Youngok Kim

Accurate distance estimation of visually small objects using consumer-grade monocular cameras remains challenging when targets occupy only a limited number of pixels in the image. Under such extreme small-object conditions, sparse visual information and geometric projection sensitivity significantly degrade both object localization reliability and ranging accuracy. This problem is particularly critical in consumer electronic vision systems, such as smartphone cameras, smart monitoring devices, and consumer mobile robots, where lightweight monocular sensors are widely adopted. This paper proposes a monocular vision framework that jointly addresses detection uncertainty and geometric estimation sensitivity under pixel-sparse conditions. Unlike conventional approaches that treat detection and distance estimation as independent modules, the proposed framework explicitly considers their interaction. To improve detection robustness, in this paper, a Small Object Detection Transformer Framework (SODTF) is proposed, which integrates multi-branch feature extraction with transformer-based global contextual modeling to enhance feature representation for visually small targets. Based on the detected bounding boxes, an Inclined-Axis Monocular Distance Estimation Model (IAMDEM) is proposed to accurately recover target distance. The proposed geometric model incorporates adaptive reference point correction and slope-aware horizontal compensation to mitigate projection errors caused by camera tilt and target misalignment. Experimental results in multiple indoor and outdoor environments demonstrate that the proposed framework achieves reliable detection under extreme pixel-sparse conditions and significantly improves distance estimation accuracy. The method achieves average distance estimation errors of 0.1528 m and 0.1052 m in two testing environments, while maintaining lightweight computational complexity suitable for real-time deployment in consumer vision systems.

In dynamic Internet of Things (IoT) and mobile ad hoc network (MANET) environments, routing based on a single wireless protocol often fails to simultaneously satisfy reliability, latency, energy-efficiency, and routing-overhead requirements. To address this limitation, this paper proposes an Adaptive Heterogeneous Routing (AHR) protocol for ZigBee–BLE multi-radio MANETs. Built upon the Ad hoc On-demand Distance Vector (AODV) routing framework, AHR formulates route selection as a reliability–latency-aware heterogeneous link arbitration problem, where ZigBee and BLE interfaces are dynamically selected on a hop-by-hop basis according to physical-layer link conditions. A probabilistic Link Availability Probability (LAP) model is developed to quantify the probability that a wireless link satisfies the required SINR threshold under log-normal shadowing. Based on this model, two routing strategies are designed. The Hybrid-Power strategy prioritizes statistically stable links to improve packet delivery reliability, whereas the Hybrid-Speed strategy minimizes an expected-delay cost that jointly considers transmission latency and link availability. By penalizing unstable high-rate links, the expected-delay formulation prevents the routing algorithm from selecting marginal BLE links solely due to their nominal data rate. Extensive Monte Carlo simulations are conducted under three deployment areas, incorporating simplified MAC/channel-access overhead, energy consumption, goodput, 95% confidence intervals, and statistical significance tests. The results show that the proposed AHR framework provides a flexible trade-off among packet delivery ratio, end-to-end delay, energy consumption, and routing overhead. In particular, Hybrid-Power achieves the most robust packet delivery performance in sparse and large-scale deployments, while Hybrid-Speed reduces the delay of successfully delivered packets when reliable high-rate BLE links are available. These results demonstrate that AHR can serve as an adaptive cross-layer routing mechanism for heterogeneous IoT networks with diverse QoS requirements.

Trilateration is widely adopted in Received-signal-strength-Indicator (RSSI)-based localization owing to its simplicity. However, its performance degrades for noisy or biased distance estimates, especially with high geometric dilution of precision (GDOP) or multipath effects. To improve robustness, we propose an Annulus-Constrained Least Squares Trilateration (AC-LST) scheme, which models distance uncertainty as σ-level annular regions around anchor nodes derived from two-way RSS measurements. This converts the localization task into a constrained nonlinear least squares problem with the solution restricted to the overlapping annular zones. The optimization problem was solved using Sequential Quadratic Programming (SQP), ensuring fast convergence under geometric and measurement constraints. An adaptive σ-selection mechanism was also introduced, dynamically adjusting the constraint width based on link reliability to improve resilience against variable noise conditions. Simulation and experimental results showed that the AC-LST scheme outperformed conventional least squares trilateration methods, especially in high-noise and high-GDOP environments. The proposed scheme integrates statistical confidence modeling and geometric constraint. Across the three deployment layouts, AC-LST reduces the median localization error by 23–97% compared with conventional LST, and effectively eliminates catastrophic failures under high-GDOP conditions. Compared with the state-of-the-art CLST method, AC-LST achieves comparable median accuracy while exhibiting improved stability and stronger robustness to measurement noise and geometry distortion. Overall, AC-LST offers the best accuracy–robustness trade-off among the evaluated RSSI-based trilateration methods.

The field of human activity recognition (HAR) using channel state information (CSI) is advancing rapidly through recent research. In this study, we propose and investigate a novel lightweight and fast-response gated recurrent unit (LFR-GRU) model that can distinguish continuous dynamic and static postures over time of a person without electronic devices in an indoor environment by utilizing the propagation characteristics of Wi-Fi signals. The CSI time-series amplitude data used in this study are collected from the characteristics of continuous movements and postures in two directions (horizontal and vertical) using an 802.11n Intel network interface card and a Wi-Fi access point. The proposed LFR-GRU model is evaluated using criteria such as F1-score, total number of parameters, learning time, processing speed, and memory usage, and its performance is compared with conventionally available models like 1D convolutional neural networks (CNNs), long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), and Bi-GRU. The experiment results show that the proposed LFR-GRU model achieved an F1-score accuracy of 94.16% when trained with one person’s data and 96% accuracy when trained with two people. Additionally, it demonstrates comprehensively improved performance compared to other time-series models in terms of weight reduction, computational cost, and fast response (the model’s ability to respond quickly to input) speed. Based on these results, it can be confirmed that the proposed LFR-GRU model has sufficient potential to real-time HAR applications in indoor environments.

