The resurgence of large-scale conventional warfare, exemplified by the ongoing conflict in Ukraine, has highlighted the critical importance of modern technologies in enhancing situational awareness and target acquisition on the battlefield. In particular, the integration of unmanned aerial vehicles (UAVs) and artificial intelligence (AI)-driven computer vision systems has emerged as a key enabler of real-time intelligence and precision engagement. This paper presents an approach to enhance object detection models on aerial imagery for military applications. Initial experiments revealed shortcomings in detecting certain object classes, particularly in complex environments and under variable lighting conditions. To address these issues, the paper investigates impacts of cross-dataset training to improve the robustness and accuracy of object detection models. Through selective label integration and careful dataset curation, the paper demonstrates that incorporating assets form external sources significantly enhances generalization and detection performance. The results underline the potential of leveraging large-scale annotated datasets to augment domain-specific applications with minimal additional labeling cost.

In the information age, often referred to as the era of the digital revolution, the collection and the processing of data is an essential capability in both civilian and military environments. Thanks to the rapid development of technology, the intertwining of machine learning and remote sensing provides us with many opportunities, especially in the areas of information collection, processing and interpretation. The importance of effective management and use of information is unquestionable - information is the cornerstone of strategic decisions in many areas. The article will examine the opportunities provided by technology.

Object detection and tracking on RGB and thermal imaging is getting increasingly more advanced and important in the field of computer vision. The increasing image quality combined with decreasing camera prices during recent years, created new opportunities to gather data. They are now used in many different fields, for example: medical, military, law enforcement or industrial applications. Thermal cameras are useful in darkness and in situations when there is not enough light for cameras that operate in the visual spectrum to function properly. Leveraging the information provided by colored images for labeling, this paper approach aims to enhance the accuracy and efficiency of object recognition in thermal imagery. The synergy between thermal and colored imagery offers a comprehensive solution, addressing challenges posed by limited thermal resolution and ambiguity in object boundaries. By harnessing the complementary strengths of colored and thermal imaging, this paper opens avenues for applications in surveillance, autonomous systems, and other fields where reliable object detection and tracking under challenging thermal conditions are paramount. For this experiment different models were trained using MMDetection and MMRotate modules from MMLab framework. We explored and compared multiple models like: Rotated Faster R-CNN (two-stage anchor-based detection), Single Shot Alignment Network (S2ANET one stage anchor detector), and FCOSR (one-stage anchor-free detector)

Since their invention, drones have been used in many areas, both civil and military. Among these uses, some of the more important ones are reconnaissance, patrolling and object recognition. In most of the applications, objects in images are detected using neural networks. To carry out this process, datasets with a lot of training samples should be collected. Unlabeled data takes a lot of time and human resources to process. Carrying out real drone measurements is a long process and is affected by many external factors, such as weather conditions. A simulation environment can provide a solution to this problem, where different drone measurements can be simulated in varied environments. One of the advantages of the simulation environment is that the objects on the acquired images can be automatically labeled, which is a very time-consuming process in the case of manually labeled images. An additional advantage is that it can perform measurements on diverse terrains with realistic objects with the help of specified configurations, thus generating a large and diverse set of training samples with minimal human intervention. The paper contains information on how to generate different terrains, navigate them with a virtual drone, taking pictures and labeling them at the same time. Development possibilities include dynamically changing the appearance of objects, as well as introducing moving objects into the simulation. It has been proven that it is possible to generate a diverse training set with the help of the proposed simulation.

