Multi-view Detection Method for UAVs Based on Probabilistic Fusion

Human detection by a single unmanned aerial vehicle (UAV) is easily limited by the angle of view, which cannot obtain enough feature information about objects. To solve this problem, this paper proposes a multi-view detection method for UAVs based on probabilistic fusion. The method comprehensively utilizes human features detected by UAVs from different perspectives to improve the detection effect. Firstly, images are selected from the Okutama-action dataset and a human detector is trained based on Yolov8. After that, the SIFT algorithm is employed for global feature matching to obtain a perspectives transformation matrix between images from different perspectives, so that the distance between objects from different images can be measured in the same coordinate system. Then, object association is achieved using the concept of global nearest neighbor. Finally, a probability-based approach is utilized to fuse the detection results from two UAVs, and the results are compared with common fusion methods. The experimental results demonstrate that the proposed method can improve detection efficiency: The accuracy of the object association method is higher than 90%, and the fusion method can reduce missed detection.