Information technology has been actively utilized in the field of imaging diagnosis using artificial intelligence (AI), which provides benefits to human health. Readings of abdominal hemorrhage lesions using AI can be utilized in situations where lesions cannot be read due to emergencies or the absence of specialists; however, there is a lack of related research due to the difficulty in collecting and acquiring images. In this study, we processed the abdominal computed tomography (CT) database provided by multiple hospitals for utilization in deep learning and detected abdominal hemorrhage lesions in real time using an AI model designed in a cascade structure using deep learning, a subfield of AI. The AI model was used a detection model to detect lesions distributed in various sizes with high accuracy, and a classification model that could screen out images without lesions was placed before the detection model to solve the problem of increasing false positives owing to the input of images without lesions in actual clinical cases. The developed method achieved 93.22% sensitivity and 99.60% specificity.

Ultrawideband (UWB) radar has emerged as a compelling solution for noncontact human activity recognition. This study proposes a novel framework that leverages adaptive signal decomposition and video-based deep learning to classify human motions with high accuracy using a single UWB radar. The raw radar signals were processed by empirical wavelet transform (EWT) to isolate the dominant frequency components in a data-driven manner. These components were further analyzed using the Hilbert transform to produce time–frequency spectra that capture motion-specific signatures through subtle phase variations. Instead of treating each spectrum as an isolated image, the resulting sequence was organized into a temporally coherent video, capturing spatial and temporal motion dynamics. The video data were used to train the SlowFast network—a dual-path deep learning model optimized for video-based action recognition. The proposed system achieved an average classification accuracy exceeding 99% across five representative human actions. The experimental results confirmed that the EWT–Hilbert-based preprocessing enhanced feature distinctiveness, while the SlowFast architecture enabled efficient and accurate learning of motion patterns. The proposed framework is intuitive, computationally efficient, and scalable, demonstrating strong potential for deployment in real-world scenarios such as smart healthcare, ambient-assisted living, and privacy-sensitive surveillance environments.

Human Activity Recognition (HAR) is a key technology with applications in healthcare, security, smart environments, and sports analytics. Despite advances in deep learning, challenges remain in building models that are both efficient and generalizable due to the large scale and variability of video data. To address these issues, we propose a novel Dual-Teacher Knowledge Distillation (DTKD) framework tailored for HAR. The framework introduces three main contributions. First, we define static and dynamic activity classes in an objective and reproducible manner using optical-flow-based indicators, establishing a quantitative classification scheme based on motion characteristics. Second, we develop subset-specialized teacher models and design a hybrid scoring mechanism that combines teacher confidence with cross-entropy loss. This enables dynamic weighting of teacher contributions, allowing the student to adaptively balance knowledge transfer across heterogeneous activities. Third, we provide a comprehensive evaluation on the UCF101 and HMDB51 benchmarks. Experimental results show that DTKD consistently outperforms baseline models and achieves balanced improvements across both static and dynamic subsets. These findings validate the effectiveness of combining subset-aware teacher specialization with hybrid scoring. The proposed approach improves recognition accuracy and robustness, offering practical value for real-world HAR applications such as driver monitoring, healthcare, and surveillance.

Vehicle in-cabin occupancy detection supports restraint-system control, child presence detection, and intelligent cabin sensing, with child presence detection linked to safety assessment frameworks such as the European New Car Assessment Program. High-dimensional radar representations show strong potential, but their physical basis can be difficult to interpret. Engineered-feature approaches improve interpretability, yet few explicitly separate the linear-delay component from the residual structure of the complex reflection response. This study proposes an interpretable representation derived from the complex reflection ratio within a fixed time-domain region of interest (ROI). The method extracts direct-delay cues, frequency-domain Hilbert residual signature (FHRS) scalar features, and continuous wavelet transform (CWT)-based residual-delay and log-magnitude-derivative shape features. In the original experiment, the residual-centered condition achieved 97.72% accuracy and a 97.73% macro F1-score, whereas the direct-delay-only condition achieved 100.00%. A separately acquired 20-trial challenge experiment used Empty Seat, Person, and Human-like Object-loaded Seat conditions, with a human-shaped test object used for the latter. In this challenge, the direct-delay-only, residual-centered, and all-feature conditions achieved accuracies of 79.44%, 96.92%, and 97.39%, respectively. These results demonstrate that residual features provide nonredundant discriminative information when direct-delay cues alone are insufficient. Participant-independent, cross-vehicle, and cross-installation generalization remain to be verified.

Human action recognition (HAR) often struggles to capture important temporal cues distributed across an entire video when relying solely on a single sampled clip. To overcome this limitation, this study proposes a framework that constructs two temporal views from the same video and explicitly learns the prediction consistency between them. Specifically, the prediction-level agreement (AG) loss was introduced to align the class probability distributions of the two views. In addition, conditional gating was applied to adaptively control the contribution of AG loss according to the sample-wise prediction confidence, thereby reducing unstable alignment in temporally ambiguous or information-insufficient segments. The proposed framework was evaluated using both convolutional neural network (CNN)- and Transformer-based backbones on three representative action-recognition benchmark datasets, and it generally improved the performance over the single-view baseline across backbone–dataset combinations. Further empirical analyses, including training behavior, motion magnitude, temporal prediction stability, and qualitative case studies, were conducted to examine the effectiveness and behavior of the proposed two-view framework from multiple perspectives.