We implement transfer learning (TL) to attain efficiency in training and feature extraction by freezing the connection weights of shallow layers of deep learning (DL) models, already trained for a given task, to be used in another. This eliminates the need to retrain the whole model from scratch for the new task, resulting in significantly reduced training times, while maintaining the classification accuracy at acceptable levels. Specifically, we investigate bidirectional feature transfer between the human activity recognition (HAR) and user identity recognition (UIR) tasks through the use of wearable motion sensors. Both tasks have recently drawn considerable attention and are significant because of their relevance in a broad range of applications. Knowledge transfer between tasks via TL brings many advantages. We employ the features extracted by our recently proposed hybrid DL model which has proven superior to commonly used state-of-the-art DL models. The model can extract features more effectively because of its parallel-running 2D CNN and LSTM branches that receive complementary input types, unlike many hybrid models that implement a series connection that receives a single type of input. To our knowledge, TL between the above-mentioned tasks has not been implemented to date. We demonstrate high-accuracy bidirectional feature transfer between the HAR and UIR domains that shortens the training time by a factor that varies between 14.3 to 34.4, without much degradation in the classification accuracy. To provide a typical figure, the accuracy degrades by 2.45% for a group of three dynamic activities. The key technical challenge of this work is maintaining the classification accuracy at acceptable levels while benefiting from the advantages that TL offers. We expect the results of this work to have impact by way of reducing the demanding computational requirements in training resource-limited wearable edge devices.

We present a novel heuristic fall-detection algorithm based on combining double thresholding of two simple features with fuzzy logic techniques. We extract the features from the acceleration and gyroscopic data recorded from a waist-worn motion sensor unit. We compare the proposed algorithm to 15 state-of-the-art heuristic fall-detection algorithms in terms of five performance metrics and runtime on a vast benchmarking fall data set that is publicly available. The data set comprises recordings from 2880 short experiments (1600 fall and 1280 non-fall trials) with 16 participants. The proposed algorithm exhibits superior average accuracy (98.45%), sensitivity (98.31%), and$F$-measure (98.59%) performance metrics with a runtime that allows real-time operation. Besides proposing a novel heuristic fall-detection algorithm, this work has comparative value in that it provides a fair comparison on the relative performances of a considerably large number of existing heuristic algorithms with the proposed one, based on the same data set. The results of this research are encouraging in the development of fall-detection systems that can function in the real world for reliable and rapid fall detection.

We propose a new methodology to attain invariance to the positioning of body-worn motion-sensor units for recognizing everyday and sports activities. We first consider random interchangeability of the sensor units so that the user does not need to distinguish between them before wearing. To this end, we propose to use the compact singular value decomposition (SVD) that significantly reduces the accuracy degradation caused by random interchanging of the units. Second, we employ three variants of a generalized classifier that requires wearing only a single sensor unit on any one of the body parts to classify the activities. We combine both approaches with our previously developed methods to achieve invariance to both position and orientation, which ultimately allows the user significant flexibility in sensor-unit placement (position and orientation). We assess the performance of our proposed approach on a publicly available activity data set recorded by body-worn motion-sensor units. The experimental results suggest that there is a tolerable reduction in accuracy, which is justified by the significant flexibility and convenience offered to users when placing the units.

We propose techniques that achieve invariance to the positioning of wearable motion sensor units on the body for the recognition of daily and sports activities. Using two sequence sets based on the sensory data allows each unit to be placed at any position on a given rigid body part. As the unit is shifted from its ideal position with larger displacements, the activity recognition accuracy of the system that uses these sequence sets degrades slowly, whereas that of the reference system (which is not designed to achieve position invariance) drops very fast. Thus, we observe a tradeoff between the flexibility in sensor unit positioning and the classification accuracy. The reduction in the accuracy is at acceptable levels, considering the convenience and flexibility provided to the user in the placement of the units. We compare the proposed approach with an existing technique to achieve position invariance and combine the former with our earlier methodology to achieve orientation invariance. We evaluate our proposed methodology on a publicly available data set of daily and sports activities acquired by wearable motion sensor units. The proposed representations can be integrated into the preprocessing stage of existing wearable systems without significant effort.

