Enhancing UCAV operations with AI-driven point cloud semantic segmentation for precision gimbal targeting in defense industry

Abstract

The widespread integration of technological advancements has fundamentally transformed the field of artificial intelligence, significantly enhancing the reliability of AI model outputs. This progress has led to the widespread use of artificial intelligence in various sectors, including automotive, robotics, healthcare, space technologies, and defense industries. Particularly in the field of aerial combat, target identification and engagement operations still heavily rely on human operator intervention. Within the scope of this thesis, the aim is to automate the complex and error-prone laser designation process using 3D point clouds and deep learning algorithms. The primary dataset for the study consists of 3D point clouds obtained by processing gimbal images of the Bayraktar AKINCI Unmanned Combat Aerial Vehicle (UCAV) using photogrammetric methods. For initial evaluations and parameter optimizations, the DublinCity 3D LiDAR point cloud data was used. The DublinCity dataset was created using Airborne LiDAR methodology in the capital city of Ireland, Dublin, in 2015. This dataset is hierarchically organized into 13 classes, including buildings, vegetation, ground, and undefined, divided into four main categories. Within these main categories, there are subcategories such as windows, doors, trees, and others. For this study, we used the PointNet++ and RandLA-Net algorithms, two widely recognized approaches for point cloud segmentation. Both algorithms are designed to process point clouds directly and deliver segmentation results. However, a key difference lies in their handling of data: while RandLA-Net can incorporate both geometric and color information, PointNet++ traditionally relies only on geometric features. To address this limitation, we modified the PointNet++ algorithm to utilize color attributes, allowing for a more comprehensive analysis. This enhancement represents a significant contribution of our research. By comparing the improved PointNet++ with RandLA-Net, we observed noticeable differences in their performance, particularly in how they handle datasets with combined geometric and color information. In tests conducted using only geometric features in the RandLA-Net algorithm, an accuracy rate of approximately 94% was achieved. When color information for points was also provided to the algorithm, the accuracy rate significantly increased to approximately 97%. In tests conducted with the PointNet++ algorithm, an accuracy rate of 94% was observed when only geometric features were used. However, the accuracy rate increased significantly to approximately 96% when the PointNet++ algorithm was enriched with color information. The results of this research highlight two main contributions. Primarily, combining point clouds produced from different sources with AI-driven decision-making processes provides substantial benefits for the defense industry in aerial combat activities, such as target identification, monitoring, and neutralization. Secondly, modifying the PointNet++ algorithm, which originally relied exclusively on geometric data, to include color information has greatly enhanced the accuracy of learning and decision-making in 3D point cloud processing tasks. This research seeks to offer a dependable and effective approach to reduce human involvement in laser designation procedures, especially by utilizing data gathered from the Bayraktar AKINCI UCAV. Upcoming research will aim to enhance precision by incorporating higher resolution point cloud data and examining different deep learning algorithms. Furthermore, analyzing data gathered from advanced RADAR systems and enhanced photogrammetric point clouds will be a key emphasis for upcoming studies. Additionally, this thesis features an extensive preprocessing stage, enhancing the DublinCity dataset and photogrammetrically produced Bayraktar AKINCI UCAV information by organizing them into separate categories. The intricate framework of the DublinCity dataset, consisting of 13 categories, provided a varied basis for assessing segmentation algorithms. In this process, distinguishing highly specific classes (e.g., windows, doors) created a chance to examine the connection between detail level and algorithm precision. This part of the research highlights the difficulties and constraints associated with handling high resolution, intricate data. Apart from defense applications, laser, RADAR, and gimbal systems offer numerous potential uses. The techniques and strategies created in this research can be readily modified for use in other fields. For instance, in disaster management, these systems could be employed for automatic debris identification and directing rescue teams during natural disasters like earthquakes or floods. In urban design and infrastructure management, these technologies can greatly enhance procedures such as 3D city visualization and automated extraction of building inventories. In agriculture and forestry, they might be used to improve soil productivity, identify damaging structures, and track plant health. Likewise, in the preservation of cultural heritage, these systems can aid in 3D mapping of archaeological locations, identifying relics, and comprehensive documentation of historical items. In summary, this thesis illustrates the successful use of deep learning algorithms in automating laser targeting processes for aerial combat applications involving UAVs within the defense sector. These studies not only improve the effectiveness of current technologies but also establish a base for creating autonomous systems. The findings of this research possess significant theoretical and practical implications. Upcoming developments in this area seek to integrate more intricate datasets, including those from radar technologies, and evaluate different algorithms to foster additional innovation.