Posters

Abstract

General summary. The proliferation of wind farms raises naturally the need for minimizing their environmental impact, and an essential part of this is the protection of wildlife. Regarding the demand for eliminating bird collisions with the blades of wind turbines, several innovative solutions have been proposed. A key ingredient of an optimal solution should be a method that estimates the distance between a wind turbine and an approaching bird. Because a robust estimation of the distance allows the operator to engage deterrence measures in time, and if necessary to stop the rotor of the wind turbine. The classical approach to distance estimation is using the principles of computer stereo vision. In this case, an algorithm compares the images of an object seen by two cameras, and from the correspondences it computes an estimation of the distance between the object and the camera pair. However, recent advances in robotics and autonomous navigation indicate that estimating distance from a single camera is also feasible. The use of a single camera is highly desirable for a wildlife protection system because it halves its complexity and results in obvious advantages. Herein we present preliminary results from our research in this field. Our method is exploiting information from a single image that is recorded by a high-resolution camera. Using principles of computer vision, the developed algorithm is able to produce distance estimations in real time. The efficiency of the algorithm allows to combine estimations in successive video frames, that further increase its accuracy. We have conducted experiments in a controlled environment, and the results demonstrate that our approach is feasible and appropriate for a bird protection solution. Method. Considering that detecting a bird in the video frame includes localization information, we extract adequate features of this region. The features are processed taking into account the intrinsic parameters of a camera that has been calibrated in advance. The output is further processed with statistical techniques, which use the previous estimations to produce a final smooth estimation of the distance. Results. We conducted experiments in a controlled environment using a drone as target. We recorded video using a 4K camera, and manually localized the drone in the frames to minimize errors. Moreover, the drone was equipped with GPS, which provided accurate distance measurements. Regarding the accuracy of the estimations, the average relative error was 20%. Conclusions. We plan to conduct more experiments, to assess the efficiency of the method in real-life conditions. But the evidence so far indicates that we approach a breakthrough in wildlife protection, by designing a system that uses half of the cameras compared to a system based on traditional stereo vision. This is an important advantage, because the cameras are a large part of the production cost, and they have also a considerable maintenance cost. Learning objectives. Our report will be most useful to wind farm operators and ornithologists that apply solutions for wildlife protection. It will disseminate the design and performance of an innovative protection system, which can estimate bird distance, and take adjustable measures for collision avoidance.