Automatic fall detection using radar technology significantly advances assisted living and smarter healthcare solutions. In this paper, a novel unsupervised method for detecting fall incidents in human daily activities is proposed. By analyzing ultra-wideband radar returns, a radar time series is obtained and used to find the time-frequency signatures of different activities. These signatures are subsequently binarized and used as input to a deep stacked auto-encoder network for latent feature extraction. The latent features are then clustered through a clustering layer that leverages an auxiliary distribution function to fine-tune the samples clustered into fall and non-fall groups. The proposed fall clustering method is compared against several clustering approaches in terms of accuracy of clustered samples, normalized mutual information and adjusted rand index metrics. It is shown that the proposed method realizes automatic latent feature learning from the radar data and can distinguish fall from non-fall classes by uncovering patterns in a data set with no pre-existing labels. The results show that the proposed unsupervised fall detection method outperforms the other approaches in terms of providing higher accuracy of the clustered data samples. The advantage of the proposed method is that it does not need a large dataset to achieve distinctiveness between clusters.

Dependable visual drone detection is crucial for the secure integration of drones into the airspace. However, drone detection accuracy is significantly affected by domain shifts due to environmental changes, varied points of view, and background shifts. To address these challenges, we present the DrIFT dataset, specifically developed for visual drone detection under domain shifts. DrIFT includes fourteen distinct domains, each characterized by shifts in point of view, synthetic-to-real data, season, and adverse weather. DrIFT uniquely emphasizes background shift by providing background segmentation maps to enable background-wise metrics and evaluation. Our new uncertainty estimation metric, MCDO-map, features lower postprocessing complexity, surpassing traditional methods. We use the MCDO-map in our uncertainty-aware unsupervised domain adaptation method, demonstrating superior performance to SOTA unsupervised domain adaptation techniques. The dataset is available at: https://github.com/CARG-uOttawa/DrIFT.git.

Personalized monitoring systems (PMS) are used for Decision Making (DeM) to support humans and fully autonomous decision-making in several applications, such as health monitoring and management. However, it is still challenging to design decision-making in PMS. In this work, we propose a systematic modeling approach, called DEMDE, for decision-making design in PMS during the design phase. DEMDE guides the design of a specific Bayesian network (BN) from an instantiated domain model for PMS based on context-aware data fusion using a probabilistic domain model (general BN). We evaluated our proposal by developing a BN for decision-making about sending an alert of high contamination risk in cell culture. The case study demonstrated the application of the DEMDE concepts and the modeling process, including model evaluation through sensitivity analysis to assess the robustness and reliability of the modeled BN.

Depth completion and object detection are two crucial tasks often used for aerial 3D mapping, path planning, and collision avoidance of Uncrewed Aerial Vehicles (UAVs). Common solutions include using measurements from a LiDAR sensor; however, the generated point cloud is often sparse and irregular and limits the system's capabilities in 3D rendering and safety-critical decision-making. To mitigate this challenge, information from other sensors on the UAV (viz., a camera used for object detection) is utilized to help the depth completion process generate denser 3D models. Performing both aerial depth completion and object detection tasks while fusing the data from the two sensors poses a challenge to resource efficiency. We address this challenge by proposing a novel approach to jointly execute the two tasks in a single pass. The proposed method is based on an encoder-focused multi-task learning model that exposes the two tasks to jointly learned features. We demonstrate how semantic expectations of the objects in the scene learned by the object detection pathway can boost the performance of the depth completion pathway while regressing the missing depth values. Experimental results show that the proposed multi-task network outperforms its single-task counterpart, particularly when exposed to defective inputs.

This paper presents the 4-th edition of the “drone-vs-bird” detection challenge, launched in conjunction with the the 17-th IEEE International Conference on Advanced Video and Signal-based Surveillance (AVSS). The objective of the challenge is to tackle the problem of detecting the presence of one or more drones in video scenes where birds may suddenly appear, taking into account some important effects such as the background and foreground motion. The proposed solutions should identify and localize drones in the scene only when they are actually present, without being confused by the presence of birds and the dynamic nature of the captured scenes. The paper illustrates the results of the challenge on the 2021 dataset, which has been further extended compared to the previous edition run in 2020.

