Research Project “DRONES FOR FINDING AVALACHE-BURIED” (D-FAB)

D-FAB project aims at developing a vertical takeoff and landing (VTOL) drone for assisting Alpine rescue teams in locating avalanche buried and semi-buried people, with the aim of shortening the localization time thus improving the chances of saving their lives.

The project has to solve a number of problems, relevant to different disciplines:

1. Drone design. Sturdy, rugged drone mechanical design, suitable for operation in harsh winter alpine conditions. Typical operating conditions—together with the characteristics of the payload needed for victims localization—impact on the system design at different levels, from the power storage, to the resistance against water and humidity, low temperatures, and to the ability to fly in a controllable and stable way even in presence of significant winds as it may be the case on high altitude scenarios. Also thanks to the expertise of DII members in designing and realizing autonomous vehicles and control systems (also considering aerodynamics aspects)[3–6], and in manufacturing structural components in composite materials, it will be realized a prototype of a lightweight, resistant, and foldable VTOL drone structure—which is not yet available as off-the-shelf solution.
2. Scenario and ego-state reconstruction, by applying sensor fusion techniques to a set of lower level measurements, including GPS location, inertial navigation, height-over-ground detection, apparent wind direction and intensity, ground morphology. More precisely, this will be addressed by developing algorithms that exploit computer vision and image processing techniques to deal with the huge quantity of information that UAV collects (possibly offloaded to a ground station for elaboration) [22,23]. It will consist in the design of a nonlinear filter, which combines image gradient features and Gaussian process (GP) modeling. The image gradient features allow capturing detailed information regarding the structure of the investigated classes of objects (human figures, semi-buried personal items) whereas the GP model fed with image gradient features permits to yield a statistical estimate of the presence of objects of interest for any position within the image. The image/video acquisitions will be performed from predetermined altitude very close to the ground. Accordingly, the GP model will be trained with a library of predefined objects with adequate spatial resolutions. In order to speed up the algorithm, a fast and simple screening technique based on application of adaptive image threshold will be adopted in order to reduce the areas to be analyzed by the detector.
3. Localization of buried and semi-buried avalanche victims, possibly using standard ARVA beacons together with image analysis and classification, by using the very approach described in B and coherently fused with the other sensed data, for incrementally updating a map of the avalanche area, marking the presence of beacons, items (skis, backpacks), and semi-buried human figures. Regarding beacon signal analysis, there are still few examples of automatic searching. One of the key problems arises from the uncertainty about the transmitter orientation. In fact, the direction of the transmitter antenna changes radically the field shape. From a general point of view, with respect to classical far–field identification problem, in this case we have to identify 6 states for each transmitter, instead of 3, while we can only collect 3 measurements (the H-field vector components). Amongst other solutions based on complex antenna arrangement and circuitry, it is worth recalling the solution proposed in [7, 8]—based upon Bayesian estimation theory and Kalman filters—which is going to be adopted in the project, much likely to prove compatible with the constraints for a drone on-board system.
4. Navigation. The operating scenario presents harsh conditions and stringent time constraints. Consequently, in our envisioned application, the operators set goals and constraints for the mission and define search areas for the drone. The drone executes the mission autonomously and whenever a meaningful event takes place (e.g., localization of a victim or completion of the scanning mission on the assigned area), it releases markers to pinpoint locations of interest and send notification to the rescue team. During the mission execution operators are relieved of the task of guiding the drone and can shift their attention to other goals (e.g., setting up a protected area for emergency care or dig out a victim that the drone has already located). Mission planning and control in a partially unknown environment with mission critical goals is a difficult task, for which robust and reliable solution are not yet available. The problem is hybrid in nature since the continuous dynamics of the drone is combined to a “logical” formulation of constraints and goals [20] but the hostile environment conditions and the need for flexible response to unanticipated events, requires robust solutions such as can be offered by probabilistic approaches [21] for the specification of the mission and for the search of the solution space. As far as the navigation is concerned, the drone is associated with a set of basic maneuvers (symbols), each one implemented by a suitable control law and associated with a change in the state of the system (e.g., a change in the coordinates). Navigation is implemented by a sequence of symbols decided by the planning algorithms and executed by the control algorithm. The correct sequence is identified by an optimization framework that breaks the high level goals down into a sequence of elementary changes each one implemented by a control symbol. The algorithm will fully exploit the information coming from Task 4 (environmental condition monitoring, surface reconstruction, localization, presence of beacons) and the detected environmental features to plan optimal routes for patrolling the area of interest. According to the stochastic models based on the precious experience of the Alpine Rescue Service as well as on the changing weather conditions (e.g., the presence of sudden wind gusts), the goals and the changes in the state by each symbol have a probabilistic nature. So, the mission planner decides the course of actions that maximizes the probability of fulfilling the goals and meeting the constraints. A mission supervisor monitors the execution of the mission and triggers alternative plans if significant deviations are detected with respect to the planned progress.
5. Human-Machine Interface. Most of the operational efficacy of our solution lies in the ability for the human operators to easily and effectively and securely define the mission objectives and constraints, and to receive readable and unambiguous notifications of the meaningful events. The harshness of the operational scenario (irregular illumination, presence snow and gales, heavy garments and glass that obstruct the visibility of screen) can put a strain on the most readable interfaces if these factors are not adequately accounted for. What is more, the use of gloves and the low temperature of the fingertips can seriously reduce the precision and the efficacy of using the touch screen. These considerations discourage the straightforward utilization of existing smart phone and tablet apps that have become the natural companion of low cost commercial drones. A close cooperation with a sample of potential users belonging to the rescue teams will allow us to the requirements of the interface, which will be fine-tuned with field tests on potential operational scenarios.
6. Legal aspects: concerning both legal standards for drones operability and civil and criminal liability rules for controlling and preventing human-induced avalanches. As for the first subtopic, operating a drone poses significant and unprecedented legal issues, considering that the first attempt in regulating drone flight is extremely recent (the first Italian regulation of this new aerial technology was issued by ENAC in 2013 and the UE is currently preparing a uniform discipline across member states –http://europa.eu/rapid/press-release_IP-14-384_en.htm). More specifically, the issues to be considered are: role of the civil aviation authority over drone operations; safety and manufacturing standards as well as product liability rules; rules of civil liability applicable to drone-induced damages in the event of mid-air collisions or surface crushes; privacy issues, since data collected by UAVs must comply with the applicable data protection rules, and data protection authorities must monitor the subsequent collection and processing of personal data. As regards the second subtopic, a vast debate flourishes in legal literature today on the civil, criminal and administrative rules aimed at controlling the risk of avalanches, since the diffusion of the practice of ski ride and extra slopes skiing (also encouraged by winter tourism marketing strategies) increases the statistical occurrence of avalanches and calls for the application of more severe rules to skiers engaged in such practice, also inducing a serious concern for the social allocation of the costs associated to the rescuing of avalanche-buried skiers. This debate needs to be analyzed and reconstructed in order to provide the research project with an in-depth analysis of the law and economics of avalanche prevention, which is the field targeted by the innovative product at the core of the research project.

Clearly the operational efficiency of the proposed system strongly depends on the integration of these interdisciplinary aspects, as it is typical for mechatronics devices. For these reasons, the team proposing this research has been selected in order to provide a set of competences encompassing all the problems A.–F.