Through-wall human monitoring using data-driven models with doppler information

Through-wall human monitoring within a highly cluttered environment is a problem of current interest. Example applications include law enforcement, disaster search-and-rescue, and urban military operations. The purpose is to clearly monitor humans through building walls using a radar system. Doppler-based sensors offer an inexpensive way to detect moving targets in the presence of stationary clutters. It also provides information regarding motions of the human by micro-Doppler returns. In this dissertation, the applications of data-driven model (DDM) are investigated for locating human subjects and classifying their activity using Doppler sensors. DDM is a mathematical model trained by a set of data that describe the input-output relationship. It is suitable for real-time applications. As DDM, an artificial neural network (ANN) and a support vector machine (SVM) are considered. A collection of Doppler sensors is studied to localize humans in two ways: the use of spatially distributed Doppler sensors and the use of a single-sensor array. Furthermore, the feasibility of classifying human activities is studied with the obtained Doppler information. First, an ANN is proposed to track humans using the Doppler information measured by a set of spatially distributed sensors. The ANN estimates the target position and velocity given the observed Doppler data from multiple sensors. A point-scatterer model is used for the training data generation. For the verification of the proposed method, a toy car and a human moving in a circular track are measured in line-of-sight and through-wall environments. Second, an array-processing algorithm is proposed to estimate the number of targets and their Direction-of-Arrival (DOA) based on ANN when the available number of sensor elements is small. Using software beamforming, a number of overlapping beams are simultaneously formed. The received signal strengths from all the beams produce a unique signature in accordance with the target locations, as well as the number of targets. The identification of the number of targets and their locations is carried out sequentially via ANNs. For the verification of the algorithm, both line-of-sight and through-wall measurements are performed using loudspeakers driven by audio tones and moving humans. Third, an SVM is proposed to classify activities of a human subject using the measured Doppler information. MicroDopplers from moving limbs of human subjects contain significant information regarding their activities. Seven different human activities of twelve human subjects are measured in the laboratory using a Doppler radar. Six microDoppler features are extracted from the resulting spectrograms. A decision-tree based SVM is used for the classification of seven activities based on the features. Diverse situations such as combination of different activities, oblique angle case, and throughwall case are also discussed.