Motion Data Processing

Motion Skeleton Model

Current motion capturing technologies (e.g., Microsoft Kinect, Vicon Vantage, Asus Xtion…) record human motions at high frame-per-second rates. Motions are recorded as a series of 3D coordinates of body joints in space and time. The recorded motion data can be processed and utilized in a variety of applications, for example, in sports for comparing the performance of athletes, in security for identifying special-interest persons, in medicine for determining the success of rehabilitative treatments. All these applications require an effective and efficient (sub)motion-to-motion similarity matching.

DEMO #1: Action recognition

Action recognition demoClassification task: recognize labels of any user-selected short motion. A short motion (sub)sequence is selected as a query. Retrieval engine compares the query with 2K+ categorized motion samples to obtain the ranked list of nearest matches from which the most probable action label is obtained. The demonstration currently supports 130 label categories.

»Try the action recognition demo

DEMO #2: Subsequence search

subseq_searchSubsequence search task: given a short query, search for similar subsequences within long motions. A total motion sequence of 12 hours can be searched. A short query motion is specified by the user using the query-by-example paradigm. Query-similar sub-motions are retrieved and displayed, and ordered according to their similarity score. A new query can be selected from the retrieved results.

»Try the subseqeunce search demo

LSMB19: Dataset for benchmarking search and annotation

LSMB19 Dataset and BenchmarkWe propose a benchmark to evaluate search and annotation algorithms. The benchmark contains a motion dataset of 2 very long and continuous unsegmented 3D skeleton sequences, training and testing data for two modalities (cross-subject and cross view), 98 search queries and ground truth labels.

»Visit LSMB19 benchmark homepage

Related Publications