Feature modeling and recognition of object behavior in 3-D space

Visual-based behavior recognition models object movement involved spatial-temporal variance to recognition and describe behaviors. In object tracking phase, we estimate affine transformation matrix by matching points to obtain 3-D movement parameters. In order to obtain state sequence which is insensitive to the duration time, we proposed a time invariant behavior feature by piecewise linear fitting method based on polygon approximation (PA). In the recognition phase, based on the dynamic time warping (DTW) behavior templates are trained and testing sequences are recognized due to the online algorithm. Interactive behavior recognition experiments of single rigid body indoor show that the proposed algorithm can modeling behavior in 3-D space via single camera, and can accurately recognize object behaviors with no pre-segmentation videos.