In Drone-vs-Bird Detection Challenge in conjunction with the 4th International Workshop on Small-Drone Surveillance, Detection and Counteraction Techniques at IEEE AVSS 2021, we proposed a YOLOV5-based object detection model for small UAV detection and classification. YOLOV5 leverages PANet neck and mosaic augmentation which help in improving detection of small objects. We have combined the challenge dataset with one of the publicly available UAV air to air dataset having complex background and lighting conditions for training the model. The proposed approach achieved 0.96 Recall, $0.98 mAP_{0.5}$, and $0.71 mAP_{0.5:0.95}$ on the 10% randomly sampled dataset from the whole dataset.

Automatic detection of a falling person based on noncontact sensing is a challenging problem with applications in smart homes for elderly care. In this article, we propose a radar-based fall detection technique based on time-frequency analysis and convolutional neural networks. The time-frequency analysis is performed by applying the short-time Fourier transform to each radar return signal. The resulting spectrograms are converted into binary images, which are fed into the convolutional neural network. The network is trained using labeled examples of fall and nonfall activities. Our method employs high-level feature learning, which distinguishes it from previously studied methods that use heuristic feature extraction. The performance of the proposed method is evaluated by conducting several experiments on a set of radar return signals. We show that our method distinguishes falls from nonfalls with 98.37% precision and 97.82% specificity, while maintaining a low false-alarm rate, which is superior to existing methods. We also show that our proposed method is robust in that it successfully distinguishes falls from nonfalls when trained on subjects in one room, but tested on different subjects in a different room. In the proposed convolutional neural network, the hierarchical features extracted from the radar return signals are the key to understand the fundamental composition of human activities and determine whether or not a fall has occurred during human daily activities. Our method may be extended to other radar-based applications such as apnea detection and gesture detection.

This paper analyzes the structure of the feature space of radar data collected from real subjects either still or in motion and provides dimensionality reduction and modeling through genetic programming. Three movement classes: sedentary and still, sedentary with movements, and walking contained in the returns obtained from a single channel continuous wave phase-modulated radar are considered. Unsupervised methods are used for finding the intrinsic dimensionality of the space of the original features and nonlinear mappings are used to obtain lower dimensional representations. The classification results for the original and the reduced dimension data are similar, thus, indicating the redundancy of the eliminated features. The white-box models obtained through genetic programming is then compared with the conventional black-box models obtained through supervised classification using random trees, extreme learning machines and multilayer perceptron. For this problem, the explicit white-box models obtained with genetic programming produced equal or better classification accuracies than those obtained with black-box approaches. In addition to explainability, the genetic programming models found have the additional advantage of involving only a few relevant predictors, exhibiting good feature selection capabilities.

The advancement of UAS (Unmanned Aircraft Systems) technologies, and the rise in potential misuse of these vehicle platforms, and in particular small UAS (sUAS), has highlighted the demand for a robust and reliable solution for detection and classification of the aircraft (commonly referred to by Drone)vs. other flying objects. Most of the existing research addresses this problem either by extracting micro-Doppler from radar data or features from visual data. But these solutions do not perform well in all weather conditions and beyond a particular distance. To solve the problem of classifying small UASs and differentiating them from the birds, we propose a novel approach by merging classical Interactive Multiple Model tracking (IMM) filter and state of the art Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) unit. The IMM extracts the kinematic features of the target flight trajectory and the RNN with LSTM learn the complex sequence of flight maneuvers. We generated synthetic trajectories emulating birds and drones flights using 3D kinematic models for training. The paper demonstrates that the classification accuracy of 99.3% was achieved with five-fold cross-validation on a network with convolutional layers, LSTM layers, and the dense output layer.

