Disordered Quantum many-body Systems (DQS) and Quantum Neural Networks (QNN) have many structural features in common. However, a DQS is essentially an initialized QNN with random weights, often leading to non-random outcomes. In this work, we emphasize the possibilities of random processes being a deceptive quantum-generating model effectively hidden in a QNN. When we choose weights in a QNN randomly the unitarity property of quantum gates is unchanged. As we show, this can lead to memory effects with multiple consequences on the learnability and trainability of QNN one would not expect from a classical neural network with random weights. This phenomenon may lead to a fundamental misunderstanding of the capabilities of common quantum generative models, where the generation of new samples is essentially averaging over random outputs. While we suggest that DQS can be effectively used for tasks like image augmentation, we draw the attention that overly simple datasets are often used to show the generative capabilities of quantum models, potentially leading to overestimation of their effectiveness.

Human activity recognition (HAR) is a long-standing problem in artificial intelligence with applications in a broad range of areas, including healthcare, sports and fitness, security, and more. The performance of HAR in real-world settings is strongly dependent on the type and quality of the input signal that can be acquired. Given an unobstructed, high-quality camera view of a scene, computer vision systems, in particular in conjunction with foundation models, can today fairly reliably distinguish complex activities. On the other hand, recognition using modalities such as wearable sensors (which are often more broadly available, e.g., in mobile phones and smartwatches) is a more difficult problem, as the signals often contain less information and labeled training data is more difficult to acquire. To alleviate the need for labeled data, we introduce our comprehensive Fitness Multimodal Activity Dataset (FiMAD) in this work, which can be used with the proposed pre-training method MuJo (Multimodal Joint Feature Space Learning) to enhance HAR performance across various modalities. FiMAD was created using YouTube fitness videos and contains parallel video, language, pose, and simulated IMU sensor data. MuJo utilizes this dataset to learn a joint feature space for these modalities. We show that classifiers pre-trained on FiMAD can increase the performance on real HAR datasets such as MM-Fit, MyoGym, MotionSense, and MHEALTH. For instance, on MM-Fit, we achieve a Macro F1-Score of up to 0.855 when fine-tuning on only 2% of the training data and 0.942 when utilizing the complete training set for classification tasks. We compare our approach with other self-supervised ones and show that, unlike them, ours consistently improves compared to the baseline network performance while also providing better data efficiency.

Head-mounted displays (HMDs) are increasingly used in safety-critical fields such as surgery, aviation, and industrial manufacturing. As major manufacturers shift toward video-see through (VST) designs to deliver unified AR and VR experiences, they also replace direct visual access to the real world with a video feed. This design choice raises concerns about its impact on user performance. This study investigates the isolated impacts of VST and optical-see through HMDs on user real-world perceptual–motor performance by comparing two leading HMDs, Apple Vision Pro (AVP) and HoloLens 2 (MHL2) against unencumbered vision using the Purdue Pegboard Test (PPT), a standard assessment of manual dexterity. Twenty participants completed tasks across three conditions (AVP, MHL2, and Baseline), while we recorded dexterity scores, cognitive load, system usability, VR sickness, and subjective feedback. Movement data were also collected via Apple Watches. Study results with 20 participants revealed that dexterity scores significantly declined under the AVP condition across all subtests. This was accompanied by significantly higher cognitive load and a notable drop in RMS acceleration values (observed in the RMS analysis of a subset of 13 participants). The analysis on dexterity score yielded a significant difference between MHL2 and Baseline only for a single subtest of the PPT (Left Hand). Post-task interviews revealed greater discomfort, visual fatigue, and reduced task confidence with AVP. These findings suggest that current VST HMDs impose a hidden ergonomic cost undermining user performance in tasks where precision, and comfort are essential. For AR applications designed to enhance user performance, such as assistive tools, training systems, or task guidance interfaces, designers must account for and mitigate this performance degradation through counterbalancing strategies to offset the visual and cognitive burden introduced by VST HMDs.

