For variance-based RTI (VRTI), variance can be caused by two types of motion: extrinsic motion and intrinsic motion. Extrinsic motion is defined as the motion of people and other objects that enter and leave the environment. Intrinsic motion is defined as the motion of objects that are intrinsic parts of the environment, objects which cannot be removed without fundamentally altering the environment. If a significant amount of windowed variance is caused by intrinsic motion, then it may be difficult to detect extrinsic motion.
We noticed the effects of intrinsic motion on VRTI in a repeat of the identical experiment reported in Wilson 2009 that we performed in May, 2010. This new experiment was performed at the same location and using the identical hardware, number of nodes, and software. Yet, in the new experiment, variance-based radio tomographic imaging (VRTI) does not always locate the person walking inside the house as accurately as reported in Wilson 2009. Sometimes the position estimate error is as large as six meters. Investigation of the experimental data quickly indicates the reason for the degradation: periods of high wind. The old experiment (we call Experiment 1 in the following) is performed on a clear winter day, while the new experiment (Experiment 2) is performed on a windy day in late spring. From the video recorded, we can see there are no leaves on branches and no wind is observed during Experiment 1. However, for Experiment 2, we observe from our recorded video that wind causes grass, leaves and branches to sway. The wind also causes the PVC stands supporting the nodes to move. The swaying of leaves and branches and the movement of the PVC stands are intrinsic parts of the environment, which cannot be avoided, even without people present in the environment.
To reduce the effects of intrinsic motion, we propose two robust estimators: subspace variance-based radio tomography (SubVRT) and least squares variance-based radio tomography (LSVRT). The estimates from VRTI and LSVRT are shown in the Figures above. The details of these two estimators can be found here . The experimental data of Experiment 1 and Experiment 2 can be downloaded for academic use.