Earticle

Notwithstanding the growing role of speech emotion recognition (SER) in human-computer interaction, deep-learning models often display a gender imbalance, recognizing female speech less accurately. We explore a gender-aware conformer-transformer framework to solve this, which first learns gender-related patterns under noisy conditions instead of undergoing the standard end-to-end training. These patterns are then fed into a transformer emotion decoder via cross attention. A conformer encoder was pretrained on gender classification under heavy noise with a robustness of 97.73%. This framework achieved a weighted accuracy of 88.64% and narrowed the gender gap (91.1% recognition accuracy for female speech versus 85.2% for males) on the Ryerson audio-visual database of emotional speech and song dataset. The emotion recognition dropped, as expected, on the crowd-sourced emotional multimodal actors dataset without adaptation, owing to domain differences. However, the gender classification was reasonably accurate (F1 = 0.386). We conclude that using explicit gender information improves fairness under matched conditions, although cross-dataset generalization remains challenging.

Purpose: The purpose of this study was to investigate the availability of software for rehabilitation with the Kinect sensor by presenting an efficient algorithm based on machine learning when classifying the motion data of the PNF pattern if the subjects were wearing a patient gown. Methods: The motion data of the PNF pattern for upper extremities were collected by Kinect sensor. The data were obtained from 8 normal university students without the limitation of upper extremities. The subjects, wearing a T-shirt, performed the PNF patterns, D1 and D2 flexion, extensions, 30 times; the same protocol was repeated while wearing a patient gown to compare the classification performance of algorithms. For comparison of performance, we chose four algorithms, Naive Bayes Classifier, C4.5, Multilayer Perceptron, and Hidden Markov Model. The motion data for wearing a T-shirt were used for the training set, and 10 fold cross-validation test was performed. The motion data for wearing a gown were used for the test set. Results: The results showed that all of the algorithms performed well with 10 fold cross-validation test. However, when classifying the data with a hospital gown, Hidden Markov model (HMM) was the best algorithm for classifying the motion of PNF. Conclusion: We showed that HMM is the most efficient algorithm that could handle the sequence data related to time. Thus, we suggested that the algorithm which considered the sequence of motion, such as HMM, would be selected when developing software for rehabilitation which required determining the correctness of the motion.

The number of healthcare products available for pets has increased in recent times, which has prompted active research into wearable devices for pets. However, the data collected through such devices are limited by outliers and missing values owing to the anomalous and irregular characteristics of pets. Hence, we propose pet behavior recognition based on a hybrid one-dimensional convolutional neural network (CNN) and long short- term memory (LSTM) model using pet wearable devices. An Arduino-based pet wearable device was first fabricated to collect data for behavior recognition, where gyroscope and accelerometer values were collected using the device. Then, data augmentation was performed after replacing any missing values and outliers via preprocessing. At this time, the behaviors were classified into five types. To prevent bias from specific actions in the data augmentation, the number of datasets was compared and balanced, and CNN-LSTM-based deep learning was performed. The five subdivided behaviors and overall performance were then evaluated, and the overall accuracy of behavior recognition was found to be about 88.76%.

본 논문에서는 독성물질에 노출된 물벼룩의 행동 패턴에 따라 독성도를 판단하는 방법을 제안하였다. 다양한 조건에서 의 물벼룩의 그림자 추적 장치와 인공지능 기법을 이용하여 행동 패턴을 분석하여 범주화하고 범주화된 자료를 이용하 여 행동 패턴의 변화만으로 독성도를 평가하였다. 이를 위하여 중크롬산칼륨(potassium dichromate, K2Cr2O7), 염화칼륨 (potassium chloride, KCl), 염화카드뮴(cadmium chloride, CdCl2), 염화암모늄(ammonium chloride, NH4Cl)의 4가지 물 질을 선정하여 생태독성시험 절차에 따라 실험을 진행하였으며 EC50 및 TU1~5에 해당하는 농도를 산정하였다. 또한 물 벼룩 이동 경로 검출용 광 감지 시스템을 제작하여 각 TU에 해당하는 농도에서의 물벼룩의 행동 패턴 데이터를 수집하 였다. 물벼룩의 움직임을 1초에 한번을 측정하여(x, y) 좌표로 인식하여 분석을 위한 데이터로 활용하였고 이 데이터는 머신 러닝에서 인공 지능을 학습하는 용도와 학습된 인공 지능의 정확성을 검증하는 용도로 사용하였다. TU1~5 각각에 대하여 5번의 생태독성실험을 실시하여 25개의 학습 데이터를 이용하여 50번을 반복적으로 학습하였다. 학습된 인공지 능의 성능을 시험하기 위해서 5개의 검증 데이터를 사용하였고 5개 모두 분류에 성공하여 100%의 정확성을 도출하였다.

