CLEAR: Counting and Location Estimate using Ambient Radio signals
This project uses backscattered signal from ambient RFID tags to count the number of occupants and their locations in a room using machine learning and inverse problem solutions.
Figure showing RFID setup in a room for assisted living application.
There are two linear problem models developed, using:
The Left figure shows a setup arrangement. Its 1/6 scaled model was simulated using CST software, with the result shown on the right using the scattering model-based reconstruction.
The problem model from both the approximations look like linear inverse problem of form y=Ax, with x as the reflectivity/ occupancy at each voxel in the imaging domain. There are various methods to solve this problem, depending on the characteristics of A. Our model has highly ill-conditioned and fat A matrix. Size of y is of the order of (#Tags * #Receivers * #Frequencies), and size of x is equal to number of voxels. Thus, it is highly underdetermined problem with size(y) « size(x). Also, in order to solve in real-time, the algorithm should be computationally efficient.
We have used fastest matched-filtering algorithm, and also employed sparsity based solutions, orthogonal matching pursuit (OMP) and fast iterative shrinkage thresholding algorithm (FISTA)3. Following figure shows we can estimate true object number and accurate location for most cases with one person standing in a 4m x 4m room. The imaging results are calculated using OMP algorithm, and show detected number of clusters using connected component labeling4, and their centroid location. For the shown result our algorithm can accuractely detect 1 cluster, with median error of 0.36 m.
While generated image can provide correct number estimate for upto 2 objects, greater than that, the linear problem model is not valid. So we’re using an intelligent neural network based approach, that is able to detect upto 5 object with high accuracy of 93% as shown in the following image. Following figure shows confusion table with true vs predicted number of occupants. Python’s Pytorch library is used to achieve the results.