A Multiscale-Metric-Learning Model for Obstacle Classification in Autonomous Driving

Research on vision-based obstacle classification models is crucial for advancing autonomous driving technology. High-precision obstacle classification models empower vehicles to avoid obstacles effectively, enhancing both safety and robustness. However, existing classification models typically rely on a single deep learning architecture for feature extraction, limiting their ability to capture multiscale information within images. Features extracted from a single scale fail to contain the full range of information required to classify obstacles of varying sizes accurately. As a result, these models often demonstrate reduced robustness when confronted with obstacles of diverse sizes. Moreover, previous deep-learning-based models for obstacle classification rely heavily on labeled data for supervised learning, ignoring the underlying similarities across samples from different categories. This oversight can impair the model’s generalization performance on new data. To address these limitations, we propose a multiscale-metric-learning (MSML) model for obstacle classification in autonomous driving. The MSML model combines principles from both metric learning and multiscale feature fusion. By incorporating multiscale features, the MSML model captures richer contextual information within obstacle images, thereby improving its overall interpretability and generalization capability. Metric learning further refines the model by organizing the feature space so that samples from the same category cluster closely together, while those from different categories are separated. This approach enhances the model’s ability to distinguish different obstacle types, improving better robustness against noise and improved generalization to new data. Experimental results on real-world datasets reveal the MSML model’s strong performance, achieving Recall, Accuracy, and F1-Score values of 0.988, 0.988, and 0.988, respectively.