This article presents a multi-DOF motion sensing system consisting of a permanent magnet (PM) and a magnetic tensor sensor (MTS) comprising a 3 x 3 array of three-axis digital magnetic flux density (MFD) sensors, and the methods to measure their relative MTS-PM position/orientation in a 3-D space. The design enables redundant differential measurements of MFD vectors and gradient tensor components to account for the singularities due to matrix inversion, providing a basis to explore different methods for multi-DOF estimation of the PM position/pose. Formulated as a two-stage linear least-square (LS) problem to take advantage of the dipole simplicity and exploit its physics revealed by its inverse model to guide the design of a fully connected artificial neural network (ANN) to account for the MTS measurement noise and un-modeled factors, a prototype multi-MTS system capable of 5-DOF motion measurements is developed and evaluated experimentally along with a study analyzing the parametric effects on the estimation accuracy; both stationary and moving sensor scenarios are considered. Enhanced with a fully connected ANN, an accuracy within a root-mean-square error (RMSE) of 40 mu m spatial position and 0.1 degrees pose can be uniquely obtained without subtracting a predetermined geomagnetic field in both fixed and moving multi-MTS scenarios, representing a significant improvement over the (0.5 mm, 1 degrees) RMSE of the single-MTS.

This article offers an impedance sensing method taking advantage of the conductivity changes due to muscle contraction to estimate muscle-driven joint torques through a convolutional neural network (CNN), where the input images are derived from a finite set of boundary voltage measurements. Guided by a physical model combining the forearm biomechanics and the muscle electric field along with the CNN criteria considering the receptive fields (RFs), the effects of two image formats [for quasi-static (QS) and dynamic (DYN) states] on the CNN performance are experimentally studied on eight human subjects' forearms using a prototype impedance sensing system. By comparing the CNN-estimated torques with that measured on a haptic device, the findings verify that the impedance-based method can estimate the joint torques driven by both the deep and superficial muscles within 9% errors of the three degrees-of-freedom wrist torque and 10% error of the gripping torque and that it is feasible to share data among a similar group to reduce data collection and time when training a CNN for uses on a new subject.

This article presents a three-degree-of-freedom (3-DOF) magnetic sensor, referred to here as a pantographic exoskeleton (PGE) sensor, for monitoring in real time the internal human-joint motion in the sagittal plane. With two sets of embedded magnetic sensors and a permanent magnet, the PGE wearable on a healthy leg or lower extremity exoskeleton (LEE) independently measures the 2-DOF translations and the joint angle. Two sensor estimation methods, which are the model-based and the artificial neural network (ANN), are experimentally analyzed in the presence of measurement noise. As an illustration, the PGE sensors are evaluated for sit-to-stand (STS) exercises, where the real-time measurements are verified by comparing with the joint angles determined by a commercial VICON motion capture system, and the translational deviations measured on a 4-DOF platform manipulated to follow a specified internal motion trajectory of an ankle joint during STS. With the ANNs appropriately trained to account for measurement noise, the PGE sensors can track the joint angles while measuring the internal motions of both legs with or without the LEE demonstrating that the PGE sensor has the potential to serve as an indicator of stroke rehabilitation where patients lack force perception and suffer an increased risk of falls due to the weak affected leg.

This article offers an image processing method to optimize the design of a three-level color machine vision algorithm illustrated in the context of presenting a whole chicken carcass for subsequent handling, which overcomes some common problems in image segmentation, feature identification, and pose estimation for classifying, identifying, and locating poultry meat products. First, artificial color contrast and principal component analysis are optimized to enhance the contrast between similar colors to extract targeted regions more effectively. Second, the boundaries of poultry products are mostly smooth curves that are proposed as features for object recognition. Third, image feature points are found using template matching, from which the object (size, location, and pose) can be accurately determined. The algorithm simplifies the complex edges of a poultry product into a finite number of arc-centers, which greatly improves the efficiency/accuracy of the matching for part-presentation. Evaluated with 100 randomly posed samples, the algorithm has a success rate of 93%. The 7% failures were primarily due to missing points, which can be eliminated by incorporating more template points.

