identification and engagement on stationary, moving and maneuvering targets in the air and land domains. Using the streamed videos from monochrome and color cameras, a machine vision concept accomplishes object detection by recognizing the object-specific edges and shapes which yield unique geometric features. Object recognition with descriptive identification at reasonable accuracy is performed using the Jaccard index. The region proposal detectors with the intersection over union is implemented by a convolutional neural network. The machine learning models, implemented by the low power system-on-module, were trained. These machine learning algorithms and models support feature detection and matching, localization, and, three-dimensional spatiotemporal reconstruction. The two-and three-dimensional geometry with object and background contrast and color mappings are considered. Subsequent tracking and locking on maneuverable targets is performed using spatio-temporal three-dimensional analytics. The airborne machine acquisition module with trained models is demonstrated. A concept suits monochrome, color and thermal analog and digital cameras. Experiments in typified operational environments demonstrate acceptable accuracy, convergence, low latency and near real-time capabilities for different object categories and classes within developed custom datasets.