CI Symposium 2025: Toward Safer Southern African Roads with Camera-Only Assistive Perception скачать в хорошем качестве

CI Symposium 2025: Toward Safer Southern African Roads with Camera-Only Assistive Perception

Presentation Slides: https://koacloud.its.hawaii.edu/s/xx9... Speaker: Moseli Mots’oehli PhD Candidate, Information and Computer Science, University of Hawaiʻi at Mānoa Abstract: Autonomous perception and driving in Southern Africa are constrained by data scarcity, limited computing resources, and poor infrastructure. This talk outlines a practical path to assistive autonomy using camera-only inputs (no LiDAR): core perception (detection, drivable area, depth, lanes), rapid iteration with weak/pseudo-labels, a simple train-and-evaluate workflow, and weather/lens-artifact simulation to diversify data. I’ll share progress from the AfricanDrive Dataset for end-to-end autonomous driving, focusing on data collection, image processing, camera calibration, camera poses, depth estimation, and pseudo-labeling with state-of-the-art models, along with nuScenes-like exports that enable benchmarking. The aim is dependable assistance today that improves safety on regional roads, while laying a grounded foundation for future autonomy under Southern African constraints. Bio: Moseli Mots’oehli is a PhD candidate in Computer Science at UH Mānoa, supervised by Kyungim Baek and Huaijin Chen. His work focuses on camera-only autonomous driving in low-resource settings in Southern Africa, active learning, and learning under annotation noise. He previously earned a master’s in Data Science from the University of Pretoria.