Active-Learning Method: An Effective Way to Generate Ground Truth Data to Test & Validate ADAS Function Development

Machine learning exerts a significant influence on the autonomous driving industry, enabling the development of self-driving vehicles. However, the performance of these models heavily relies on the quality and diversity of the training data. In situations where valid data is scarce, models struggle to make informed decisions. To address this obstacle, active learning methodologies are utilized for the purpose of choosing the most informative data frames from an extensive reservoir of unlabeled data. Uncertainty scores are calculated using methods like Least Confident and Entropy-based sampling, and frames with high uncertainty scores are manually annotated and added to the training dataset. This iterative process improves the model’s performance over time. This research study centers on assessing the application of active learning in diminishing the manual labor needed for data labeling through the selection of frames containing valuable information. A pre-trained YOLOv3 model is utilized to calculate confidence values for detected objects in each frame. The frames with the highest informativeness measures, determined through Least Confident and Entropy-based sampling strategies, are chosen for manual labeling. The labeled data is then used to train YOLOv5 and YOLOv8 models for label prediction. This study achieves an accuracy of 80% in 10 iterations using the Least Confident sampling strategy, and an accuracy of 73% in 10 iterations using the Entropy-based sampling strategy for YOLOv5 models. For YOLOv8, an accuracy achieved is 57% in 6 iterations using the Least Confident sampling strategy, and an accuracy of 52% in 6 iterations using the Entropy-based sampling strategy. The paper concludes by highlighting open problems in the field and suggesting future research directions.