Falls are a major cause of fatal and nonfatal injuries in people aged 65 years and older. Radar has the potential to become one of the leading technologies for fall detection, thereby enabling the elderly to live independently. Existing techniques for fall detection using radar are based on manual feature extraction and require significant parameter tuning in order to provide successful detections. In this paper, we employ principal component analysis for fall detection, wherein eigen images of observed motions are employed for classification. Using real data, we demonstrate that the PCA based technique provides performance improvement over the conventional feature extraction methods.
Recent research and developments for in home radar monitoring have shown real promise of the technology in detecting normal and abnormal gross-motor activities of humans inside their residences and at private homes. Attention is now paid to challenges in system integration, operations, and installations. One important question touches on the required number of radar units for a given residence and whether eventually one radar unit per room would become the nominal approach. Towards addressing this question and assessing the effectiveness of radar unit to sense adjacent rooms and hallways of the same residence, this paper examines through-wall radar monitoring where the radar signal faces both wall attenuation and dispersion. We show that typical interior walls do not significantly alter the radar time-frequency (TF) signature of a fall, and the radar signal return is slightly weakened by wall penetration. Additionally, we show that there is a wide variation of the TF feature values associated with fall motions which confuse a classifier, trained with generic subjects, and cause it to falsely declare a different motion.
Multi-window spectrograms offer higher energy concentration in contrast to the traditional single-window spec- trograms. However, these quadratic time-frequency distributions were not introduced to deal with randomly undersampled signals. This paper applies sparse reconstruction techniques to provide time-frequency represen- tations of nonstationary signals using the Hermite functions as multiple windows, under randomly sampled or missing data. The multi-window sparse reconstruction approach improves energy concentration by utilizing the common local sparse frequency support property across the different employed windows.
Missing samples in the time domain introduce noise-like artifacts in the ambiguity domain due to their de facto
zero values assumed by the bilinear transform. These artifacts clutter the dual domain of the time-frequency
signal representation and obscures the time-frequency signature of single and multicomponent signals. In order
to suppress the artifacts influence, we formulate a problem based on the sparsity aware kernel. The proposed
kernel design is more robust to the artifacts caused by the missing samples.
This paper considers compressive sensing for time-frequency signal representation (TFSR) of nonstationary radar signals
which can be considered as instantaneously narrowband. Under-sampling and random sampling of the signal stem from
avoiding aliasing and relaxing Nyquist sampling constraints. Unlike previous work on compressive sensing (CS) and
TFSR based on the ambiguity function, reduced observations in the underlying problem are time-domain data. In the
reconstruction process, Orthogonal Matching Pursuit (OMP) is used. Since the frequency index in the first iteration of
OMP is the same as the one obtained by finding the frequency position of the highest Spectrogram peak, it becomes
necessary to consider several OMP iterations to improve over Spectrograms performance. We examine various methods
for estimating IF from higher number of OMP iterations, including the S-method. The paper also applies CS for signal
time-frequency signature estimations corresponding to human gait radar returns.