Discovering Unusual Behavior Patterns from Motion Data

Abstract

As there are more and more surveillance cameras installed in public places a challenging problem is to discover unusual behavior patterns from a huge amount of video data However this task is currently only feasible for human beings because both object recognition and intention detection are still difficult for computer vision algorithms Recently with the release of low-cost depth cameras motion data containing 3D coordinates of skeleton joints can be directly captured thus facilitating following analysis tasks In this work we devise a complete system flow which includes steps of coordinate transformation normalization segmentation feature extraction and dimensionality reduction so as to achieve the purpose of discovering unusual behavior patterns Note that prior works generally require to predefine (either normal or abnormal) behavior patterns and then utilize data classification techniques for further analysis Instead we propose in this work to utilize data clustering techniques to discover unusual behavior patterns so that the inconsistencies of defining behavior patterns in various scenarios can be eased Specifically we adopt a density-based clustering technique and adjust the values of two parameters (i e radius and minimum points) to appropriately generate unusual behavior patterns Finally two datasets including MSR action recognition and AffectME (affective multimodal engagement) are used in our experiments for evaluation purposes Empirical studies show that our approach is effective for discovering unusual behavior patterns

Date of Award	2014 Jan 20
Original language	English
Supervisor	Wei-Guang Teng (Supervisor)