Labeling Data and Neural Network Training for Drone Detection Task

In this paper, a video data labelling algorithm is proposed for the detection task. In order to test the algorithm, a dataset with the drone image was collected. To collect the dataset, 23 sorties were made in different conditions at the same location. The total duration of the video is 1 hour 20 minutes 3.6 seconds, of which 1 hour of the video contained a drone. On the crowdsourcing platform, 18025 frames were manually labelled. A total of 178687 unique drone frames were obtained including inter-frame interpolation. Label fault was 3410 frames (1.91%), the total collected dataset contains 175277 good frames. Several neural networks were trained to test the dataset: RTMDet-s, YoloX-s. The collected dataset was divided into two parts: training and testing. There are 168713 images in the training sample and 7982 images in the test sample. The YoloX-s architecture showed the best result: 0.9520 $m A P_{50}$. The RTMDet-s architecture, which has the best quality on the COCO dataset, performed significantly worse: 0.7837 $m A P_{50}$. Experiments have shown that inter-frame interpolation degrades the quality of the trained neural network.