In this paper, we propose a method to determine toxicity according to the behavior pattern of Daphnia Magna exposed to toxic substances. The behavior pattern was analyzed by using the shadow tracking device and the artificial intelligence technique in various conditions and categorized. We show the toxicity can be evaluated only by the observation of the behavior pattern using the categorized data. For this purpose, we select four substances; potassium dichromate (K2Cr2O7), potassium chloride (KCl), cadmium chloride (CdCl2) and ammonium chloride (NH4Cl). Experiments were conducted according to the ecotoxicity test procedure and the density corresponding to EC50 and TU1~5 were calculated. In addition, a light sensing system for detecting the movement path of Daphnia Magna was constructed and the behavior pattern data of Daphnia Magna at the density corresponding to each TU was collected. The movement of the Daphnia Magna was measured once a second and used as data for analysis by recognizing them as (x, y) coordinates. This data was used to train artificial intelligence and to verify the accuracy of learned artificial intelligence. We conducted five ecotoxicity tests for each of TU1~5 and repeatedly learned 50 times using 25 training data. In order to test the performance of the learned artificial intelligence, five verification data were used and all five were classified successfully and 100% accuracy was obtained.

최근 극단적인 강우와 홍수의 발생 빈도가 증가함에 따라, 수자원 시설은 보다 강하고 빈번한 환경 스트레스에 노출되고 있다. 이처럼 반복적인 환경 스트레스는 구조물의 균열 형성과 확산을 가속화하며, 내구성 저하와 안전성 문제로 이어진다. 이에 본 연구에서는 건물 균열 이미지 데이터를 기반으로, 다단계 인공지능 영상 인식 가상센서를 적용한 수자원 시설 균열 자동 감지 기법을 제안하고자 한다. 첫째, YOLO11 기반 벽체 영역분할을 통해 관심 영역을 설정하고, 둘째, 동일 계열의 균열 탐지 모델로 균열 후보를 검출한다. 셋째, ResNet18 기반 이진 분류로 실제 균열 여부를 판별하며. 마지막으로 YOLO11 기반 세분화(Segmentation)모델을 적용하여 균열의 형상을 정밀하게 인식한다. 본 연구의 결과는 계층적 인공지능 분석 체계를 통해 수자원 시설 구조물의 균열을 효율적으로 감지하고 손상 위험을 조기에 파악할 수 있는 과학적 근거를 마련하는 데 기여할 것으로 기대된다.

With the increasing frequency of extreme rainfall and flooding events, water infrastructure has been increasingly exposed to stronger and more frequent environments stresses. Such repetitive external loads accelerate the formation and propagation of cracks in structures, leading to reduced durability and potential safety issues. In this study, a multi-stage artificial intelligence vision-based virtual sensor approach is proposed to automatically identify cracks in water infrastructure using building crack image data. First, the region of interest (ROI) is defined through wall segmentation based on YOLO11, followed by crack candidate detection using the same YOLO11 framework. Next, a ResNet18-based binary classification model determines the actual presence of cracks, and finally, a YOLOv11-based segmentation model is applied to precisely identify the crack geometry. The Proposed multi-stage AI analysis framework enables efficient crack identify in water infrastructure and provides a scientific foundation for the early identification of potential structural damage.

