Semi-Supervised Active Learning for Radar based Object Classification Using Track Consistency

Abstract

Development of machine learning (ML) models requires large amounts of labeled data. For safety critical automotive applications such as radar based perception, the dataset must contain various and rare corner cases, e.g. rare instances that have not been seen before. The straightforward approach of measuring and manually labeling large amounts of data to capture such corner cases is often infeasible or impractical. Thus, approaches for efficiently selecting and labeling the relevant data are essential for ML-based radar applications. In this paper, we propose a method for semi-supervised learning (SSL) for radar object type classification. We use the track consistency of tracked radar objects as a constraint to generate high-quality labels for the vast portions of the unlabeled dataset. We extend the proposed SSL approach with active learning that considers the data relevance, such that the most relevant data with the least accurate auto-labels are selected for human labeling. We show that the proposed approach achieves a saving of more than 87% of human labeling costs based on auto-labeling and relevant data selection.