The Internet of Things (IoT) technology presents a complex and challenging paradigm where a huge amount of noisy raw sensor data is collected in order to observe and detect critical events occurring on the system, and generate alarms when required. The biggest challenge of the IoT systems is that the systems collect a massive amount of uncertain data from diverse IoT devices connected through the network. In addition, some events are inferred from other events and uncertainty is propagated from parent events to the inferred events, which additionally contributes to overall system uncertainty. The observed complex events are a complex relationship of primitive events that are produced by IoT devices and collected in IoT systems. A survey performed on existing prior arts on quantifying uncertainty for complex events concluded that proposed existing solutions are unable to scale under heavy loads of incoming data. This paper presents a micro-service based notification methodology that uses complex event recognition (both complex event processing and probabilistic programming) to handle IoT systems uncertainty. In addition, the paper analyzes and recommends existing big data platforms for processing complex events in IoT systems. The current focus of our work includes research and development of the optimized deadline-based and cost-effective resource allocation algorithm in Apache Spark for Uncertain IoT Notification systems.

Detecting falls using radar has many applications in smart health care. In this paper, a novel method for fall detection in human daily activities using an ultra wideband radar technology is proposed. A time series derived from the radar scattering matrix is used as input to the the residual network for automatic feature extraction. In contrast to other existing methods, the proposed method relies on multi-level feature learning directly from the radar time series signals. In particular, the proposed method utilizes a deep residual neural network for automating feature learning and enhancing model discriminability. The performance of the proposed method is compared with that of the other methods such as support vector machine, K-nearest neighbors, multi-layer perceptron and dynamic time warping techniques. The results show that the proposed fall detection method outperforms the other methods in terms of accuracy and sensitivity values.

Classification is presented as a pre‐processing step in this study. The state of the subject is classified as the unmoving state with normal breathing (normal breathing class), unmoving state with no breathing (stop breathing class) or the state when the subject is moving (erratic signal class) before breathing estimation algorithms are applied. Estimation algorithms may be applied to obtain breathing rate if normal breathing class is detected or alarms may be generated if stop breathing is detected, and fine‐grained classification of activities may be pursued if the erratic signal is detected. Experiments were performed using a single‐channel pulse‐modulated continuous wave radar with three subjects for a total of 135 min. In each experiment, the subject was continuously monitored for 15 min and the subject performed activities that resulted in a signal that belonged to one of the three classes. Besides extracting a feature that assessed the distribution of energy of the signal in the frequency domain, a novel nonlinear time series feature extraction method based on the higher‐dimensional embedding technique was applied to ascertain periodicity of the reflected signal. Bayes classifier was used to classify each 5‐s segment of radar returns. A 30‐fold cross validation resulted in 97% of overall classification accuracy.

In real life applications, it is crucial to monitor the different kinds of human activity without interfering with their regular occupations. Contactless physiological monitoring using radars is a valuable tool, but typically it is performed when the human subjects are immobile. This paper analyzes single channel Continuous Wave (CW) Doppler radar signals in relation to three levels of human activity: i) Sedentary and still, ii) Sedentary and moving and iii) Walking. A combination of computational intelligence techniques (GammaTest, neural networks, random forest and genetic algorithms) was used for assessing the predictive ability of 43 features derived from the radar return signal, as well as of subsets of them, which were composed of highly predictive attributes. It is shown that with about one half the number of attributes it is possible to achieve high levels of classification accuracy, in some cases improving false negative ratios. While several attributes were completely irrelevant and noisy, others were required by discriminating each of the classes. There are attributes required by certain classes in particular and there are others associated to the distinction of classes with subtle differences.

Noninvasive blood pressure (BP) measurement is an important tool for managing hypertension and cardiovascular disease. However, automated noninvasive BP measurement devices, which are usually based on the oscillometric method, do not always provide accurate estimation of BP. It has been found that change in arterial stiffness (AS) is an underlying mechanism of disagreement between an oscillometric BP monitor and a sphygmomanometer. This problem is addressed by incorporating parameters related to AS in the algorithm for BP measurement. Pulse transit time (PTT) is first used to estimate AS parameters, which are fixed into a model of the oscillometric envelope. This model can then be used to perform curve fitting to the measured signal using only four parameters: systolic BP, diastolic BP, mean BP, and lumen area at zero transmural pressure. The proposed technique is independent of the experimentally determined characteristic ratios that are commonly used in existing oscillometric methods. The accuracy of the proposed technique was evaluated by comparing with the same model without incorporation of AS, and with reference BP device measurements. The new method achieved standard deviation of error less than 8 mmHg and mean error less than 5 mmHg. The results show consistency with ANSI/AAMI SP-10 standard for noninvasive BP measurement techniques.

