Pose estimation from 2D borescope inspection videos via structure from motion

A method for generating a pose estimation for an object from a 2D borescope inspection video, the method comprising:

• generating a model 3D point cloud of a predetermined area of an object using a Computer Aided Design (CAD) assembly model of the object;
• extracting a video frame sequence from a 2D borescope inspection video of the object;
• generating an estimated 3D point cloud by processing the video frame sequence using a Structure-from-Motion (SfM) algorithm;
• identifying a common coordinate system with respect to the model 3D point cloud and the estimated 3D point cloud;
• and
• computing a rough alignment by processing the common coordinate system, the model 3D point cloud and the estimated 3D point cloud using a global registration algorithm;
• and
• after generating the model 3D point cloud from the CAD assembly model and the estimated 3D point cloud from the video frame sequence, generating a fine registration pose estimation by processing the rough alignment of the object.

A computer-implemented method for generating a pose estimation for an object from a 2D borescope inspection video, comprising:

• generating a Computer Aided Design (CAD) assembly model of an object;
• generating a model 3D point cloud of the object using the CAD assembly model;
• generating a 2D borescope inspection video of the object, wherein the 2D borescope inspection video includes a plurality of video frames;
• extracting a video frame sequence from the plurality of video frames;
• generating a coarse estimated point cloud video frame sequence by applying a Structure-from-Motion (SfM) algorithm to the video frame sequence;
• generating an estimated 3D point cloud by applying a filtering algorithm to the coarse estimated point cloud video frame sequence to filter out outliers;
• identifying a common coordinate system with respect to the model 3D point cloud and the estimated 3D point cloud;
• computing a rough alignment using a global registration algorithm;
• and
• after generating the model 3D point cloud from the CAD assembly model and the estimated 3D point cloud from the coarse estimated point cloud video frame sequence, generating a fine registration pose estimation of the object by processing the rough alignment.