This thesis proposes an approach to a method, used by an Unmanned Surface Vehicle (USV) to detect obstacles, that are outside the range of onboard stereoscopic camera pair. Algorithm is based on a concept of Structure from motion using two-view geometry. It also utilizes data from other available sensors, such that are used to assess the camera pose. The proposed approach relies on finding suitable pairs in the data buffer, that contains various data of previously captured images. Search of image pairs is reduced by implementing k-means clustering with respect to other geometric constraints on camera poses. Algorithm is building a growing point cloud, which is available to obstacle detection and path planning procedures. We have conducted experiments, which evaluate the detector repeatability and determine the effect of detection uncertainty on point cloud reconstruction.