Signature descriptions for motion trajectory representation, perception, recognition and reproduction

Abstract

Motion trajectory is a compact and informative clue in characterizing the motions of humans, robots or moving objects. In this research, a novel signature mechanism is proposed in Euclidean space to serve as a 3-D motion trajectory descriptor. The signature principles are firstly presented and then the signature’s applications are explored for adaptive trajectory representation, motion perception, trajectory recognition and reproduction. In this thesis, three flexible signature descriptions are defined. The basis of the signature descriptor is the Euclidean differential invariants features. The full signature is a fundamental and complete description to the raw trajectory data. To make it computationally reliable, low order joint Euclidean invariants are employed to implement an approximate signature to avoid the noise-sensitive, high order derivatives. In addition, the signature is also proven to be able to describe irregular trajectories and complicated motions. The optimized signature is the condensation of the full signature for a more compact description. The cluster signature is a mixture model-based, probabilistic description for a motion class/pattern, which can be developed based on either the full signature or the optimized signature. High functional adaptability can be achieved from the flexible signature descriptions to meet different application requirements in characterizing different motions. Invariants are a key measure to the flexibility of a motion trajectory descriptor. Substantial descriptive invariants can be deduced from the proposed trajectory signature, which is attributed to the signature’s computational locality. The signature’s rich invariants are elaborated from different points to demonstrate the resulting flexibility. Adaptive motion perception and trajectory recognition are shown using the flexile signature descriptions. The motion perception is explored from three different perspectives, i.e. a single-signature based, salient feature perception, the dynamic time warping-based inter-signature matching for comparing two motions and the Gaussian mixture regression-based signature generalization for perceiving a motion class/pattern. Motion perception can support high level analysis for motions or robot tasks by offering an intuitive perception interface to highlight the motion features of interest. Meanwhile, three signature recognition solutions are investigated in correlation with the three signature descriptions by developing corresponding classification engines. While high recognition accuracy can be obtained by the nonlinear matching of the full signature, the linear matching of the optimized signature can show higher efficiency. The Bayesian signature recognition method is more accurate and efficient, as it can take advantage of prior knowledge. The trajectory recognition is helpful for understanding the meaning of a motion, which can benefit the human-robot interactions and intelligent robot task learning. A typical application of the signature is investigated to support effective robot task description. Free form 3-D motion trajectories extracted from spatiotemporal motions are appropriate to characterize a kind of long-term and temporally continuous robot task. The signature descriptor is employed to serve as a generic and generalized task description to be learned by a robot instead of learning the raw data directly. More importantly, a corresponding trajectory reproduction algorithm from the learned signature is formulated as well, which enables a robot to re-run the learned task by instantiating a trajectory instance. The signature description and the reproduction algorithm can serve as the core for a potential motion trajectory-oriented robot learning system. In addition, the supports to cognitive robot learning offered by the motion perception and trajectory recognition are also pointed out. Experiments are conducted to verify the signature’s effectiveness. Human sign language is generally adopted as the experiment entry. Apart from using a set of robot vision setups for motion tracking and trajectory extraction, a public available sign dataset is also used. Various experiments are designed for both small-scale and large-scale applications, in which the real data, synthetic data, noisy data and multi-user data are used to test the signature’s performance. The experiment results show the signature’s capabilities and flexibility that can satisfy adaptive motion trajectory representation, perception, recognition and reproduction. Index Terms: Signature, motion trajectory, trajectory descriptor, trajectory representation, motion perception, trajectory recognition, trajectory reproduction, robot learning, robot vision.

Date of Award	16 Feb 2009
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	You Fu LI (Supervisor)