Search

Tactical reconnaissance using small unmanned aerial vehicles has become a common military scenario. However, since their sensor systems are usually limited to rudimentary visual or thermal imaging, the detection of camouflaged objects can be a particularly hard challenge. With respect to SWaP-C criteria, multispectral sensors represent a promising solution to increase the spectral information that could lead to unveiling camouflage. Therefore, this paper investigates and evaluates the applicability of four well-known hyperspectral anomaly detection methods (RX, LRX, CRD, and AED) and a method developed by the authors called local point density (LPD) for near real-time camouflage detection in multispectral imagery based on a specially created dataset. Results show that all targets in the dataset could successfully be detected with an AUC greater than 0.9 by multiple methods, with some methods even reaching an AUC relatively close to 1.0 for certain targets. Yet, great variations in detection performance over all targets and methods were observed. The dataset was additionally enhanced by multiple vegetation indices (BNDVI, GNDVI, and NDRE), which resulted in generally higher detection performances of all methods. Overall, the results demonstrated the general applicability of the hyperspectral anomaly detection methods for camouflage detection in multispectral imagery.

The expected increase of unmanned aerial vehicles (UAVs) in civil airspace is driving research activities in the field of UAV Sense and Avoid. In terms of the application of unmanned air cargo transport, new multifunctional on-board systems can be envisaged, which allow to avoid both, other air traffic as well as hazardous weather areas. In order to prevent icing and turbulence, such a weather function should be able to detect dangerous weather phenomena. In addition, the weather function could offer the possibility to operate UAVs in compliance with visual flight rules (VFR) by observing legal minimum cloud distances. In this paper we propose a combination of AESA (Active Electronically Scanned Array) based weather radar and electro optical-cloud detection onboard UAV in order to increase the weather detection capabilities of both sensor systems. Besides the individual systems, the combination of both sensor systems and their advantages are described in detail. Furthermore, application scenarios are defined and shown using sample images from test flight campaigns.

We propose an approach to enhance the overall detection performance in unmanned aerial reconnaissance with imaging sensors. This is achieved by generating near global optimal UAV trajectories utilizing previously developed sensor performance models. Here, we are taking into account UAV, sensor and environment characteristics as well as mission objectives. The approach can be briefly summarized as follows: First, we apply coverage path planning to determine the sensor footprint path within a predefined recon area. Second, the sensor performance model is utilized to calculate discretized two-dimensional perception maps for each footprint, which is a state-dependent representation of the detection performance. For each perception map, we identify suitable sensor positions of high detection performance. From here, we systematically connect these positions of consecutive perception maps to create a transition graph. With use of optimal control, we assign costs (e.g. recon duration) to these connections while taking several requirements and constraints into account. Dynamic programming is utilized to determine the optimal path in the graph with the lowest cost. Finally, we smooth the path by cubic spline motion primitives, which leads to the reference flight trajectory for the UAV. We evaluate our approach through a series of simulated vehicle detection reconnaissance tasks with a fixed wing UAV. Our approach improves the detection performance by 3% compared to a trajectory generated with a local optimal greedy optimization and the reconnaissance time is reduced by 20%. In comparison with a primitive trajectory, the improvement in detection performance is even up to 10% with an associated increase in reconnaissance time of at least 15%.

In this paper, we consider the task of simultaneously tracking several spatially distributed ground objects (i.e., vehicles, or animals) with a single UAV equipped with an imaging sensor. The problem regarded is to actively schedule the sequence of single target observations in such a way that the kinematic state (position and velocity) of the objects to be tracked can be estimated with minimal error. The main difficulty in this task is that there is not one single optimal scheduling policy that is appropriate for all geospatial relations between the UAV and the target objects, as the quality of the target localization depends on the relative position of the object to the sensor. While many established scheduling and planning algorithms can generally be applied to this problem, their suitability varies depending on the specifics of the tracking scenario. In order to track multiple objects with only one sensor in changing situations, a trade-off has to be found between a) collecting sensor data in the right place and at the right time to achieve the lowest estimated error of a single object, and b) not focusing perception to a single object and thus neglecting observations of the others. Even if additional environment descriptors are neglected, relevant literature has proven scheduling problems akin to the one at hand to be NP-hard. We identify criteria according to which a scheduling policy can be selected and conclude with an evaluation of the procedure in simulation experiments.