We present a quantum inspired image augmentation protocol which is applicable to classical images and, in principle, due to its known quantum formulation applicable to quantum systems and quantum machine learning in the future. The augmentation technique relies on the phenomenon Anderson localization. As we will illustrate by numerical examples the technique changes classical wave properties by interference effects resulting from scatterings at impurities in the material. We explain that the augmentation can be understood as multiplicative noise, which counter-intuitively averages out, by sampling over disorder realizations. Furthermore, we show how the augmentation can be implemented in arrays of disordered waveguides with direct implications for an efficient optical image transfer.

Automatic and precise fitness activity recognition can be beneficial in aspects from promoting a healthy lifestyle to personalized preventative healthcare. While IMUs are currently the prominent fitness tracking modality, through iMove, we show bio-impedence can help improve IMU-based fitness tracking through sensor fusion and contrastive learning. To evaluate our methods, we conducted an experiment including six upper body fitness activities performed by ten subjects over five days to collect synchronized data from bio-impedance across two wrists and IMU on the left wrist. The contrastive learning framework uses the two modalities to train a better IMU-only classification model, where bio-impedance is only required at the training phase, by which the average Macro F1 score with the input of a single IMU was improved by 3.22 % reaching 84.71 % compared to the 81.49 % of the IMU baseline model. We have also shown how bio-impedance can improve human activity recognition (HAR) directly through sensor fusion, reaching an average Macro F1 score of 89.57 % (two modalities required for both training and inference) even if Bio-impedance alone has an average macro F1 score of 75.36 %, which is outperformed by IMU alone. In addition, similar results were obtained in an extended study on lower body fitness activity classification, demonstrating the generalisability of our approach.Our findings underscore the potential of sensor fusion and contrastive learning as valuable tools for advancing fitness activity recognition, with bio-impedance playing a pivotal role in augmenting the capabilities of IMU-based systems.

Manufacturing industries strive to improve production efficiency and product quality by deploying advanced sensing and control systems. Wearable sensors are emerging as a promising solution for achieving this goal, as they can provide continuous and unobtrusive monitoring of workers’ activities in the manufacturing line. This article presents a novel wearable sensing prototype that combines IMU and body capacitance sensing modules to recognize worker activities in the manufacturing line. To handle these multimodal sensor data, we propose and compare early, and late sensor data fusion approaches for multichannel time-series convolutional neural networks and deep convolutional LSTM. We evaluate the proposed hardware and neural network model by collecting and annotating sensor data using the proposed sensing prototype and Apple Watches in the testbed of the manufacturing line. Experimental results demonstrate that our proposed methods achieve superior performance compared to the baseline methods, indicating the potential of the proposed approach for real-world applications in manufacturing industries. Furthermore, the proposed sensing prototype with a body capacitive sensor (BCS) and feature fusion method improves by 6.35%, yielding a 9.38% higher macro F1 score than the proposed sensing prototype without a BCS and Apple Watch data, respectively.

In this work, we propose an open-source scalable end-to-end RTL framework FieldHAR, for complex human activ-ity recognition (HAR) from heterogeneous sensors using artificial neural networks (ANN) optimized for FPGA or ASIC integration. FieldHAR aims to address the lack of apparatus to transform complex HAR methodologies often limited to offline evaluation to efficient runtime edge applications. The framework uses parallel sensor interfaces and integer-based multi-branch convolutional neural networks (CNNs) to support flexible modality extensions with synchronous sampling at the maximum rate of each sensor. To validate the framework, we used a sensor-rich kitchen scenario HAR application which was demonstrated in a previous offline study. Through resource-aware optimizations, with FieldHAR the entire RTL solution was created from data acquisition to ANN inference taking as low as 25% logic elements and 2% memory bits of a low-end Cyclone IV FPGA and less than 1% accuracy loss from the original FP32 precision offline study. The RTL implementation also shows advantages over MCU-based solutions, including superior data acquisition performance and virtually eliminating ANN inference bottleneck.

