Object recognition for SLAM in floor environments using a depth sensor

Depth sensors have been increasingly used for object recognition in recent years. However, it is very challenging for simultaneous localization and mapping (SLAM) to make use of the forward scenes from a depth sensor. To this end, we introduce the object recognition framework for SLAM in indoor environments based on the extraction of an object-level descriptor. The proposed object-level descriptor can be obtained based on the surface appearances acquired from a depth sensor without any training. To express the surface normal distribution, a well-known descriptor, fast point feature histogram (FPFH), with a small sampling radius is used to define basic shape elements of a plane, a cylinder and a sphere. The object-level descriptor to recognize the objects can be obtained using these shape elements. Several experiments on arbitrary objects on the floor show the proposed scheme is useful in object recognition and generation of the feature map.