This work investigates inter-subject and inter-activity variability of a given activity dataset and provides some new definitions to quantify such variability. The definitions are sufficiently general and can be applied to a broad class of datasets that involve time sequences or features acquired using wearable sensors. The study is motivated by contradictory statements in the literature on the need for user-specific training in activity recognition. We employ our publicly available dataset that contains 19 daily and sports activities acquired from eight participants who wear five motion sensor units each. We pre-process recorded activity time sequences in three different ways and employ absolute, Euclidean and dynamic time warping distance measures to quantify the similarity of the recorded signal patterns. We define and calculate the average inter-subject and inter-activity distances with various methods based on the raw and pre-processed time-domain data as well as on the raw and pre-processed feature vectors. These definitions allow us to identify the subject who performs the activities in the most representative way and pinpoint the activities that show more variation among the subjects. We observe that the type of pre-processing used affects the results of the comparisons but that the different distance measures do not alter the comparison results as much. We check the consistency of our analysis and results by highlighting some of our activity recognition rates based on an exhaustive set of sensor unit, sensor type and subject combinations. We expect the results to be useful for dynamic sensor unit/type selection, for deciding whether to perform user-specific training and for designing more effective classifiers in activity recognition.

This study provides a comparative assessment on the different techniques of classifying human activities performed while wearing inertial and magnetic sensor units on the chest, arms and legs. The gyroscope, accelerometer and the magnetometer in each unit are tri-axial. Naive Bayesian classifier, artificial neural networks (ANNs), dissimilarity-based classifier, three types of decision trees, Gaussian mixture models (GMMs) and support vector machines (SVMs) are considered. A feature set extracted from the raw sensor data using principal component analysis is used for classification. Three different cross-validation techniques are employed to validate the classifiers. A performance comparison of the classifiers is provided in terms of their correct differentiation rates, confusion matrices and computational cost. The highest correct differentiation rates are achieved with ANNs (99.2%), SVMs (99.2%) and a GMM (99.1%). GMMs may be preferable because of their lower computational requirements. Regarding the position of sensor units on the body, those worn on the legs are the most informative. Comparing the different sensor modalities indicates that if only a single sensor type is used, the highest classification rates are achieved with magnetometers, followed by accelerometers and gyroscopes. The study also provides a comparison between two commonly used open source machine learning environments (WEKA and PRTools) in terms of their functionality, manageability, classifier performance and execution times.

Three-dimensional models of environments can be very useful and are commonly employed in areas such as robotics, art and architecture, facility management, water management, environmental/industrial/urban planning and documentation. A 3-D model is typically composed of a large number of measurements. When 3-D models of environments need to be transmitted or stored, they should be compressed efficiently to use the capacity of the communication channel or the storage medium effectively. We propose a novel compression technique based on compressive sampling applied to sparse representations of 3-D laser range measurements. The main issue here is finding highly sparse representations of the range measurements, since they do not have such representations in common domains, such as the frequency domain. To solve this problem, we develop a new algorithm to generate sparse innovations between consecutive range measurements acquired while the sensor moves. We compare the sparsity of our innovations with others generated by estimation and filtering. Furthermore, we compare the compression performance of our lossy compression method with widely used lossless and lossy compression techniques. The proposed method offers a small compression ratio and provides a reasonable compromise between the reconstruction error and processing time.