Emergency scenario recognition has diverse applications, including monitoring older adults living alone and detecting crimes. Predominantly, emergency scenario recognition methods rely on cameras, presenting privacy issues, especially in bathrooms and homes. Additionally, most studies concentrate on recognizing simple isolated activity such as falling down and lying down. However, real-life emergency scenarios often more than just isolated activity. This paper introduces a method for classifying emergency scenarios, including falling down, lying down, as well as the combined activities such as falling down and standing up, lying down and standing up, falling down and crawling, and lying down and crawling. Falling and lying down are similar scenarios, and we consider that the people may stand up or crawl afterward. Our approach hinges on the micro-Doppler features derived from continuous-wave radar signals. We first process the human-generated radar echo signals to produce micro-Doppler spectrogram images. Subsequently, we propose an AI model utilizing multi-head attention mechanisms for image training and testing, demonstrating strong performance on our small datasets while reducing computational complexity compared to precious vision transformer models. Preliminary results highlight the efficacy of our approach, with an average recognition accuracy exceeding 95%.

Sensor-related indoor localization has attracted considerable attention in recent years. The accuracy of conventional fingerprint solutions based on a single sensor, such as a Wi-Fi sensor, is affected by multipath interferences from other electronic devices that are produced as a result of complex indoor environments. Light sensors and magnetic (i.e., geomagnetic) field sensors can be used to enhance the accuracy of a system since they are less vulnerable to disturbances. In this paper, we propose a deep feedforward (DFF)-neural-network-based method, termed DFF-WGL, which integrates the data from the embedded Wi-Fi sensor, geomagnetic field sensor, and light sensor (WGL) in a smart device to localize the device in an indoor environment. DFF-WGL does not require complex and expensive auxiliary equipment, except for basic fluorescent lamps and low-density Wi-Fi signal coverage, conditions that are easily satisfied in modern offices or educational buildings. The proposed system was implemented on a commercial off-the-shelf android device, and performance was evaluated through an experimental analysis conducted in two different indoor testbeds, one measuring 60.5 m² and the other measuring 38 m², with 242 and 60 reference points, respectively. The results indicate that the model prediction with an input consisting of the combination of light, a magnetic field sensor, and two Wi-Fi RSS signals achieved mean localization errors of 0.01 m and 0.04 m in the two testbeds, respectively, compared with any subset of combination of sensors, verifying the effectiveness of the proposed DFF-WGL method.

The technology for human activity recognition has diverse applications within the Internet of Things spectrum, including medical sensing, security measures, smart home systems, and more. Predominantly, human activity recognition methods have relied on contact sensors, and some research uses inertial sensors embedded in smartphones or other devices, which present several limitations. Additionally, most research has concentrated on recognizing discrete activities, even though activities in real-life scenarios tend to be continuous. In this paper, we introduce a method to classify continuous human activities, such as walking, running, squatting, standing, and jumping. Our approach hinges on the micro-Doppler (MD) features derived from continuous-wave radar signals. We first process the radar echo signals generated from human activities to produce MD spectrograms. Subsequently, a bidirectional gate recurrent unit (Bi-GRU) network is employed to train and test these extracted features. Preliminary results highlight the efficacy of our approach, with an average recognition accuracy exceeding 90%.

Device-free localization (DFL) is a technology designed to determine the positions of targets without the need for them to carry electronic devices. It achieves this by analyzing the shadowing effects of radio links within wireless sensor networks (WSNs). However, obtaining high precision in DFL often results in increased energy consumption, severe electromagnetic interference, and other challenges that impact positioning accuracy. Most DFL schemes for accurate tracking require substantial memory and computing resources, which make them unsuitable for resource-constrained applications. To address these challenges, we propose an intelligent mesh cluster (IMC) algorithm that achieves accurate tracking by adaptively activating a subset of wireless links. This approach not only reduces electromagnetic interference but also saves energy. The IMC algorithm leverages geometric objects, such as meshes and mesh clusters formed by wireless links, to achieve low computational complexity. By scanning a subset of mesh cluster-related wireless links near the DFL target, the algorithm significantly reduces the computational requirements. The target’s location estimate is determined based on the connection information among the mesh clusters. We conducted numerous simulations to evaluate the performance of the IMC algorithm. The results demonstrate that the IMC algorithm outperforms grid-based and particle filter-based DFL methods, confirming its effectiveness in achieving accurate and efficient localization.

The position information and posture information of a device-free human in a confined indoor environment has multiple uses in health monitoring and other areas. In this study, we aim to use a single-input-single-output frequency-modulated continuous-wave (FMCW) radar at 24 GHz band for simultaneous localization and posture estimation of one device-free human target. This device is easy to deploy in new setups. We use image formation to convert temporal measurements of the radar signal into image-like data, and convert the problem of simultaneous position estimation and posture perception into an image classification problem. Leveraging the use of convolutional neural networks (CNN) for image classification, we design a variety of tests to explore the most favorable parameters of the CNN model, and then use the best practice model to accomplish the classification task involving a fusion of localization and pose recognition. To explain the primary causes of inaccuracies, we examine not only cases based on a fused position and posture dataset but also cases based on position-only and posture-only datasets. On both real-life datasets, our proposed scheme can achieve 98% classification accuracy and less than 1-m localization accuracy within 0.95 cumulative error probability within the area of interest, outperforming traditional classification methods.