Self-calibration of binocular camera extrinsic parameters based on structured feature points selection

The extraction of feature points is crucial to computer vision tasks like self-calibration of binocular camera extrinsic parameters, pose estimation and structure from motion (SFM). In the context of autonomous driving, there are numerous unstructured feature points, as well as structured feature points with shapes such as L-type, Y-type, Star-type and centroid. Typically, feature points are extracted without discrimination and used as inputs for feature-based visual algorithms in a generalized manner. However, the influence of the structural characteristics of these feature points on the performance of such algorithms remains largely unexplored. To address this issue, we propose a multi-stream feature point classification network based on circular patches extraction (CPE). CPE uses concentric circles centered on a given feature point to extract the intensity distribution features around that point. Subsequently, a series of circular patches are converted into square patches according to the order of radius and polar angle. Then, we have a multi-stream feature point classification network, where each stream receives a square patch as input to learn the intensity distribution features and classify the feature points into Y-type, centroid and unstructured categories. Finally, the influence of points with structure and without structure on related autonomous driving visual algorithms was verified in the experiment. Experimental results indicate that our proposed network can effectively classify based on the structure of feature points, which can enhance the performance of feature-based vision algorithms.