User dependence remains one of the most difficult general problems in Human Activity Recognition (HAR), in particular when using wearable sensors. This is due to the huge variability of the way different people execute even the simplest actions. In addition, detailed sensor fixtures and placement will be different for different people or even at different times for the same users. In theory, the problem can be solved by a large enough data set. However, recording data sets that capture the entire diversity of complex activity sets is seldom practicable. Instead, models are needed that focus on features that are invariant across users. To this end, we present an adversarial subject-independent feature extraction method with the maximum mean discrepancy (MMD) regularization for human activity recognition. The proposed model is capable of learning a subject-independent embedding feature representation from multiple subjects datasets and generalizing it to unseen target subjects. The proposed network is based on the adversarial encoder-decoder structure with the MMD to realign the data distribution over multiple subjects. Experimental results show that the proposed method not only outperforms state-of-the-art methods over the four real-world datasets but also improves the subject generalization effectively. We evaluate the method on well-known public data sets showing that it significantly improves user-independent performance and reduces variance in results.

In this work, we present the design and implementation of a micro watt level power consumption, human body capacitance based sensor for recognizing and counting gym workouts. The concept also works when the device is attached to a body part which is not directly involved in the activity's movement. In contrast, most of the widely used motion sensing based approaches require placing the sensor on the moving body part (e.g. for analyzing leg based gym exercises the sensor needs to be placed on the leg). We described the physical principle behind the ubiquitous electric coupling between human body and environment, and explored the capability of this sensing modality in gym workouts. We evaluated our sensor with 11 subjects, performing 7 popular gym workouts each day over 5 days with our sensor being placed at 3 different body positions, including a non-contact position, where the sensor is placed in the subject's pocket. Results showed that our sensing approach achieved an average counting accuracy of 91%, which is highly competitive with commercial devices on the market. The mean leave one user out workout recognition f-scores obtained were of 63%, 56%, 45% for sensors located on wrist, on calf and in pocket, respectively. As every subject performed activities over multiple days changing shoe height, shoe and clothes type, we demonstrate that full body activity counting and to some extent recognition is feasible, regardless of personal habit of movement speed and scale.

Proper place recognition on an environment that can change over time is fundamental for long-term SLAM. In such scenarios the observations obtained in the same region can drastically differ due to changes caused by semi-static objects, such as doors, furniture, etc. In this work, we extend a strategy that represents environment regions using words, based on spatial density information extracted from laser readings. This time, in order to deal with changes in the environment, our method not only builds words representing the real observations made by the robot, but also alternative multi-level words to account for possible changes in a place's observations generated by non-static objects. Place recognition is made by searching matches of sequences of N consecutive words (both real or alternatives). Experiments performed in real and simulated scenarios are shown, and demonstrate the advantages associated to the use of multi-level words.

Estimating the robot localization is a fundamental requirement for applications in robotics. For many years, Monte Carlo Localization (MCL) has been one of the most popular approaches to solve the global localization when using range finders, like sonars or lasers. It generally weights the estimates about the robot state by comparing raw sensor readings with simulated readings computed for each estimate. In this paper, we propose an observation model for localization that associates a kernel density estimate (KDE) to each point in the space. This single-valued density measure is independent of orientation, what allows an efficient pre-caching step, substantially boosting the computation time of the process. Using the gradient of the densities field, our strategy is able to estimate orientation information that helps to restrict the localization search space. Additionally, we can combine densities obtained by kernels of different sizes and profiles to improve the quality of the acquired information. We show through experiments in comparison with traditional approaches that our method is efficient, even working with large sets of particles, and effective.

Place recognition is the frond-end of Simultaneous Localization and Mapping (SLAM). Topological representations depend on good association of vertices, which ultimately depends on the front-end. In this paper, we consider a robot lost in an unknown environment trying to construct a topological map to localize itself using a laser range finder and odometry information. The algorithm makes use of an efficient observation model based on kernel density estimates (KDEs) to detect loops. The observation model separates the map into regions denominated words, classified based on the density of free space, number of observations and segment orientation. Loop closing results from the matching of sequences of N consecutive words (n-grams). The proposed approach is orders of magnitude faster than a sequence of Iterative Closest Point (ICP) matches. The method is evaluated varying input parameters in real and simulated scenarios.