This paper presents an approach for image-based detection of clouds on-board unmanned aerial vehicles (UAVs), taking into account changing environmental conditions and different sensor types. Therefore, different image processing methods for cloud detection are introduced. In order to separate manned and unmanned aircraft in airspaces where air traffic is expected to operate according to visual flight rules (VFR), the visibility of unmanned platforms must be ensured. Besides reduced visibility, clouds can additionally indicate an increased risk of icing and turbulence. Thus, entry into clouds must be prevented and the required minimum distances to clouds have to be complied with. In order to initiate appropriate evasive maneuvers, the cloud detection component as part of a Sense and Avoid system must deliver robust results. Our previous investigations on cloud detection algorithms have shown that the quality of the results depends very much on environmental conditions like lighting conditions, visibility, cloud type or ground texture. The requirement to deliver reliable detection results under as many weather scenarios as possible is therefore difficult to fulfill. This goal shall be achieved by using different sensor configurations and a performance-based selection of cloud detection algorithms. Apart from introducing this concept of adaptive and context-sensitive cloud detection, this paper gives an overview of complementing cloud detection methods that can be used for this purpose.

To assess the mission effectiveness and efficiency of future airborne systems-of-systems with autonomous components, appropriate performance models and metrics are required. In a first attempt, such metrics were derived systematically by subject-matter experts (SMEs) decomposing a specific mission and its tasks hierarchically, until measurable criteria were reached. Weights of the criteria were obtained through the Fuzzy Analytic Hierarchy Process (FAHP), using linguistic variables for the pairwise comparison of criteria on all decomposition levels.

Visual communication, and especially that based on gestures, can be an alternative to radio-based command and control links to small unmanned aerial vehicles (UAVs). In order to allow this natural form of communication to be an equivalent substitute to a conventional data link with regard to the efficient transfer of task data, the human-UAV-interaction must consist of as few gestural parameters as possible to keep the mental load for the user low, but at the same time provide enough information for the UAV to estimate the users intention. For this purpose, an experimental study under laboratory conditions with N = 20 German-speaking participants has been conducted to generate a dataset with 120 verbal formulations for six common UAV related tasks. The collected data has been evaluated regarding the distribution of sentences, words and parts of speech. Results indicate that users prefer short task descriptions prioritizing verbs and nouns (84.15%) with a mean number of 2.55 (SD = 1.57) words per sentence. All tasks considered were described with an average of 1.73 sentences (SD = 1.00). A slot filling approach is presented to map parameters in user intents.

Aerial surveillance systems have become an important aspect in a range of civilian and military applications. For instance, reliable location data of individuals or objects is essential for use cases such as traffic surveillance, search and rescue, relief efforts after natural disasters, or patrol and border control missions. Unmanned aerial vehicles (UAVs) provide great value in such situations but conventionally require at least one human operator whose insights and intuition are necessary for performing the task. With this contribution, we conceptualize a functional design that leverages performance models to track multiple targets and generate situational awareness autonomously. Further, we devise a testing harness for different scheduling schemes and provide some preliminary experimental data on basic scheduling policies found in literature.

Vehicle detection on aerial imagery plays an important role in autonomous systems especially on board of UAVs (Unmanned Aerial Vehicles). Deep learning based object detectors are often used to overcome the resulting detection challenges. To achieve the necessary variance, real training data are extended or completely replaced by synthetic data sets from virtual simulation environments. Differences between these two domains can lead to performance differences in the later real application, which are called reality gap. Our current research interests are focused on the identification of image properties that on the one hand allow a distinction between real and synthetic domains and on the other hand are related to detection performance. We present a method to rank the most relevant image descriptors using a machine learning based classification chain. In the first part of the investigations real and synthetic image pairs from own UAV flights and from the publicly available KITTI/Virtual KITTI data set will be analyzed. Finally, it will be shown which image descriptors are responsible for the false detections of a purely synthetic trained YOLOv3 vehicle detector and which conclusions can be drawn for future training data generation.

Sense and Avoid is gaining importance in the integration of unmanned aerial vehicles (UAVs) into civil airspace. In order to be seen by other air traffic participants operating under visual flight rules (VFR), own visibility must be granted. This includes the observance of legal minimum cloud distances. An airborne electro-optical cloud detection and avoidance system could cope with this task on unmanned platforms. In this paper we propose an approach for cloud distance estimation using computer vision algorithms. In addition to the functional architecture, results from a simulation environment are shown and the demands for future developments are addressed.