Skeleton-based Object Representation And Recognition In Images

The manufacture of artificial intelligence robots has been a dream of scientists for a long time. An essential step to achieve this goal is recognizing objects in images. The objec-t's skeleton contains not only the shape feature but also the topological structure, therefore it is a useful and essential descriptor for object recognition and many related applications, such as content-based image retrieval systems, character recognition systems, analysis of biomed-ical images and surveillance systems. In recent years, due to ease of capturing the articulated objects, the skeleton is widely applied to model human pose motion for human computer interaction and somatosensory gaming. Consequently, the research on skeleton-based object representation and recognition will lead to the rapid progress of computer vision.The works of this dissertation concentrate on addressing the skeleton-based recogni-tion tasks, including skeleton pruning, shape classification and clustering and human pose correction. The main contributions of this dissertation are summarized as follow:1. A novel significance measure, called bending potential ratio (BPR), is proposed for skeleton pruning, in which the pruning decision for a skeletal branch is determined by the contribution of its corresponding contour segment to the overall shape. Such the contribution depends on the particular location of the segment within the whole contour based on the fact that a segment may be considered to be insignificant in one place on the contour while it may be considered as feature elsewhere. The BPR measurement is integrated in a skeleton growing scheme to obtain the pruned skeleton, which meanwhile ensures the connectivity of the skeleton. Our experi-ments demonstrate that the proposed algorithm can remove the redundant skeleton branches and generate accurate skeletons which are useful for shape matching.2. A novel skeleton pruning approach is proposed, which differs from the traditional ones fundamentally. It casts skeleton pruning as a tradeoff between skeleton sim-plicity and shape reconstruction error, formulated within a Bayesian framework. Shape reconstruction error is measured as the area overlap between the reconstruct-ed and original shapes. Skeleton simplicity is measured to be inversely related to skeleton length. A simple greedy algorithm is applied to approximate the maxi-mum of the Baycsian posterior probability which defines an order for iteratively removing the end branches to obtain the pruned skeleton. Presented experimental results obtained without any parameter tuning clearly demonstrate that the resulting skeletons are stable to boundary deformations and intra class shape variability. 3. A skeleton-based approach to shape classification is proposed. It models a shape class by a supervised learned tree-union. Each node in the tree-union stores not only the examples of skeletal junctions but also their statistic distributions. These information are then used to classify new shapes according to Bayesian rule. The classification accuracies on two well known shape data sets are higher than the state-of-the-art approach which shows the effectiveness of the proposed approach.4. A skeleton-based approach is proposed to address the problem of shape clustering by discovering the common structure which captures the intrinsic structural infor-mation of shapes belonging to the same cluster. Unlike the traditional clustering algorithms only consider the pairwise similarity, the proposed approach adopts ag-glomerative hierarchical framework, in which the common structure is updated dur-ing clustering and used to improve the similarity between shapes in next merging iteration. The experimental results show that the proposed approach can discover the common structure of shapes of the same cluster, detect the outlier node automat-ically, be not sensitive to parameter setting, and achieve the best clustering results on four shape data sets.5. A new algorithm for pose correction from the initially estimated skeletons from Kinect depth images is present. It shows that exemplar-based approach serves a promising direction for pose correction and learning the inhomogeneous systemat-ic bias by random forest regression is the essential key. Cascaded regression and motion consistency are also applied to improve pose correction. The experimental results show that the proposed pose correction algorithm indeed significantly im-proves the accuracy of pose recognition and is much better than the one employed in the Kinect system.All the problems addressed in this dissertation are fundamental in computer vision, so other vision tasks and applications can benefit from the models and the algorithms proposed in this dissertation.