Image Segmentation and Region Classification in Automotive High-Resolution Radar Imagery

Image segmentation and classification of surfaces and obstacles in automotive radar imagery are the key technologies to provide valuable information for path planning in autonomous driving. As opposed to traditional radar processing, where clutter is considered as an unwanted return and should be effectively removed, autonomous driving requires full scene characterization. Hence, clutter carries necessary information for situational awareness of the autonomous platform and needs to be fully assessed to find the passable areas. In this paper, we proposed a method of automatic segmentation of automotive radar images based on two main steps: unsupervised image pre-segmentation using marker-based watershed transformation, followed by the supervised segmentation and classification of regions containing objects and surfaces based on the use of statistical distribution parameters. Several distributions were considered to characterize returns from specific region types of interest within the scene (denoted as classes) in calibrated radar imagery-the extracted distribution parameters were assessed for their ability to distinguish each class. These parameters were then used as features in a multivariate Gaussian distribution model classifier. Both the performances of the proposed supervised classification algorithm and the automatically segmented results were investigated using F1-score and Jaccard similarity coefficients, respectively.