This study addresses the limitations of existing autonomous vehicle datasets, particularly the need for greater specificity of weather conditions and road types. We utilized X-CAR to highlight the challenges of extreme weather and non-urban road conditions on autonomous driving systems. Our dataset comprises recordings under seven distinct weather and lighting conditions across four road types. Notably, our research focuses on differentiating between various lighting and weather conditions and road types, which often need improvement in many existing datasets.
We used the X-CAR platform to collect 360-degree image information and LiDAR point clouds at 10Hz. Due to the constraints of time and resources, we used algorithmic prediction to generate ground truth data via the Co-DETR 2D prediction algorithm. We validated the accuracy of the Co-DETR algorithm through partial manual annotation. However, it is undeniable that in some extreme conditions, the algorithm-generated ground truth can lead to results deviating from expectations and real-world situations. Therefore, we conducted a scaled manual annotation and controlled experiments, ensuring the highest level of accuracy.
After the manual annotation, we validated our initial conclusions and trained a model based on YOLOv8x, focusing on weak environmental conditions. The final model underwent multiple iterations and achieved satisfactory accuracy. The enhanced model demonstrated a significant increase in detection accuracy compared to the original YOLOv8x model. At the same time, our analysis identifies weather conditions that markedly reduce detection accuracy, providing focal points for future dataset enhancements. / Master of Science / This study explores the limitations of current autonomous vehicle datasets, particularly their lack of detail regarding weather conditions and road types. We used X-CAR to examine how extreme weather and light conditions affect autonomous driving systems. Our dataset includes recordings from seven different weather and lighting conditions across four types of roads. Due to time and resource constraints, we used an algorithm to predict ground truth data with the help of Co-DETR. While not all data was fully annotated, we manually labeled part of the data to create an actual ground truth. This allowed us to verify our previous findings and train a model based on YOLOv8x, focusing on challenging conditions. The improved model showed much higher accuracy in detecting objects than the original YOLOv8x model. This study highlights the significant impact of weather conditions on detection accuracy and suggests areas for future improvements in datasets.
| Identifer | oai:union.ndltd.org:VTETD/oai:vtechworks.lib.vt.edu:10919/120687 |
| Date | 24 July 2024 |
| Creators | Du, Xuelai |
| Contributors | Mechanical Engineering, Eskandarian, Azim, Sandu, Corina, Taheri, Saied |
| Publisher | Virginia Tech |
| Source Sets | Virginia Tech Theses and Dissertation |
| Language | English |
| Detected Language | English |
| Type | Thesis |
| Format | ETD, application/pdf |
| Rights | In Copyright, http://rightsstatements.org/vocab/InC/1.0/ |
Page generated in 0.0023 seconds