Managing resources on a multi-modal sensing device for energy-aware state estimation

Sensor management;State estimation;partially observable Markov decision process (POMDP);Markov models;Embedded implementation;Vehicle detection;Energy scalability

Cohen, David ; Jones ; Douglas L.

关键词: Sensor management;    State estimation;    partially observable Markov decision process (POMDP);    Markov models;    Embedded implementation;    Vehicle detection;    Energy scalability;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/45372/David_Cohen.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Multi-modal sensing devices are becoming more and more prevalent in everyday life. Whether it be in the form of a smartphone, mobile computing device, remote sensor node, or a sensor-packed robot, they are used almost everywhere. Often these devices run on battery power or on energy harvested from the environment. In these situations, energy is at a premium, and resources must be intelligently managed to balance energy consumption and system performance. We develop a methodology for joint sensor scheduling and state estimation on an energy-constrained device. Our approach is similar to existing sensor scheduling methods for hidden Markov models. We extend these methods, and cast the problem as a standard partially observable Markov decision process (POMDP), for which numerous exact and approximate solutions are well known. We then demonstrate optimal sensing policies on a vehicle detection application. A sensing platform is developed consisting of an ultra-low power MSP430 Micro Controller Unit (MCU), a high-performance ARM-based MCU, a passive infrared motion sensor, and a camera. This platform is capable of 100× energy scalability between sensing modalities. Appropriate POMDP model parameters are extracted from real data traces, and these are used to evaluate the expected performance of optimal sensing policies across a range of energy levels. These policies are then run on real data in order to compare actual performance to theoretical performance. We show that this performance gap is small in most cases, demonstrating both the theoretical and practical value of our sensor management techniques.