Multi-Feature Based Skeletal Motion Retrieval

As the production of 3D games are growing more and more quickly, computer science and technology plays an important role in many applications, just like computer animation, interactive virtual reality, film production and video game development. What’s more, computer is widely used in education, national defence and satellite research. The increasing improvement of computer software and hardware provides a convenient way in developing these products for us.With a vast amount of skeletal motion data at hand, we need to find out the substitutive characteristics which can describe the whole motion sequence well, to search for the data wanted accurately and efficiently, and retrieve the target motion sequence which users need.In this work, we propose an effective scheme for content-based retrieval of skeletal motion data. The first important novelty is that we extract and describe multiple features of the motion sequences, at various levels of semantic abstraction. Furthermore, in order to facilitate efficient feature matching, we reduce the dimensionality of the feature descriptors through principal component analysis and cluster analysis, and concisely represent a motion sequence by multiple motion histograms, one for each feature. Finally, the retrieval is achieved by measuring and sorting the similarity between the motion histograms of the query and those of the motion sequences in the database. Outstanding performance of the proposed algorithm is well demonstrated by experiments.The main contribution we make in this work can be described as follows,1、Multi-feature extraction of motion data.Different from the traditional single motion feature extraction methods, we comprehensively consider the geometrical characteristics of the human body skeletal structure which can more truly reflect the intrinsic properties of the movement, and select four representative geometrical characteristics to describe the motion sequence more accurately in order to improve the retrieval precision.2D geometric features can only describe local geometric structure of the movement well, while three dimensional space-time characteristics we extracted in this paper are based on spatial location relationships, they can effectively express the independence movement information of the joint points, and reflect the interaction of movement properties from the perspective of the global geometric structure.2、Dimensionality reduction of feature descriptorsDue to the dimension extracted from movement characteristics is large, so we reduce the dimensionality of the feature descriptors through principal component analysis and cluster analysis, in order to avoid dimension disaster and for higher accuracy with least cost of feature matching.3、Feature matching of motion dataIn order to adapt to the needs of large-scale database, we propose the presentation of motion histogram after preprocessing the motion data with dimensional reduction and cluster analysis, and build the histogram after computing the occurring frequency of each category. At last, we compute the Euclidean distance between any two histograms in order to match the query sequence and target sequence in the database.4、Evaluation metric of experimental resultsWe propose two metrics MAP and P@n to evaluate the similarity between the query sequence and target sequence in the database.