Automatické vyhodnocení přesnosti přistání pilota padákového kluzáku

Abstract

This bachelor’s thesis focuses on the design and implementation of an extension to a system for au- tomatic evaluation of paraglider pilot landing accuracy. The existing system uses camera recordings of the landing area; however, determining the frame of the first ground contact and the correspond- ing contact point is time-consuming and relies on manual evaluation. The aim of this thesis was to design a method that automates these steps using computer vision and deep learning techniques. The thesis first analyzes the requirements for the camera recordings and formulates landing moment detection as a precise short-event detection problem in video. For temporal landing moment detec- tion, a model inspired by the E2E-Spot architecture was proposed. The model processes sequences of cropped images of the paraglider and predicts the landing probability for each frame. In the experimental evaluation, variants using the Gate Shift Module and a local temporal block based on 1D convolution were compared. The best result on the test set was achieved by the E2E-Spot variant with the GSM module, reaching an average mAP0,1,2,4 of 81.42 %. The second part of the proposed solution addresses contact point localization in the image. A YOLOv8 model was fine-tuned to identify the foot in contact with the ground by distinguishing between the touch and no_touch classes. The precise contact position was then estimated using heatmap regression with a lightweight fully convolutional network called SimpleUNet. The final contact point is selected as the local maximum of the predicted heatmap inside the bounding box classified as touch. Experiments on real data show that the proposed approach is suitable for automated landing moment detection and subsequent contact point localization in image coordinates. The main limitations were also identified, including the limited dataset size, the influence of occlusions, and the need for future evaluation of the complete pipeline on long untrimmed video recordings.