Indexing Large Human-Motion Databases

Abstract

Data-driven animation has become the industry standard for computer games and many animated movies and special effects. In particular, motion capture data recorded from live actors, is the most promising approach offered thus far for animating realistic human characters. However, the manipulation of such data for general use and re-use is not yet a solved problem. Many of the existing techniques dealing with editing motion rely on indexing for annotation, segmentation, and re-ordering of the data. Euclidean distance is inappropriate for solving these indexing problems because of the inherent variability found in human motion. The limitations of Euclidean distance stems from the fact that it is very sensitive to distortions in the time axis. A partial solution to this problem, Dynamic Time Warping (DTW), aligns the time axis before calculating the Euclidean distance. However, DTW can only address the problem of local scaling. As we demonstrate in this paper, global or uniform scaling is just as important in the indexing of human motion. We propose a novel technique to speed up similarity search under uniform scaling, based on bounding envelopes. Our technique is intuitive and simple to implement. We describe algorithms that make use of this technique, we perform an experimental analysis with real datasets, and we evaluate it in the context of a motion capture processing system. The results demonstrate the utility of our approach, and show that we can achieve orders of magnitude of speedup over the brute force approach, the only alternative solution currently available.

Less is More: Similarity of Time Series under Linear Transformations

“…Table 2 gives the results. In the literature uniform scaling has been reported to perform superior for certain types of datasets [6,7,4] but to our surprise, LTD performs consistently comparable to DTW (and in some cases clearly better. Only with the synthetic control charts DTW seems better, but we will comment on this case at the end of the next subsection.)…”

“…In [7] uniform scaling (US) of a query y (of length m) to a new length p is considered (but no offset), to better match a reference series x (of length n) under the Euclidean distance:…”

“…It has also been reported that uniform scaling performs superior for applications such as motion capturing, gait recognition or query-by-humming [6,7,4]. Yet, for the general case, linearly transformed time goes almost unconsidered: in [3] (published after work on uniform scaling [6,7,4]) a broad comparison of time series similarity measures is carried out, but the list of distance measures does not contain a measure dedicated to the likely setting of just linear transformations.…”

Querying Complex Spatio-Temporal Sequences in Human Motion Databases

“…One typical example is the retrieval of motion capture data from a large motion database, which is very useful in computer animation. Many approaches for matching high dimensional numerical time series are based on numerical similarities, in which the high dimensional time series data are typically decomposed into a number of trajectories with low dimensionality (1-dimensional to 3-dimensional) [7], [18]. Warping distances such as DTW distance are preferred to match those low dimensional trajectories to handle time shifting and scaling of time series.…”

“…This essentially converts the data into a high dimensional time series and the distance between any pair sequence is then measured based on the distance of the two time series. However, existing studies on the distance measures of time series [4], [5], [6], [7] have been focused on time series with low dimensionality (usually 1-dimensional to 3-dimensional trajectories). They cannot be easily applied to high dimensional time series because of the following three reasons.…”