The thesis explores the security challenges of the 21st century and the military applications of technological advancements, with a particular focus on the integration of Unmanned Aerial Vehicles (UAVs) and Artificial Intelligence (AI). UAVs play a critical role in modern military operations by enabling efficient reconnaissance, target identification, and marking, while minimizing risks to human lives. AI offers innovative solutions for developing autonomous systems and processing data, especially through neural networks and deep learning algorithms. The aim of the thesis is to investigate how the training datasets required for UAVs can be supplemented with artificially generated images, particularly for devices with limited or inaccessible source material. The research examines the effectiveness of synthetic samples in training deep learning models for object detection in real-world scenarios. The findings contribute to the development of UAV-supported reconnaissance systems, reducing the time and cost associated with data collection while improving algorithm performance in diverse military environments. The study highlights the innovative role of AI and UAVs in modern warfare and underscores the strategic importance of technological advancements in the global security landscape.

Unmanned Aerial Vehicles (UAV) imagery-based object detection has become more and more widespread, especially in fields like agriculture, city planning, construction oversight or even military reconnaissance. This research focused on a review of the state-of-the-art object detection algorithms which use oriented bounding boxes. The images were collected using DJI Tello and Zll SG906 Pro 2. The images were annotated using Computer Vision Annotation Tool (CVAT). The different detectors were trained using MMDetection and MMRotate modules from MMLab framework. The advantages and disadvantages of different deep neural network approaches are aimed to be compared: Rotated Faster R-CNN (two-stage anchor based detector), Single Shot Alignment Network (S2ANet) (one-stage anchor detector) and FCOSR (one-stage anchor-free detector). On neural network models two series of measurements were performed. In the first version, all the image samples were trained during the epochs. In the second version the images were grouped in tasks, and in each epoch a new tasks were load to train. Based on the measurement results, it can be concluded that better results were achieved with the Rotated Faster R-CNN model and the S2ANet model, the FCOSR performance lagged behind the previous models.

The paper focuses on satellite image based object detection. Satellite images-based object detection is being used for an increasingly wide range of tasks. For object detection deep neural networks are used. The training set was created from the DOTA dataset by splitting and scaling the images. As neural network, in this research YOLO architecture is used, more specifically the Yolov4 and Yolov4-tyni configurations. The test set was created from satellite-based images on different scales. The dataset for testing was managed and labelled in CVAT annotation program. In the paper we try to synthesize the results achieved during the training and testing of the mentioned neural networks. Objects are very diverse in size, the recognition rate during testing is greatly influenced by how much the image size is proportional to the size of the images used during training. The images in the testing dataset were scaled to a different size. We tried to evaluate how much the scale of the images and size of the object affect the rate of recognition.

Drone imagery based object supervising has become more and more widespread. In the paper the Single shot Alignment Network is used to classify and localize the objects. The images were acquired by using two types of drones, DJI Tello and Zll SG906 Pro 2 in about thirty classes, and about nineteen were processed and detailed in the paper. The objects labeling was realized in CVAT labeling tool. For neural network management the mmdetection framework was used, and the obtained results were detailed on s2a-net. The paper focuses on the preparation of the neural network system to be used for agricultural machine detection. The network was trained on a PC with reduced processing capabilities. The dataset was cut in smaller tasks. An architecture is proposed to be used in future for dataset management during the training process.

Artificial neural networks (ANN) are widely used in solving problems like image processing, data mining, or classification. Hardware accelerators are used for increasing the performance and efficiency of neural networks. An option for implementing such an accelerator is the usage of an FPGA-based system, although developing neural networks for FPGAs is very time-consuming and requires hardware design knowledge to do it. This problem tried to be solved by creating a framework that should speed up the design process. At the same time, there is an overall outlook on some efficiency optimization and speed-up options as well. The framework is written in Python and generates a C++ code whit HLS directive. This code can be compiled by Vivado HLS into a hardware descriptive language and packaged as an IP. The Vivado tool can generate a bit file that can be uploaded onto the FPGA device.Among other things, the paper presents a comparison of different approximations of nonlinear transformations (basis functions and activation functions) in terms of accuracy, required resource, and delay needed for evaluating the transformation. The generated neural network module was integrated into a system that was developed by the authors. Using that system, the neural network module was tested and compared to the models implemented in Python.