Illustrated in the context of home-based stroke rehabilitation, this paper presents the design and development of magnetic flexonic sensor nodes (FSNs) with configurable parameters for simultaneous sensing of normal forces and displacement. The FSN is composed of a magnetic sensor in flexible elements that can be embedded in wearable devices and rapid-prototyped at a low cost to facilitate distributed force sensing applications. As an illustration, a shoe insole with 19 embedded FSN for monitoring the plantar force distribution and arch height has been designed, prototyped, and experimentally evaluated by comparing it with a commercial dynamometric platform. The FSN-embedded insole directly measures the arch height and captures the transients of the plantar forces in five different zones in real-time. Because of the simplicity and low cost, FSNs have the potential to facilitate fabricating low-cost sensing systems, for example, for home-based stroke rehabilitation where the insole stiffness and arch height of the embeddable electronic sensors can be configured according to the patient's specific abnormal plantar distribution to adapt to different range and distribution and wearable configuration during different rehabilitation state. © 2016 IEEE.

This article presents a binocular visuotactile sensor (BVTS) and its model for reconstructing the 3-D deformation due to mechanical contact on its elastomer surface. The model explicitly accounts for the refraction effects of the elastomer on the depth map reconstructed from the imaged markers. To help gain physical insights into the refraction effects on the depth map reconstruction and experimentally analyze the effectiveness of the model-based BVTS, a prototype BVTS has been developed for experimental investigation. The refractive indexes of the elastomer, the acrylics, and their combination are experimentally determined from the model-based reconstruction and verified with that measured by a spectroscopic ellipsometer; the differences are within 3%. As a basis for evaluation, the reconstructed depth maps are compared with that of the two other commonly used practices, stereovision (SV) calibration without considering refraction and integration camera model that lumps the refraction effects with the SV parameters during calibration. By explicitly modeling the refraction physics using ray tracing, the model-based BVTS faithfully reconstructs the points on the actual height and is capable of dynamically tracking small torsional displacements with balls of different radii (3-15 mm) with a depth error below 4%.

This paper presents a distributed soft contact sensing method based on a sensitivity mapping function, which relates the change in measured voltages to that in the elastomer conductivity due to contact force acting on its surface. The sensitivity-image-based sensing system uses a small number of boundary electrodes with a multiplexer to create different electric-field patterns to generate a series of sensitivity images for machine learning, significantly reducing the number of training data typically obtained with single-point indentation measurements. The mapping function, which does not rely on the knowledge of the electric and conductivity fields during online sensing, can be trained with only a small amount of measured data in the order of the square of electrode number. The proposed method has been experimentally evaluated on two 16-electrode prototypes trained with 129 and 80 data for two typical applications, which are tactile perception on a flat surface and contact force measurement in a model knee joint, respectively. The former verifies the measurement accuracy of the contact position and force magnitude while the latter demonstrates the application of this method for measuring the internal joint forces between two curved surfaces.

Motivated by the increased needs of home-based rehabilitation for stroke patients, more and more interest has been drawn towards developing body-fixed sensors for monitoring affected joint motions. Although relatively accurate bone geometries can be obtained by scanning technologies, most human joints are approximated by simple circles and spheres to reduce the highly nonlinear kinematics to a tractable form for motion studies; many human joint-motion sensing challenges remain open. This article presents a novel magnetic tensor sensor (MTS) for noncontact tracing a human joint trajectory and a physics-based measurement model implemented on an artificial neural network (ANN) to account for un-modeled factors. The effects of input configurations and datatypes on measurement accuracy of an MTS/ANN have been numerically investigated with published data and experimentally evaluated on a prototype pantographic exoskeleton worn on a human shank/foot. As demonstrated experimentally, the MTS/ANN system calibrates the sensor intrinsic parameters, accounts for the environmental magnetic effects on the measurements, and can be trained with both offline and user-specific data.

Motor imagery (MI) brain–computer interface (BCI) systems have broad application prospects in rehabilitation and other fields. However, to achieve accurate and practical MI-BCI applications, there are still several critical issues, such as channel selection, electroencephalogram (EEG) feature extraction and EEG classification, needed to be better resolved. In this paper, these issues are studied for lower limb MI which is more difficult and less studied than upper limb MI. First, a novel iterative EEG source localization method is proposed for channel selection. Channels FC1, FC2, C1, C2 and Cz, instead of the commonly used traditional channel set (TCS) C3, C4 and Cz, are selected as the optimal channel set (OCS). Then, a multi-domain feature (MDF) extraction algorithm is presented to fuse single-domain features into multi-domain features. Finally, a particle swarm optimization based support vector machine (SVM) method is utilized to classify the EEG data collected by the lower limb MI experiment designed by us. The results show that the classification accuracy is 88.43%, 3.35–5.41% higher than those of using traditional SVM to classify single-domain features on the TCS, which proves that the combination of OCS and MDF can not only reduce the amount of data processing, but also retain more feature information to improve the accuracy of EEG classification. © 2021, The Author(s).