Multiple human tracking and identification with wireless distributed pyroelectric sensors
The advantage of thermal human tracking over the optical counterpart lies in its theoretical and practical illumination invariance. The goal of our research is to develop a prototype wireless distributed pyroelectric sensor system, which can track multiple human objects inside a room, while maintaining their identities under all-illumination circumstances. It involves two sub-problems: multiple thermal source tracking and thermal object identification.
Throughout construction of the prototype sensor system, the following topics have been addressed, investigated, and managed:
1. Signal processing balance between photonics and electronics. Given the feasibility of visibility modulation using Fresnel lens arrays for the pyroelectric sensor, we have explored different visibility coding schemes and sensor configurations & deployments to achieve higher sensing efficiency and efficacy, better tracking precision, and more representative human thermal features.
2. Algorithm trade-off between performance and cost, for real-time system implementation. We have investigated and tested various signal processing, event inference, data-object-association, and tracking techniques, for the multiple human tracking purpose, as well as supervised regression and unsupervised expectation-maximization learning schemes, to the human thermal feature clustering and identification end. Computational choices and compromises have been made to neutralize the conflicts between tasks and resources of the sensor system.
3. Distributed computation and communication management amongst host, master, and slave modules. We have developed a prototype sensor system, consisting of several sensing enabled slaves, one master, and one host. Efforts have been made in assigning all the computing, from event detection/capture and signal digitization to communication synchronization and error rejection to data fusion/association/learning and task synthesis; amongst these three kinds of modules, so as to draw an efficient management of computation and communication resources out of their heterogeneous capabilities and purposes.
The main accomplishments of this thesis include the following four components. (1) A real-time demo of one human object tracking under a Bayesian tracking scheme with wireless distributed pyroelectric sensors. (2) A real-time demo of multiple human object tracking under an EM-Bayesian tracking scheme with wireless distributed pyroelectric sensors, whose unique two-column optic geometry is designed to facilitate the process of data-object-association. (3) A real-time demo of path-dependent human identification based on linear regressions of spectral features of event signals. (4) A real-time demo of path-independent human identification based on Hidden Markov Model's accommodation of digital event index sequences that are chosen as statistical features of human objects.
This thesis proposes and explores the unexploited frontier and terrain of multiple human object tracking and identification in pyroelectric terms. On account of what we accomplished, it can be heralded that the low cost, low power consumption yet reliable pyroelectric detector will, in the near future, rise to be a mainstream human detection instrument, beside its video and audio counterparts, for extensive applications from human-machine interfaces to human biometrics.