This paper presents a novel algorithm for radar-based estimation of vital signs in a noncontact, privacy friendly manner while subjects are in motion. Unlike the traditional methods that merely use the Fourier spectrum of the output of the radar receiver to obtain estimates of breathing and heart rates, the proposed algorithm uses time-frequency approach. From the Time-Frequency Representation (TFR) of the output of a pseudo-random binary Phase Modulated Continuous Wave (PMCW) radar, frequency of the maximum amplitude at every time instant is estimated and a timeseries of dominant frequencies is formed. MUSIC algorithm is then applied to estimate the vital signs from this series. The proposed algorithm is demonstrated using simulated and real data. Simulated data is obtained through modeling the output of a PMCW radar. Real data is obtained by monitoring a walking subject for 10 minutes in a realistic setting with a 24.125 GHz PMCW radar. The vital sign estimates obtained using the proposed method are found to match closely the estimates from wearable devices that were applied to provide the ground truth for breathing and heart rates.

Tag anticollision is critical to the performance of many radio frequency identification systems in industrial applications. Tree-based schemes are popular arbitration algorithms for tag collision, as they are scalable and easy to implement. In these schemes, collided tags are recursively divided into two groups until each group contains only one tag. However, we have determined that previous pure tree-based schemes do not fully utilize the information that can be collected; they allow only the responding tags to be involved in the splitting process. In this paper, we propose a novel pure tree-based method that we named parallel splitting with retrieve (PSR), which allows both the nonresponding and unidentified tags to also be involved in the splitting process. PSR consists of two mechanisms: parallel splitting (PS) and retrieve. The PS mechanism can reduce the total number of tag collisions and the retrieve mechanism can reduce the extra idle slots caused by PS. In comparison with conventional pure tree-based schemes, the proposed PSR algorithm achieves higher efficiency without the need to estimate the number of tags; simulation results indicate an efficiency level of 0.41. The PSR algorithm also has similar performance to state-of-the-art algorithms that do not need to estimate the number of tags, but it is simpler to implement and is thus preferred for resource constrained systems.

This paper describes a new approach to accelerate the simulation of the steady-state response of nonlinear circuits using the harmonic-balance (HB) technique. The approach presented in this paper focuses on the direct factorization of the Jacobian matrix, of the HB nonlinear equations, using a graphical processing unit (GPU) platform. The computational core of the proposed approach is based on developing a block-wise version of the KLU factorization algorithm, where scalar arithmetic operations are replaced by block-aware matrix operations. For a large number of harmonics, or excitation tones, or both, the Block-KLU (BKLU) approach effectively raises the ratio of floating-point operations to other operations and, therefore, becomes an ideal vehicle for implementation on a GPU-based platform. Motivated by this fact, we develop a GPU-based framework to implement the BKLU. The proposed approach yields speedup by up to 89 times over conventional direct factorization on CPU.

In radio frequency identifications (RFIDs), an idle slot can be interrupted by the reader and takes less time than the other slots. In this case, the widely used binary splitting (BS) protocol is not optimal.$M$-ary splitting ($M>2$) with a carefully selected$M$outperforms BS, because it can take into account smaller duration of the idle slots. In this paper, we provide optimal value of$M$($M_{opt}$) for varying slot sizes that yields minimum identification time. The proposed scheme is called optimal splitting (OS). We also revise OS to deal with mobile tags and propose the optimal splitting with arrival tags blocking (OS-ATB) protocol. Extensive simulations show that OS and OS-ATB consistently yield better performance than the other splitting and tree-based schemes.