In this study, surfaces with different properties are differentiated with simple low-cost infrared (IR) emitters and detectors in a location-invariant manner. The intensity readings obtained from such sensors are highly dependent on the location and properties of the surface, which complicates the differentiation and localization process. Our approach, which models IR intensity scans parametrically, can distinguish different surfaces independent of their positions. The method is verified experimentally with wood, Styrofoam packaging material, white painted wall, white and black cloth, and white, brown, and violet paper. A correct differentiation rate of 100% is achieved for six surfaces and the surfaces are localized within absolute range and azimuth errors of 0.2 cm and 1.1/spl deg/, respectively. The differentiation rate decreases to 86% for seven surfaces and to 73% for eight surfaces. The method demonstrated shows that simple IR sensors, when coupled with appropriate processing, can be used to differentiate different types of surfaces in a location-invariant manner.

We describe a new method for position estimation of planar surfaces using simple, low-cost infrared (IR) sensors. The intensity data acquired with IR sensors depends highly on the surface properties and the configuration of the sensors with respect to the surface. Therefore, in many related studies, either the properties of the surface are determined first or certain assumptions about the surface are made to estimate the distance and the orientation of the surface relative to the sensors. We propose a novel method for position estimation of surfaces with IR sensors without the need to determine the surface properties first. The method is considered to be independent of the type of surface encountered since it is based on searching the position of the maximum value of the intensity data rather than using absolute intensity values. The method is verified experimentally with planar surfaces of different surface properties. An intelligent feature of our system is that its operating range is made adaptive based on the maximum intensity of the detected signal. The absolute mean range error for the method resulting in the lowest errors is 0.15 cm over the range from 10 to 50 cm. The cases where the azimuth and elevation angles are nonzero are considered as well. The results obtained demonstrate that IR sensors can be used for localization to an unexpectedly high accuracy without prior knowledge of the surface characteristics.

This study investigates the use of low-cost infrared sensors in the differentiation and localization of commonly encountered target primitives in indoor environments, such as planes, corners, edges, and cylinders. The intensity readings from such sensors are highly dependent on target location and properties in a way which cannot be represented in a simple manner, making the differentiation and localization process difficult. In this paper, we propose the use of angular intensity scans and present an algorithm to process them. This approach can determine the target type independent of its position. Once the target type is identified, its position can also be estimated. The method is verified experimentally. An average correct classification rate of 97% over all target types is achieved and targets are localized within absolute range and azimuth errors of 0.8 cm and 1.6/spl deg/, respectively. The proposed method should facilitate the use of infrared sensors in mobile robot applications for differentiation and localization beyond their common usage as simple proximity sensors for object detection and collision avoidance.

This study investigates the processing of sonar signals using neural networks for robust differentiation of commonly encountered features in indoor environments. The neural network can differentiate more targets, and achieves high differentiation and localization accuracy, improving on previously reported methods. It achieves this by exploiting the identifying features in the differential amplitude and time-of-flight characteristics of these targets. An important observation follows from the robustness tests, which indicate that the amplitude information is more crucial than time-of-flight for reliable operation. The study suggests wider use of neural networks and amplitude information in sonar-based mobile robotics.

This study investigates the processing of sonar signals using neural networks for robust differentiation of commonly encountered features in indoor robot environments. The neural network can differentiate more targets with higher accuracy, improving on previously reported methods. It achieves this by exploiting the identifying features in the differential amplitude and time-of-flight (TOF) characteristics of these targets. Robustness tests indicate that the amplitude information is more crucial than TOF for reliable operation. The study suggests wider use of neural networks and amplitude information in sonar-based mobile robotics.

Presents a method for surface profile determination using multiple sensors. Our approach is based on morphological processing techniques to fuse the range data from multiple sensor returns in a manner that directly reveals the target surface profile. The method has the intrinsic ability of suppressing spurious readings due to noise, crosstalk, and higher-order reflections, as well as processing multiple reflections informatively. The algorithm is verified both by simulations and experiments in the laboratory by processing real sonar data obtained from a mobile robot. The results are compared to those obtained from a more accurate structured-light system, which is however more complex and expensive.

