An FPGA implementation of the matching pursuit algorithm for a compressed sensing enabled e-Health monitoring platform

Highlights

• Using Compressed sensing technique to reduce data dimension and energy consumption.

• Hardware implementations on Zybo bord of the Matching pursuit algorithm considering:    - Low computational times,    - Low energy consumption,    - High-quality signal recovery.

• System evaluation using ECG data collected using Shimmer device.

Abstract

Wireless monitoring of physiological signals is an evolving direction in personalized medicine and home-based e-Health systems. There are several constraints in designing such systems, with two of the most important being energy consumption and data compression. Compressed Sensing (CS) is an emerging data compression technique that can be used to overcome those constraints. This work presents a low-complexity CS hardware implementation on a Field-Programmable Gate Array (FPGA) for the reconstruction of compressively sensed signals using the matching pursuit (MP) algorithm, targeting health-care applications. The proposed hardware design is based on pipeline optimization of the Programmable Logic (PL) implementation performed on the Zynq FPGA, which provides a significant performance enhancement, namely an increased processing speed and a reduced computational time since it is 115x faster than the Matlab implementation and 75x faster than the Processing System (PS) implementation carried out on the same Zynq FPGA device, while achieving alternative a high-quality signal recovery with a Peak Signal to Noise Ratio (PSNR) of 23.8 dB. Comparisons against other state-of-the-art methods showed that the low complexity of the MP algorithm can be exploited for providing almost similar results to more complex algorithms using 87–583 less Digital Signal Processor (DSP) cores, 28–540 less Block RAMs and 10,300 to 84,700 less Look-Up Table (LUT) slices.

Introduction

Advances in technology are affecting multiple aspects of the information technology field and the related industry. Performance level enhancement, computational time reduction, ease of manufacturing, and portability of the final product are currently driving the industry. Apart from advances in hardware, newer and more efficient methods and algorithms are continuously proposed as well. One such technique with the potential to benefit multiple fields and applications is Compressed Sensing (CS) [1], [2], [3]. Compressed sensing is a signal processing method for an efficient acquisition and reconstruction of a signal. It takes advantage of the fact that most signals are sparse in nature. When used under specific conditions [4], this characteristic of sparseness, provides a very powerful tool for CS against standard compression methods. CS has been used in many applications such as image and video processing [5], [6], [7], where the authors used CS to recover images or video from fewer data. It has also been applied in medical imaging, e.g. to reduce Magnetic Resonance Imaging (MRI) acquisition time [8], [9], as well as in wireless communications for sparse channel estimation [10], [11]. With the emergence and proliferation of health care applications using wearable sensors that require low energy consumption, CS is capable of providing enhanced performance in the field of physiological signals transmission [12]. Recently, owing to population aging and to the increase of life expectancy in worldwide, wireless body sensors that enable pervasive and remote health care monitoring have attracted considerably scholarly attention. Works on fall detection and human activity prediction for elderly people [13], [14], [15], as well as electrocardiography (ECG) and body temperature monitoring [16], [17], [18], have been proposed. In sum, wireless body sensors promise to allow inexpensive, continuous, and remote health monitoring.

In the present study, we propose a module for wearable sensor-based remote health care application that exploits CS techniques and programmable hardware architectures to achieve enhanced performance, Fig 1(a) illustrates the overview of the system. One aspect of the proposed work is to exploit compressed sensing in order to create a system capable of reducing the number of samples used for signal construction, which would thus minimize the consumption of energy during transmission. The second aspect of our study focuses on the hardware implementation in a portable embedded system and the acceleration of computations. To this end, we used the Shimmer platform [19] which can provide various signals: accelerometer, gyroscope, ECG, and electromyogram (EMG). The ECG is a physiological signal that reflects the cardiac electrical activity and contains substantial information about the human body. Furthermore, it has sparse signal characteristics, which can be represented as several significant atoms in a suitable basis. The CS theory ensures that, in this case, the signal can be recovered from a small number of random projections using linear programming algorithms under certain conditions. In addition, an implementation on System-on-Chip (SoC) [20] can be used to accelerate the real-time performance of the system by adapting the algorithm in order to exploit the pipeline, the parallelism, and co-design (software/hardware). For the target application, we opted to use the Zynq SoC [21], as, alongside the characteristics and advantages of traditional FPGA platforms, it additionally provides the flexibility and the choice between different interfaces, optimization, programmable logic (PL), and processing system (PS). The biggest advantage of the Zynq SoC is that it combines the software programmability of an ARM-based CPU with the hardware programmability of an FPGA. The signal reconstruction algorithm used when employing CS techniques has a major impact on the performance of a system. Currently, several reconstruction algorithms which are defined either in the context of convex optimization, or greedy approaches, are available [22]. Due to their low implementation cost and high recover speed, the most commonly used greedy algorithms are the Matching Pursuit [2] algorithm and its derivative Orthogonal Matching Pursuit (OMP) [23]. In this study, aiming to accelerate the computations in the proposed system by exploiting the pipeline while achieving high quality signal recovery, we present a hardware implementation of the basic Matching Pursuit (MP) algorithm. In response to the current demand of lower complexity and lower energy consumption, the proposed approach can be effectively employed in the developed health care applications. The proposed schemes are implemented and tested out on the Xilinx Zybo board [21]. The implementation is simulated and tested using synthetic and real ECG signals collected using the Shimmer platform. The following two implementations were developed and tested: i) A software running on the Processing System (PS) part of the Zybo board and ii) an Intellectual Property (IP) core that executed in the Programmable Logic (PL) part of the board. Multiple optimizations were used in order to reduce complexity and computational time. The remainder of this paper is organized in as follows. Section 2 provides background information in the field of Compressed Sensing, while the architecture and the methods used in the proposed health monitoring systems are described in Section 3. Furthermore, in Section 4, the performance of the proposed systems is evaluated, whereas conclusions are drawn and future research directions are outlined in Section 5.