Hierarchical Aligned Cluster Analysis for Temporal Clustering of Human Motion

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Understanding human motion requires segmenting complex movements into basic units called motion primitives.
• Discovering these primitives is challenging due to the vast number of movement combinations, variable action durations, and intricate articulated motion representations.

Purpose of the Study:

• To develop an unsupervised framework for learning human motion primitives.
• To address the challenges in temporal segmentation and clustering of complex time series data.

Main Methods:

• Introduced Hierarchical Aligned Cluster Analysis (HACA), a bottom-up hierarchical clustering framework.
• Combined kernel k-means with a generalized dynamic time alignment kernel for time series clustering.
• Utilized coordinate descent and dynamic programming for efficient optimization.

Main Results:

• HACA successfully segmented complex motions in motion capture and video data.
• The framework demonstrated effectiveness as a visualization tool for time series data.
• Performance was validated against state-of-the-art algorithms using honey bee dance data.

Conclusions:

• HACA provides an effective unsupervised approach for temporal segmentation and clustering of human motion.
• The method offers a robust framework for discovering motion primitives and low-dimensional time series embeddings.
• The availability of HACA code facilitates further research in computational motion analysis.