A hybrid approach for human motion retrieval

Retrieving similar motions from motion databases has become an essential topic of computer animation research. The use of binary geometric features and inverted indexes is one of the efficient solutions to this problem. This approach can be used with variation and inexactness algorithms for fuzzy searches. However, close similarity searches are not possible. In addition, the process is not automatic since it needs user input for selecting binary features to use. In another efficient approach, k-d tree with medium sized numerical based feature sets and shortest path search on a directed graph are used. However, it does not propose any tool for fuzzy searches. In this thesis, we propose a hybrid approach that utilizes numerical based feature sets, k-d tree based indexing structure and inverted index based motion matching technique. Our hybrid approach does not need user input, can be used in environments requiring close similarity of motions and can be used with variation and inexactness algorithms. Our results show that our hybrid approach is useful and efficient for similarity searches on motion databases.