Aerial Synthetic Aperture Radar (SAR) systems are increasingly used in applications such as autonomous navigation of Unmanned Aerial Vehicles (UAVs), where accurate Above-Ground-Level (AGL) altitude information is critical for both image formation and safe flight operations. Small UAVs often lack payload capacity or power budget for dedicated altimetry sensors, motivating alternative solutions leveraging existing SAR data. This paper proposes a novel method to estimate AGL altitude directly from SAR range-profile data, without any additional hardware. The approach exploits the characteristic energy step in the range profiles caused by the nadir ground echo to infer the slant range to ground using a robust edge detector. The method is conceptually simple, computationally efficient, and readily integrable into existing SAR processing chains. Experimental validation on real airborne X-band SAR data from a fixed-wing aircraft shows that the proposed approach achieves meter-level accuracy, with a root-mean-square AGL error below 1% of the nominal flight altitude and median absolute errors on the order of a single range-resolution cell, demonstrating its suitability for terrain-relative navigation and support of autonomous flight.

Edge detection is a fundamental building block in synthetic aperture radar (SAR) image analysis. Speckle phenomena present in SAR images poses a challenge for detection and localization of edges. Hence, there is a need for robust edge detection methods specifically designed for such images. In this paper, three different models for edges present in SAR images are introduced and a new robust SAR edge detection (ROSED) method is proposed. The method uses a ratio of medians (ROM) kernel for detecting clutter boundaries combined with a modified constant false alarm rate (CFAR) target detector for point targets and line-like edges. Finally, the log likelihood ratio test is applied for precise edge localization. The results show that the proposed method is robust to biasing effects and provides more complete and accurate edges compared to other methods.

The scope of Synthetic Aperture Radar (SAR) image registration is rapidly expanding beyond traditional multi-modal applications to include emerging domains such as SAR odometry, navigation, and SAR-based SLAM, where accurate registration between sequential SAR images is essential. In this work, we explore the feasibility of using deep neural network (DNN) featurematching models for SAR-to-SAR image registration. A new dataset of SAR image pairs was constructed to facilitate training and evaluation. Three state-of-the-art DNN models-ROMA, SuperGlue, and ELoFTR-were tested. ROMA, a dense matcher, achieved high accuracy without additional training, demonstrating strong generalization. In contrast, SuperGlue and ELoFTR performed poorly with pretrained weights but showed substantial improvement after fine-tuning on the SAR dataset. SuperGlue's rotation RMSE decreased by 35.3 % (from 0.3265° to 0.2111°), and x-translation error dropped by 55.5 % (from 6.70 m to 2.9797 m). ELoFTR exhibited even greater gains, with an 82.1 % reduction in rotation RMSE and over 95% improvement in xtranslation accuracy. All models achieved sub-meter accuracy with sub-second inference times, demonstrating the potential of fine-tuned DNN matchers for real-time SAR-SAR registration tasks.

In this paper we propose a multi-session simultaneous localization and mapping algorithm using multiple input multiple output synthetic aperture radar images. Our algorithm uses only radar data to calculate odometry, loop-closure and inter-session constraints as well as to generate a map of the traversed environment. The proposed algorithm was validated through real-world data displaying better overall trajectory estimation, e.g., 79.3% improvement in terms of mean absolute error, as well as expanded radar-generated maps.

In this paper we propose a radar-only simultaneous localization and mapping algorithm based on multiple input multiple output synthetic aperture radar images. The algorithm distinguishes itself from others by depending only on radar data for generating synthetic aperture radar images for estimating traversed trajectory and building a visual representation. In our algorithm, ego-velocity (estimated using only radar data) is used for generating synthetic aperture radar images. The generated radar images are used for rotation estimation in the odometry step as well as for place recognition by exploiting the Fourier-Radon image registration approach. After the trajectory is optimized, we combine coherent and incoherent processing over the radar data for generating a map of the traversed area. The proposed concept was evaluated over multiple sequences comprising heterogeneous and dynamic environments. The results show high performance of the algorithm in terms of place recognition, attaining a balanced f-score in the range of 0.86–0.96. Moreover, the algorithm also achieves good results in terms of simultaneous localization and mapping. For example, it achieves an absolute trajectory error of 0.11 m for a trajectory of length 340 m, and 0.43 m for a trajectory of length 1092 m. Finally, we also include a case study in which we show the capability of the radar-only localization and mapping solution in operating under scenarios that are challenging for global navigation satellite systems.

In this paper, a radar-only 2D odometry estimation algorithm is presented. The algorithm stands out by proposing a statistically robust method for aggregating two sources of rotation estimation obtained using radar generated images. The proposed idea was evaluated using real-world data collected with a radar mounted on a mobile robot. When compared to two other alternative algorithms, the proposed method achieved an improvement in the average absolute relative pose error of up to 30.2% and 45% in the translation and rotation part respectively.

Being able to measure and track positions of mobile systems is an important capability in many applications, autonomous driving being one major example. With the Robust Radon Radar Odometry algorithm, this paper presents an approach for estimating the odometry of vehicles based on radar data only. The algorithm embodies a robust method for estimating the change in orientation as a key feature. The odometry algorithm is under the realm of direct methods, and it exploits properties of the Fourier transform for decoupling the changes in orientation from the changes in translation. In the first step, the Radon transform along with phase-correlation, outlier removal, robust measure of central tendency, keyframe selection and graph optimization are used in order to achieve a robust method for estimating the change in orientation, next the translation is estimated with the support of phase-correlation. The algorithm's performance was evaluated with real world data. Significant improvements in position and orientation error in terms of relative pose error and the KITTI odometry error metric are shown as compared to other direct methods for radar based odometry.

Since Unmanned Aerial Vehicles (UAVs) became available to the civilian public, it has witnessed dramatic spread and exponential popularity. This escalation gave rise to privacy and security concerns, both on the recreational and institutional levels. Although it is mainly used for leisure and productivity activities, it is evident that UAVs can also be used for malicious purposes. Today, as legislation and law enforcement federations can hardly control every incident, many institutions resort to surveillance systems to prevent hostile drone intrusion. Although drone detection can be carried out using different technologies, such as radar or ultra-sonic, visual detection is arguably the most efficient method. Other than being cheap and readily available, cameras are typically a part of any surveillance system. Moreover, the rise of deep learning and neural network models rendered visual recognition very reliable [9, 21]. In this work, three state-of-the-art object detectors, namely YOLOv4, SSD-MobileNetv1 and SSD-VGG16, are tested and compared to find the best performing detector on our drone data-set of 23,863 collected and annotated images. The main work covers detailed reportage of the results of each model, as well as a comprehensive comparison between them. In terms of accuracy and real-time capability, the best performance was achieved by the SSD-VGG16 model, which scored average precision (AP50) of 90.4

Autonomous driving has been gaining momentum in recent years and is today one of the hottest areas of research and development in the mobility sector. One of the basic tasks to cover in the field of autonomous driving is lane detection. Considering that lane keeping and controlled lane change are low level autonomy tasks, those tasks are essential to any project aiming to achieve a reasonable level of autonomous driving. As a students’ playground, the University of Applied Sciences Upper Austria - along with its partners - is currently establishing a model car based future mobility race event. To make this happen, a ROS based model car is equipped with various known and newly to be developed algorithms enabling certain capabilities. Given the described topical context, in this paper two feature-based lane detection algorithms, namely Hough Line Transform algorithm and Sliding Window algorithm, are developed, tested and compared.