레이다 센서는 환경적 변수의 영향을 최소화하고 사생활 침해 우려가 없어 기존의 시스템과 차별화되며, 데이터의 크기가 작아 실시간 비접촉식 제어에 유리하다. 따라서, 본 논문에서는 양손이 자유롭지 않은 상황에서 차량 트렁크를 편리하게 제어하기 위한 CW (continuous wave) 레이다 센서 기반 발동작 인식 시스템을 제시한다. 이를 위해 24 GHz 대역의 CW 레이다 센서를 활용하여 5가지 발동작을 측정하고, 단시간 푸리에 변환을 적용하여 스펙트로그램을 추출하였다. AI 분류기 설계는 연산 복잡도와 메모리 사용량을 최소화하기 위해 BNN (binarized neural network) 기반 구조로 설계하였고, 이를 CNN (convolutional neural network)과 비교 분석하였다. 제안된 모델은 98.20%의 정확도를 달성하였고, Raspberry Pi 4에서 구현되어 실시간 추론을 성공적으로 수행함으로써 저사양 임베디드 환경에서의 효율적인 동작 가능성을 입증하였다.

The radar sensor distinguishes itself from existing systems by minimizing sensitivity to environmental factors, eliminating privacy concerns, and enabling real-time contactless control due to its compact data size. Accordingly, this paper proposes a continuous wave (CW) radar sensor-based foot gesture recognition system designed for convenient vehicle trunk control in situations where both hands are occupied. Five distinct foot gestures were measured using a 24 GHz CW radar sensor, and spectrograms were extracted through short-time Fourier transform. To minimize computational complexity and memory usage, the AI classifier was designed based on a binarized neural network (BNN), and its performance was compared and analyzed against a convolutional neural network (CNN). The proposed model achieved an accuracy of 98.20% and was successfully implemented on a Raspberry Pi 4, demonstrating its capability for efficient real-time inference in low-spec embedded environments.

As life expectancy becomes longer, reduction of human muscular strength threatens quality of human life. Many robotic devices have thus been developed to support and help human daily life. This paper deals with a new type of in-wheel actuator that can be effectively used for the robotic devices. BLDC motor, drive board, brake, ARS (Attribute Reference System), and torque sensor are combined in the single actuator module. The torque sensor is used to recognize human intention and the in-wheel actuator drives walking aids in our system. Its feasibility was tested with the active walking aid device equipped with the in-wheel actuator. Based on it, we designed an admittance filter algorithm to react on uphill and downhill drive. By adjusting mass, damping, and spring parameters in accordance with the ARS output, it provided convenient drive to the old on uphill and downhill walks.

스마트폰을 이용한 인간 활동 인식은 컴퓨터 지능 분야에서 뜨거운 연구 주제이다. 스마트폰에는 다양한 센서가 장착되어 있다. 이러한 센서의 데이터를 융합하면 응용프로그램에서 많은 활동을 인식할 수 있다. 그러나 이러한 장치는 활용 가능한 센서 수가 제한되기 때문에 리소스가 적으며, 최적의 성능과 효율적인 특징 추출을 달성하기 위해서는 특징 선택 및 분류 방법이 필요하다. 이 논문에서는 이러한 요구사항에 따라 스마트폰-기반 HAR 체계를 제안한다. 이 논문 에서 제안된 방법은 가속도 센서, 자이로 센서, 기압 센서에서 시간-도메인 특징을 추출하며, 커널 판별 분석(KDA)과 SVM을 적용하여 높은 정확도로 활동을 인식한다. 이 방법은 각 활동에 대해 각 센서에서 가장 관련성이 높은 특징을 선택한다. 우리의 비교 결과는 제안된 시스템이 이전의 스마트폰-기반 HAR 시스템보다 성능이 우수함을 보여준다.

Human activity recognition(HAR) using smartphones is a hot research topic in computational intelligence. Smartphones are equipped with a variety of sensors. Fusing the data of these sensors could enable applications to recognize a large number of activities. However, these devices have fewer resources because of the limited number of sensors available, and feature selection and classification methods are required to achieve optimal performance and efficient feature extraction. This paper proposes a smartphone-based HAR scheme according to these requirements. The proposed method in this paper extracts time-domain features from acceleration sensors, gyro sensors, and barometer sensors, and recognizes activities with high accuracy by applying KDA and SVM. This approach selects the most relevant feature of each sensor for each activity. Our comparison results shows that the proposed system outperforms previous smartphone-based HAR systems.

