Semantic Segmentation for Autonomous Student Formula Race Track Localization
Semantická segmentace pro nalezení trati závodu autonomní studentské formule
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
Tato práce se zabývá metodou pro lokalizaci vnitřku trati pro autonomní studentskou formuli. Trať je ohraničena dopravními kužely. Navrženy byly tři různé přístupy pro lokalizaci trati: Segmentace trati se segmentační maskou, predikce hranic trati s přímou regresí a predikce hranic trati pomocí heatmap. Všechny tyto metody využivají konvolučních neuronových sítí. Pro trénování sítí byl použit námi sesbíraný anotovaný dataset. Kvantitativně jsme vyhodnotili přesnost modelů a porovnali je se stávajícím přístupem. Stávající přístup nejprve detekuje dopravní kužely YOLO detektorem a následně používá heuristický algoritmus pro nalezení trati. V práci ukazujeme, že lokalizace vnitřku trati za pomocí segmentačních masek vytvořených UNet modelem často překonává baseline metodu na komplikovaných závodních tratích.
The thesis presents a race track visual localization method for Autonomous Student Formula. The track is delineated by traffic cones. We propose three different approaches for track localization: Segmenting the race track with a segmentation mask, predicting the boundaries of the race track with direct regression, and predicting the boundaries with heatmap regression. All of these approaches uti lize convolutional neural networks. The annotated dataset used for training was collected specially for this problem. We quantitatively evaluated the accuracy of the models and compared them to the baseline approach. The baseline first detects the traffic cones by YOLO detector and then uses a heuristic algorithm to find the track. We show that localizing the race track with a segmentation mask produced by the UNet model achieves accuracy 0.93 IoU. The segmentation with our UNet model often outperforms the baseline in complicated tracks.
The thesis presents a race track visual localization method for Autonomous Student Formula. The track is delineated by traffic cones. We propose three different approaches for track localization: Segmenting the race track with a segmentation mask, predicting the boundaries of the race track with direct regression, and predicting the boundaries with heatmap regression. All of these approaches uti lize convolutional neural networks. The annotated dataset used for training was collected specially for this problem. We quantitatively evaluated the accuracy of the models and compared them to the baseline approach. The baseline first detects the traffic cones by YOLO detector and then uses a heuristic algorithm to find the track. We show that localizing the race track with a segmentation mask produced by the UNet model achieves accuracy 0.93 IoU. The segmentation with our UNet model often outperforms the baseline in complicated tracks.