In this study, physical models are used to model reflections from target primitives commonly encountered in a mobile robot's environment. These targets are differentiated by employing a multi-transducer pulse/echo system which relies on both amplitude and time-of-flight data, allowing more robust differentiation. Target features are generated as being evidentially tied to degrees of belief which are subsequently fused by employing multiple logical sonars at different geographical sites. Feature data from multiple logical sensors are fused with the Dempster-Shafer rule of combination to improve the performance of classification by reducing perception uncertainty. Dempster-Shafer fusion results are contrasted with the results of combination of sensor beliefs through a simple majority vote. The method is verified by experiments with a real sonar system. The evidential approach employed here helps to overcome the vulnerability of the echo amplitude to noise and enables the modeling of non-parametric uncertainty in real time.

A novel, flexible, three-dimensional (3-D) multisensor sonar system is employed to localize the center of a spherical target and estimate its radius of curvature. The interesting limiting cases for the problem under study are the point and planar targets, both of which are important for the characterization of a mobile robot's environment. A noise model is developed based on real sonar data. An extended Kalman filter (EKF) which incorporates the developed noise model is employed as an estimation tool for optimal processing of the sensor data. Simulations and experimental results are provided for specularly reflecting cylindrical targets.

The performance of a commonly employed linear array of sonar sensors is assessed for point-obstacle localization intended for robotics applications. Two different methods of combining time-of-flight information from the sensors are described to estimate the range and azimuth of the obstacle: pairwise estimate method and the maximum likelihood estimator. The variances of the methods are compared to the Cramer-Rao lower bound, and their biases are investigated. Simulation studies indicate that in estimating range, both methods perform comparably; in estimating azimuth, maximum likelihood estimate is superior at a cost of extra computation. The results are useful for target localization in mobile robotics.

A low-cost solid-state inertial navigation system (INS) for mobile robotics applications is described. Error models for the inertial sensors are generated and included in an extended Kalman filter (EKF) for estimating the position and orientation of a moving robot vehicle. Two different solid-state gyroscopes have been evaluated for estimating the orientation of the robot. Performance of the gyroscopes with error models is compared to the performance when the error models are excluded from the system. Similar error models have been developed for each axis of a solid-state triaxial accelerometer and for a conducting-bubble tilt sensor which may also be used as a low-cost accelerometer. An integrated inertial platform consisting of three gyroscopes, a triaxial accelerometer and two tilt sensors is described.>

An error model for a solid-state gyroscope developed previously by the authors is included in a Kalman filter for improving the orientation estimate of a mobile robot. Orientation measurement with the error model is compared to the performance when no error model is incorporated in the system. The results demonstrate that without error compensation, the error in localization is between 5-15/spl deg//min but can be improved by a factor of 5 to 7 if an adequate error model is supplied. Results from tests of this gyroscope on a large outdoor mobile robot system are described and compared to the results obtained from the robot's own radar-based guidance system. Like all inertial systems, the platform requires additional information from some absolute position sensing mechanism to overcome long-term drift. However, the results show that with careful and detailed modelling of error sources, low cost inertial devices can provide valuable orientation and position information particularly for outdoor mobile robot applications.>

A low-cost, solid-state inertial navigation system for robotics applications is described. Error models for the inertial sensors are generated and included in an extended Kalman filter (EKF) for estimating the position and orientation of a moving robot vehicle. A solid-state gyroscope and an accelerometer have been evaluated. Without error compensation, the error in orientation is between 5-15/spl deg//min but can be improved at least by a factor of five if an adequate error model is supplied. Similar error models have been developed for each axis of a solid-state triaxial accelerometer. Linear position estimation with accelerometers and tilt sensors is more susceptible to errors due to the double integration process involved in estimating position. WIth the system described here, the position drift rate is 1-8 cm/s, depending on the frequency of acceleration changes. The results show that with careful and detailed modeling of error sources, low cost inertial sensing systems can provide valuable position information.

