Viewpoint dependent model for multi-view object recognition

Abstract

The life expectancy of humans increases due to better medical care, food quality and personal hygiene. The demand for domestic robots performing (simple) household chores will increase as consequence to this trend. An important task of these robots is recognizing objects within an environment. Object models are constructed using multiple views from various viewpoints in order to increase the recognition robustness and reduce the influence of noise such as variation in illumination. The objective of this thesis is the construction of a novel viewpoint dependent model for multi-view object recognition. A benchmark is created in order to evaluate novel object recognition models for multi-view object recognition. The acquisition of object data is replaced by a scientific captured object dataset, which reduces the influence of external noise (i.e. illumination) and unequal view distributions. Edges, corners, shape, color and texture features are extracted, which describe the object data mathematically. Artificial noise is added for the construction of query objects, which are classified via the Euclidean distance measure. Based on sequence alignment, a novel viewpoint dependent object model is proposed and evaluated. Each view in a viewpoint dependent object is represented by a single value based on the location of that view in a kd-tree. The proposed sequence alignment algorithm matches objects via the optimal alignment of view sequences. The benchmark results show that multi-view sequence alignment has a higher object recognition rate compared to single-view object recognition. Novel object data is captured by moving a hand held Kinect camera around an object. A visual odometry method is implemented, which estimates the camera egomotion from registered depth and color. The ground plane is extracted from the depth data for the initial camera alignment and object segmentation. A viewpoint dependent object is constructed from the estimated egomotion and segmented object data. For object recognition each object is captured twice, where object variation occurs by changes in lighting and unequal distributed views. The scientific object datasets are replaced by the novel captured object datasets for evaluation of the object recognition performance. The benchmark results show an increased object recognition rate for the sequence alignment algorithm, despite the influence of external noise factors.