Ultrasonic Array Doppler Sensing for Human Movement Classification

Abstract

Abstract-Classification of human movements is an important problem in healthcare and well-being applications. An ultrasonic array Doppler sensing method is proposed for classifying movements from a given set. The proposed method uses velocity and angular information derived from Doppler frequencies and direction-of-arrival (DoA) by processing the signals at the receiver sensor array. Doppler frequency estimation is done by obtaining an initial estimate based on the Fourier transform in conjunction with a predictive tracker. A Root-MUSIC algorithm is used at the estimated Doppler frequencies to obtain DoA corresponding to the dominating moving object. Using speed, direction, and angle as features, a Bayesian classifier is employed to distinguish between a set of movements. The performance of the proposed method is evaluated using an analytical model of arm movements and also using experimental data sets. The proposed ultrasonic Doppler array sensor and processing methods provide a new, compact solution to human arm movement classification.

Ultrasonic circular array sensing for human arm motion classification

“…top from bottom. This extends the flexibility of the method compared to the previously considered ULA [5], making it more suitable for the application. …”

“…The proposed sensor consists of an ultrasonic transmitter and a co-located UCA receiver, thereby allowing extraction of directional information in two dimensions from a correlation analysis of the received signals. The circular configuration thus provides more angular information than a uniform linear array (ULA), as considered in previous work [5], and therefore extends the set of movement classes that may be distinguished.…”

“…Following the steps described in [5], the received signal at the p-th sensor element in the UCA as a result of N reflecting moving objects, where object i has velocity v i (t) and v s |v i (t)|, is given by…”

“…The dominant Doppler frequency is then found by employing maximization in conjunction with a tracker, using the physical (short term) predictability of movements [5]. For the sake of clarity, the method is divided into three steps which are repeated for each frame n. First, a probability density function (PDF), reflecting our belief that a trustworthy Doppler frequency is found at a certain frequency f and frame m > n, is derived using autoregressive prediction based on the Doppler frequencies found in the past.…”