Exploiting a Prior 3D Map for Object Recognition

Recognizing objects in the environment and precisely determining their positions is a fundamental component of autonomous navigation systems. This thesis presents a technique for determining both the locations and the semantic labels of new objects in a scene with respect to a prior three-dimensional (3D) map of the scene. This work aims to reduce object recognition errors in cluttered environments by isolating new objects from the known background by correlating features detected in a new photo with feature points that constitute the 3D map. Such isolation enables a neural network trained to recognize an enumerated set of objects to focus narrowly upon those portions of images that contain new objects instead of having to process the whole scene. As a result, changes in a prior map can be rapidly detected and semantically labeled, allowing for confident navigation within the ever-evolving cluttered environment. Using multiple images obtained from varying camera poses, the globally-referenced 3D positions of changes in the scene can be determined with multiple-view geometry techniques.