This research project explores the feasibility of Vision-Language Models (VLMs) as an alternative to conventional object detectors for autonomous driving in rural environments. It addresses the lack of research in rural object detection and VLM applications, aiming to bridge these literature gaps. The study modifies a rural autonomous driving dataset and evaluates Faster R-CNN, RetinaNet, and SEEM models for 2D object detection. The results reveal that both Faster R-CNN and SEEM outperform RetinaNet, with SEEM showing potential improvements when incorporating textual information. The impact of object size and IoU threshold on model accuracy is highlighted, and class-specific variations in mean Average Precision (mAP) values are explored. Despite dataset limitations, the research contributes valuable insights to the field of rural autonomous driving object detection and highlights the promise of VLMs in this context. Future work includes addressing dataset imbalances and label errors for more reliable model evaluation and deployment in rural settings, enhancing object detection systems for rural autonomous driving.
There are several limitations in the projects study that must be considered when interpreting the results. The limitations of the rural dataset introduce constraints on the conclusions that can be drawn regarding model performance. The dataset exhibits class imbalance, with certain object classes being underrepresented compared to others. This imbalance may skew the evaluation metrics, potentially giving an inaccurate representation of the true models’ ability. Additionally, the presence of label errors within the dataset further complicates the evaluation of model performance. These errors lead to misannotated objects, affecting the training and evaluation process. Another limitation, is the region which the rural dataset was gathered. As rural environments can vary greatly across different regions and countries, the findings may not generalise well to other rural settings globally. Factors such as road infrastructure, vehicle types, and traffic patterns may differ, potentially affecting the models’ ability to adapt to a more diverse rural road scene