The authors investigate the prey capture problem in an uncluttered environment for a mobile robot equipped with a wide-beam sonar system. An analytical lower bound for the mean capture time is derived by assuming complete information about the prey. Algorithms using qualitative and quantitative information are implemented for prey capture, and compared in terms of mean capture time and capture probability. Although qualitative information is sufficient, quantitative information assures more efficient prey capture to be achieved over a broader range of system parameters. Experimental results verifying the analysis and the simulations are provided.>

The problem of docking two mobile robots using a wide-beam sonar system is considered. Be- cause of the similarity to biological bats, the prob- lem is discussed in terms of prey capture in two di- mensions. The basics of the bat-like sonar system are described. Two measures of performance are consid- ered: the capture probability and the mean capture time when capture occurs. These measures are com- puted in two ways: as the ratio of speeds of the prey to the pursuer is varied and as the strategies employing either qualitative information (prey is to the left or right) or quantitative information (range and azimuth to prey) are employed. The lower bound for the mean capture time is determined from game theory, which assumes complete information about the prey. The analysis is verifled by performing experiments with a real mobile robot. Both capture probability and mean capture time are inversely related to the prey/pursuer speed ratio. It is also observed that, while qualitative information is sumcient for docking to occur, quan- titative information allows successful docking over a larger range of speed ratios.

An active wide-beam sonar system that mimics the sensor configuration of echolocating bats is described for applications in sensor-based robotics. Obstacles in a two-dimensional (2-D) environment are detected and localized using time-of-flight (TOF) measurements of their echoes. The standard threshold detector produces a biased TOF estimate. An unbiased TOF estimate is derived by a parametric fit to the echo waveform, motivated by experimental observations of actual sonar signals. This novel method forms a tradeoff between the complexity of the optimum estimator and the biased threshold detector. Using the TOF information from both methods, the range and azimuth of an obstacle are estimated. Localization is most accurate if the obstacle is located along the system line-of-sight and improves with decreasing range. Standard deviations of the range and azimuth estimators are compared to the Cramer-Rao lower bounds. The parabolic fit method has large variance but zero bias at large deviations from the line-of-sight. The system operation is generalized from isolated obstacles to extended obstacles.>

A sonar system that simulates the sensor configuration of echolocating bats is described. Two estimators for range and azimuth of an obstacle are developed, based on the received signal envelope. The standard threshold detector produces a biased time-of-flight (TOF) estimate. An unbiased TOF estimate is derived by using the echo waveform shape to estimate the object location. The estimates are most accurate if the obstacle is located along the line-of-sight, and improve with decreasing range.>

A multitransducer, pulse/echo-ranging system is described that differentiates corner and plane reflectors by exploiting the physical properties of sound propagation. The amplitudes and ranges of reflected signals for the different transmitter and receiver pairs are processed to determine whether the reflecting object is a plane or a right-angle corner. In addition, the angle of inclination of the reflector with respect to the transducer orientation can be measured. Reflected signal amplitude and range values, as functions of inclination angle, provide the motivation for the differentiation algorithm. A system using two Polaroid transducers is described that correctly discriminates between corners and planes for inclination angles within +or-10 degrees of the transducer orientation. The two-transducer system is extended to a multitransducer array, allowing the system to operate over an extended range. An analysis comparing processing effort to estimation accuracy is performed.>

A procedure to navigate a vehicle from a source to a destination through an unstructured environment using only a sonar sensor is described. The environment considered is a two-dimensional floor plan, consisting of line segments of arbitrary size that is extended into the third dimension. To ensure that a sonar-guided vehicle does not collide with an obstacle, it is necessary to consider an obstacle that produce the weakest echoes and determine the conditions under which these obstacles can be detected. In this environment, the smallest echo is produced by the line that defines the boundary between the two planes, or the edge; and distant edges become invisible. This physical principle then becomes the basis for the navigation strategy, indicating the necessary scanning pattern and maximum step size that guarantee that no collision will occur. The approach is illustrated with results produced by a vehicular robot equipped with a Polaroid sensor. A representative floor plan and the automatically determined path are shown; a blind alley was included to observe the behavior of the algorithm.>