Machine Learning Based Specular Highlight Detection Techniques To Enhanced UAV Indoor Navigation Through Slam
Date of Award
Master of Science (MS)
Unmanned aerial vehicle technology is advancing to the point where it has a lot of potential for use in various applications. While global navigation satellite systems (GPS) enable autonomous flight in outdoor contexts, dedicated, accurate wireless localization is a prospective contender for enabling indoor navigation and associated applications. This thesis represents ongoing research in Simultaneous Localization and Mapping (SLAM) to improve mapping solely through a camera to remove map noises resulting from glass reflections. This research aims to investigate Machine Learning (ML) approaches that detect glass specularity in direct and two-step approaches. We used YOLOv4, YOLOv4-tiny, and Deeplabv3 to detect glass. Furthermore, we developed a specular detection CV2-based technique to detect specular highlights in images. We trained an image segmentation and classification models to detect glass and fed the weight into our specular highlight detection pipeline to detect specular highlight only on the area labeled by models as glass. We also trained a YOLO model to detect specular highlights directly. We trained our system with various parameters such as image size, learning rate, and the best IOU results for the specular detection directly with YOLO 71.3%. In two-stage detection, the best IOU of training with glass detection and using the specular detection pipeline to detect the specularity illustrated at YOLO tiny at 70.3%. This result improved by repeating the same steps with Deeplab-v3 to detect the glass and then entering the results into the specular highlight detection pipeline and receiving an IOU of 72.3%.
Kardan, Ramtin, "Machine Learning Based Specular Highlight Detection Techniques To Enhanced UAV Indoor Navigation Through Slam" (2022). Theses and Dissertations. 4268.
Available for download on Friday, June 14, 2024