In this correspondence, a unified approach to the autoregressive (AR) modeling of power spectral densities is described. We show that by introducing auxiliary sequences, the minimization of several customary spectral distances can be performed with the exact same convenient approach, whether a gradient-descent or a Newton/quasi-Newton descent is chosen. Moreover, we extend the usual optimization of unnormalized AR coefficients to a two-step optimization of normalized AR coefficients, and provide evidence that this alternative approach can accelerate convergence and provide robustness to erroneous initializations. Convergence and modeling results are also given.

In this paper we present a novel Radio Frequency Identification (RFID) system for accurate indoor localization. The system is composed of a standard Ultra High Frequency (UHF), ISO-18006C compliant RFID reader, a large set of standard passive RFID tags whose locations are known, and a newly developed tag-like RFID component that is attached to the items that need to be localized. The new semi-passive component, referred to as sensatag (sense-a-tag), has a dual functionality wherein it can sense the communication between the reader and standard tags which are in its proximity, and also communicate with the reader like standard tags using backscatter modulation. Based on the information conveyed by the sensatags to the reader, localization algorithms based on binary sensor principles can be developed. We present results from real measurements that show the accuracy of the proposed system.

In particle filtering, resampling is the only step that cannot be fully parallelized. Recently, we have proposed algorithms for distributed resampling implemented on architectures with concurrent processing elements (PEs). The objective of distributed resampling is to reduce the communication among the PEs while not compromising the performance of the particle filter. An additional objective for implementation is to reduce the communication among the PEs. In this paper, we report an improved version of the distributed resampling algorithm that optimally selects the particles for communication between the PEs of the distributed scheme. Computer simulations are provided that demonstrate the improved performance of the proposed algorithm.

In this paper, we derive a theoretical performance evaluation scheme of Kalman filter based channel tracking and data decoding for multiple-input multiple-output orthogonal frequency division multiplexed (MIMO-OFDM) communication systems that are based on orthogonal space-time block codes. The derivation is approximate, however, it is novel and demonstrated accurate for practical scenarios. Assuming a prior distribution for the initial channel we have derived the instantaneous signal to interference and noise ratio (SINR) for consecutive transmission blocks in the absence of training by exploiting Kalman filtering to track the channel. A theoretical estimation of BER is then derived based on such instantaneous SINR values. The resulting analysis is able to study the effect of different parameters of the system such as the number of antennas, number of sub-carriers, mobile velocity and the assumed channel length on the BER performance of the system. Numerical examples confirm the validity of the theoretical analysis.

This paper presents a particle filter approach to spectral amplitude speech enhancement. Spectral amplitudes are known to exhibit inter-frame dependencies and non-Gaussian statistics; however, incorporating these properties makes closed-form solutions intractable. Using the particle filter framework allows the presented algorithm to model the speech spectral amplitudes as an autoregressive process with Laplace distributed excitation. Two variants of the standard algorithm are also presented: one that uses an interacting multiple model approach to account for transitions between active speech and silence intervals, and one that allows for phase differences between the clean speech and noise complex Fourier transform coefficients. All of the particle sampling distributions are constrained to take the measurement into account, improving sampling efficiency. In experiments using wideband speech and real recorded noise the proposed algorithm variants are shown to offer natural-sounding output speech, with objective evaluation results that compare favorably to existing particle filter speech enhancement algorithms. The multiple model variant is found to improve inter-speech noise reduction, while the phase variant improves performance when the signal-to-noise ratio is low.

In this paper, blind decoding of OSTBC based MIMO systems using sequential Monte-Carlo methods is considered. Similar receivers developed earlier have suffered from the high computational complexity. By introducing some simplification and approximation techniques, we present a new decoding algorithm that requires much less computation. In addition to that, an efficient interlaced maximum likelihood estimator approach is developed to blindly estimate the statistical parameters of the channel. While a comparative analysis on the computational complexity shows significant computational savings, the simulation results show that the performance of the proposed algorithm is not compromised.

