Details of a Researcher

Papers - THI THI ZIN

Division display 1 - 20 of about 329 ／ All the affair displays >>

Non-contact Monitoring of Dystocia in Dairy Cows Using Keypoint Detection and Semantic Segmentation Reviewed International journal

T. Murayama, Thi Thi Zin, I. Kobayashi, M. Aikawa

The 2026 IEEE International Conference on Consumer Technology – Pacific (ICCT-Pacific 2026) 2026.3

Authorship：Corresponding author Language：English Publishing type：Research paper (international conference proceedings) Publisher：IEEE

In the dairy industry, labor shortages and the economic losses caused by calving accidents are significant issues. To address these problems, we propose a non-contact monitoring system using 360-degree cameras and deep learning techniques. This study focuses on constructing an automated workflow that detects cows, estimates their poses (standing or lying), and tracks individuals without attaching sensors to the animals. We employed YOLO11 for cow detection and keypoint extraction, and compared three models for pose estimation: Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), and Semantic Segmentation (Deeplabv3+). The experimental results showed that YOLO11 achieved a high detection accuracy (mAP@0.50: 99.47%) for bounding boxes. For pose estimation, the semantic segmentation approach with a ResNet101 backbone achieved the highest accuracy of 85.1%, outperforming keypoint-based methods. These results demonstrate the potential of the proposed system for basic behavioral monitoring in calving barns.

A Study on Supporting Neurocognitive Disorder Assessment for Deaf Individuals Using a Sign Language Recognition System Reviewed International journal

N. Shibahara, Thi Thi Zin, S. Ito, N. Takahashi, N. Takemoto

The 2026 IEEE International Conference on Consumer Technology – Pacific (ICCT-Pacific 2026) 2026.3

Authorship：Corresponding author Language：English Publishing type：Research paper (international conference proceedings) Publisher：IEEE

The Mini Mental State Examination (MMSE) is widely used for screening Neurocognitive Disorder (NCD); however, ensuring diagnostic accuracy for Deaf individuals remains a challenge due to factors such as the potential subjectivity and translation errors introduced by sign language interpreters. To address this issue, this study proposes an automated MMSE scoring system employing Japanese Sign Language (JSL) recognition based on skeletal keypoints. The proposed method utilizes MediaPipe Pose and Hands to extract feature points from examination videos and employs a Long Short-Term Memory (LSTM) model to classify sign language responses. Evaluation results using 5-fold cross-validation on a dataset of Deaf individuals demonstrated a high average classification accuracy of 92.75%. Furthermore, the system successfully performed automated scoring compliant with the MMSE protocol. These results indicate that the proposed system can enable objective cognitive assessment without interpreter intervention, thereby contributing to more accurate diagnoses for Deaf individuals.

Signal-based feature analysis of behavioral trajectories for predicting calving time and classifying assistance needs Reviewed International journal

Wai Hnin Eaindrar Mg, Pyke Tin, M. Aikawa, K. Honkawa, Y. Horii, Thi Thi Zin

Computers and Electronics in Agriculture 243 2026.3

Authorship：Last author,　Corresponding author Language：English Publishing type：Research paper (scientific journal) Publisher：Elsevier B.V.

Accurately predicting calving time and recognizing when a cow needs help during delivery are essential for effective livestock management. These factors directly influence animal welfare, how labor is distributed on the farm, and overall productivity. Without close monitoring, calving complications can lead to serious health issues or even death for the cattle. Moreover, delayed assistance during difficult births (dystocia) can significantly harm both the cow and the calf. These problems remain challenging due to the subtle and highly variable nature of cattle behavior, especially within large-scale farming environments where continuous manual monitoring is impractical. This research proposes a fully vision-based, non-invasive system that relies solely on cattle trajectory data derived from images to address these challenges. To analyze signal-based behavioral trajectories associated with calving, we applied three signal-based image processing techniques aimed at predicting calving time and identifying individuals likely to require human assistance during parturition. Our system allows for continuous, automated monitoring using four surveillance cameras eliminating the need for wearable sensors or invasive equipment. We employed three analytical approaches such as amplitude analysis, frequency analysis, and power spectral density analysis (PSD) to interpret cattle movement patterns from camera-derived trajectory data. For predicting calving time, our system achieved 100 % accuracy across all methods. Specifically, the amplitude analysis predicted calving within 9 h, the frequency analysis provided predictions within 5 h, and the PSD analysis predicted calving within 6 h. Moreover, in classifying cattle requiring human assistance during parturition, our system achieved accuracy of 60 %, 60 %, and 65 % for the amplitude, frequency, and PSD analyses, respectively. Unlike conventional methods that rely on wearable sensors, manual observation, or AI models requiring extensive training, our prediction system operates without any model training phase, instead directly analyzing motion patterns from trajectory data to generate predictions. This makes our prediction simpler, more interpretable, and highly scalable, offering a practical and robust solution for improving livestock monitoring and timely intervention in modern farming environments. This work paves the way for further development of automated, non-invasive livestock monitoring technologies.

<THI THI ZIN>

Papers - THI THI ZIN