인간 행동 인식(Human Activity Recognition) 연구는 주로 한 가지 센서만을 이용하여 진행되고 있다. 본 연구에서는 실험을 통해 10종류의 동작에 대한 3축 자이로 데이터 및 가속도 데이터와 비디오 데이터를 확보하였다. 비디오 데이터에서 신체의 33개 랜드마크의 3축 위치좌표 데이터를 추출하여 동 작 데이터를 시계열 데이터로 통일하고, 1차원 합성곱 신경망을 이용하여 지도 학습을 진행하였다.

최근 인공지능, 클라우드 컴퓨팅 등 차세대컴퓨팅 기술의 발달로 인하여 행동인식 연구가 활발하게 진행되고 있다. 행동인식은 사람의 특정 행동을 인식하는 연구 분야로 주로 가속도계, 자이로스코프 등의 센서 데이터나 카메라의 영상 데이터를 기반으로 한다. 기존 연구에서는 사용자의 행동 패턴인식에서 주로 성인을 대상으로 하는 인식 알고리즘 개발이 주를 이루었다. 본 연구에서는 아동들의 놀이 행동 레벨을 4가지 등급(low level activity, light level activity, locomotion, vigorous sports)으로 나누어 분석하였다. 제안된 모델을 평가하기 62명의 아동으로부터 데이터를 수집하였으며, 신호 처리를 기반으로 신호의 특징을 추출하고 임계치를 적용하여 레벨을 분석하도록 하였다. 추후 실시간으로 분석하는 알고리즘을 적용하여 모바일 폰 기반의 모니터링 앱에 적용할 계획이다.

현대 사회에서 안전 서비스 및 시스템 환경을 구축하고 활용하는 것은 매우 중요하고 관심이 큰 분야이다. 특히, 어린이, 고령자, 여성, 장애인 및 외국인 등 사회적 취약 계층에 대한 안전 서비스 제공은 사회적 이슈가 되고 있다. 그러나, 대부분의 안전 서비스 및 시스템은 일반인을 대상으로 적용되고 있기 때문에 사회적 취약 계층을 위한 시스템 개발이 필요하다. 따라서, 본 논문에서는 실시간으로 상황을 인식하고, 신속한 대응을 할 수 있도록 데이터를 처리, 전송 하는 시스템 모듈을 개발하였다. 이를 위해, 실시간 상황 인식에 필요한 다양한 센서를 통합 모듈로 설계하고, 이를 통해 수집된 데이터를 분석하여 처리 결과를 전송하는 안전 시스템 모듈을 개발하였다.

In modern society, developing and utilizing safety service and system environments is a very important and great interest. In particular, the provision of safety services to socially vulnerable groups such as children, the elderly, women, the disabled and foreigners has become a social issue. However, since most safety services and systems are applied to the general public, it is necessary to develop systems for socially vulnerable groups. Therefore, in this paper, we developed a system module that processes and transmits data to recognize the situation in real time and respond quickly. To this end, various sensors for real-time situation recognition were designed as integrated modules, and a safety system module was developed to analyze the collected data and transmit the processing results.

후두출 환자나 발음장애가 있는 사람은 음성인식 기술을 사용하지 못하는 불편함이 있다. 이러한 문제를 극복하기 해 발성이 필요 없는 무성인식(Silent Speech Recognition)이 발하고 있다. 본 논문에서는 물리센서 (Physical Sensor)를 이용하여 의 움직임을 감지하고 한국어 무성 자음 ‘ㄱ’, ‘ㄷ’, ‘ㅂ’, ‘ㅅ’, ‘ㅈ’, ‘ㅊ’, ‘ㅋ’, ‘ㅌ’, ‘ㅍ’, ‘ㅎ’ 의 조음 치를 분류하는 방법을 제안한다. 세 개의 특징 추출 알고리즘 VAR(Variance), CC(Cepstral Coefficient), LPC(Linear Predict coding Coefficient)와 ANN(Artificial Neural Network) 분류기를 통해 수집된 데이터를 5개의 조음 치로 분류하다. 조음 치를 분류하는 실험을 통해 평균 79.5%의 정확도를 얻었다.