Motivated by the reportedly strong performance of particle filters (PFs) for noise reduction on essentially linear speech production models, and the mounting evidence that the introduction of nonlinearities can lead to a refined speech model, this paper presents a study of PF solutions to the problem of speech enhancement in the context of nonlinear, neural-type speech models. Several variations of a global model are presented (single/multiple neurons; bias/no bias), and corresponding PF solutions are derived. Different importance functions are given when beneficial, Rao-Blackwellization is proposed when possible, and dual/nondual versions of each algorithms are presented. The method shown can handle both white and colored noise. Using a variety of speech and noise signals and different objective quality measures, the performance of these algorithms are evaluated against other PF solutions running on linear models, as well as some traditional enhancement algorithms. A certain hierarchy in performance is established between each algorithm in the paper. Depending on the experimental conditions, the best-performing algorithms are a classical Rao-Blackwellized particle filter (RBPF) running on a linear model, and a proposed PF employing a nondual, nonlinear model with multiple neurons and no biases. With consistence, the neural-network-based PF outperforms RBPF at low signal-to-noise ratio (SNR).

Acoustic echo paths can change frequently in response to minor changes in the near-end room. Using echo path measurements obtained using commercial wideband hands-free terminals, it is demonstrated that the effects of these changes are not uniform across time and frequency. A subband echo canceler combined with a proportionate step-size adaptation algorithm is proposed to exploit time and frequency response characteristics of echo path changes to achieve faster tracking and higher echo attenuation. Improvements of 510 dB compared to fullband structures are observed for real changing echo environments with wideband speech inputs.

In this paper, distributed arithmetic (DA) has been used to implement a fully parallel LUT-based DA wavelet filterbank with interlaced input registers. In our scheme, decimation has been seamlessly integrated into the filter structure to achieve the same throughput performance as polyphase-based filterbanks. However, because partitioning of the filters is avoided, our scheme gives more flexibility to implement the LUT-DA structure, and consequently, lets designers maximize area utilization on LUT-based FPGAs. Our architecture has been designed for orthonormal and biorthogonal wavelets, and implemented on an Altera Stratix II FPGA. Significant reduction in terms of area requirements and increased throughput performance are achieved when compared to other DWT filterbanks based on DA, convolution or the lifting scheme.

Among the various types of existing speech enhancement algorithms, respective advantages are usually well-known and well-studied, making it possible to choose the right algorithm given a certain type of audio quality requirement (e.g. background noise intrusiveness, naturalness of the speech, etc.). Very little has however been said so far on the quality of speech enhanced by particle filters (PFs). In this paper, we show the detailed results of the analysis of PF-enhanced speech signals. This analysis is conducted using several objective measures, and is based on the comparison with other enhancement algorithms. We find that in general, PF-based algorithms yield speech signals which are among the most natural-sounding, with a residual noise that can be seen as a white noise with its variance "modulated" by the resulting speech.

Rao-Blackwellised Particle Filters (RBPFs) are a class of Particle Filters (PFs) that exploit conditional dependencies between parts of the state to estimate. By doing so, RBPFs can improve the estimation quality while also reducing the over-all computational load in comparison to original PFs. However, the computational complexity is still too high for many real-time applications. In this paper, we propose a modified RBPF that requires a single Kalman Filter (KF) iteration per input sample. Comparative experiments show that while good convergence can still be obtained, computational efficiency is always drastically increased, making this algorithm an option to consider for real-time implementations.

A novel low-complexity residual resampling scheme for particle filters is presented. The proposed scheme uses a simple but effective "particle-tagging" method to compensate for a possible error that can be caused by finite-precision quantization in the resampling step of particle filtering. The scheme guarantees that the number of particles after resampling is always equal to the number of particles before resampling. The resulting scheme is suitable for high-speed physical realization when the number of particles is a power of two.

Resampling is a critically important operation in the implementation of particle filtering. In parallel hardware implementations, resampling becomes a bottleneck due to its sequential nature and the increased complexity it imposes on the traffic of the designed interconnection network. To circumvent some of these difficulties, we propose two new resampling algorithms. The first one, called residual-systematic resampling, combines the merits of both systematic and residual resampling and is suitable for pipelined implementation. It also guarantees the fixed duration of the resampling procedure irrespective of the weight distribution of the particles. The second algorithm, referred to as partial resampling, has low complexity and reduces traffic load through the hardware network. These two algorithms should also be considered as resampling methods in simulations on standard computers.