Object point cloud comparison and matching

#U.S. Pat. No. 9,417,323, issued Aug. 16, 2016 is incorporated herein by reference.

#The present invention relates to comparing and matching point clouds for two or more objects.

#The advent of LIDAR and other methods of generating remotely-sensed 3D point representations of scenes and objects necessitates techniques for exploitation algorithms. The large data problem and unique aspects of how 3D point clouds are generated require new ways of investigation involving computational speed and size reduction without loss of information (or at least with an acceptable loss depending on the resolution required). Currently, most point cloud comparison/matching algorithms are computer vision-based approaches. These techniques work on object surface features.

#FIG. 1 is a flow diagram illustrating the steps performed in object point cloud matching according to the present invention.

#FIG. 2 is a plot illustrating the point cloud of an object.

#FIG. 3 is a diagram showing the step of voxelizing the point cloud of FIG. 2.

#FIG. 1 is a flow diagram illustrating the steps performed in object point cloud (PC) matching according to an embodiment of the present invention. At a high level the process is a way of comparing two point clouds. The process is a tool and may also be used for comparison with a set of exemplars resulting in a “best of class” and a match if the data has a high enough SNR and the quantization level for the desired differences is known. Referring to FIG. 1, in step (a) a first point cloud representing a first object is chosen. In step (b) a second point cloud representing a second object is chosen. These two objects (or in many cases more than two objects) will be compared to see how well they match. In step (c), the two point clouds are azimuth (angle) aligned.

#In step (d) the point clouds are quantized at a high level in preparation for translation alignment. High level quantization means sufficient to capture the translation accurately enough to accommodate the smallest anticipated voxel size for comparison (though the actual translation may be done in analog). In step (e) the point clouds are translation aligned.

#In step (f), each point cloud is quantized into logical arrays having differing resolutions. This may start with the high-level quantization done in step (g), the point clouds for the two objects are compared at each resolution. This is because the best resolution to determine how well the objects match varies a great deal according to how complicated the objects are and at what level of feature the match is desired (e.g. vehicle body style versus rearview mirror). Comparison results are determined at each resolution. Generally, one of the resolution levels is determined to be the best (or at least sufficient) and that comparison result is used to decide how closely the two objects match.