A patient with laryngectomy or a person with pronunciation problem has inconvenience of not using voice recognition technology. To solve it, the silent speech recognition technology has been emerged. In this paper, the proposed method detects movement of tongue using physical sensor and classifies the articulation position of Korean silent consonants such as‘ㄱ’, ‘ㄷ’, ‘ㅂ’, ‘ㅅ’, ‘ㅈ’, ‘ㅊ’, ‘ㅋ’, ‘ㅌ’, ‘ㅍ’, ‘ㅎ’. Three feature extraction algorithms such as VAR(Varian), CC(Cepstral Coefficient), LPC(Linear Predict coding Coefficient) and ANN(Artificial Neural Network) classifier can recognize collected data into five articulation positions. Experimental results show that the classification accuracy of the proposed method is about 79.5%.

현재 RC카는 남녀노소를 불문하고 정밀한 컨트롤과 질주 본능을 느끼고 싶어하는 모든 사람이 참여할 수 있는 하나의 문화로써 자리잡게 되었다. 또한 RC카는 직접 차량을 운전하면서 느끼는 체감을 작고 가벼운 몸체 하나로 경험할 수 있다는 것을 이유로 많은 사람들에게 주목받고 있다. 이러한 트렌드에 맞춰 다양한 개발 회사들은 각각의 환경에 맞는 주행, 속도 개선와 같은 기능적인 부분과 그 정교함에 초점을 두어 개발하고 있다. 물론, 주행 성능만으로 도 굉장한 파급 효과를 불러올 수 있지만, RC카를 가지고 할 수 있는 것은 부품 변경, 모터 개조 정도 밖에 없다. 따 라서, 라즈베리파이 기반의 RC카를 제작해 어플리케이션으로 컨트롤 하고 센서 인식을 통한 다양한 기능의 추가와 웹 캠을 사용해 얻을 수 있는 영상 정보를 통한 주행을 할 수 있는 RC카를 개발한다. 작은 RC카 한 대로 사람이 진입할 수 없는 다양한 건설,사고 현장에서 움직이는 로봇으로써의 사용과 사용자들의 흥미를 이끌 수 있을 만한 기능을 겸비 할수 있도록 연구를 수행한다.

Now RC cars have become a culture where everyone who wants to feel precise control and rushing instinct regardless of sex, young or old can participate. In addition, RC cars have attracted a lot of people because they can feel the feeling of driving while driving a car in a small, lightweight body. In line with these trends, various development companies are focusing on functional aspects such as driving and speed improvement for each environment and their sophistication. Of course, driving performance alone can bring a tremendous ripple effect, but what you can do with a RC car is only part replacement and motor modification. Therefore, we will develop RC car based on raspberry pie, control by application, add various functions through sensor recognition, and develop RC car that can travel through video information that can be obtained using web cam. As a small RC car, we carry out research that can combine the functions that can be used as a robot that moves at various construction sites and accident sites where people can not enter, and which can attract users' interest.

The aim of this paper is to propose a new multi-sensor target recognition method base on relative entropy evaluation theory. There are several influencing factors in the target recognition problem, which needs several sensors to work together. Then the multi-sensor target recognition problem can be regarded as a multi-attribute decision making problem. Relative entropy measure can depict the closeness of the two systems, and then this paper will use it to develop an improved TOPSIS method for the multi-sensor target recognition problem. A new characteristic index weights method is proposed, which can avoid the subjectivity of the weight of characteristic indexes. Finally, an application example is used to illustrate the effectiveness and feasibility of the proposed method.

The purpose of this study is to suggest that the movement expression activity of intellectually disabled people is effective in the learning process of LMA motion recognition based on Kinect sensor. We performed an ICT motion recognition games for intellectually disabled based on movement learning of LMA. The characteristics of the movement through Laban's LMA include the change of time in which movement occurs through the human body that recognizes space and the tension or relaxation of emotion expression. The design and implementation of the motion recognition model will be described, and the possibility of using the proposed motion recognition model is verified through a simple experiment. As a result of the experiment, 24 movement expression activities conducted through 10 learning sessions of 5 participants showed a concordance rate of 53.4% or more of the total average. Learning motion games that appear in response to changes in motion had a good effect on positive learning emotions. As a result of study, learning motion games that appear in response to changes in motion had a good effect on positive learning emotions