A5084799893- Human Pose and Action Recognition
- Context-Aware Activity Recognition Systems
- Domain Adaptation and Few-Shot Learning
- Advanced Technologies in Various Fields
- Anomaly Detection Techniques and Applications
- Multimodal Machine Learning Applications
- Technology and Human Factors in Education and Health
- Indoor and Outdoor Localization Technologies
University of Electronic Science and Technology of China
2023-2025
With the rapid development of wearable cameras, it is now feasible to considerably increase collection egocentric video for first-person visual perception. However, hindered by a shortage multi-modal activity datasets. Furthermore, catastrophic forgetting problem multimodal continual learning, as branch has not been thoroughly explored, which makes accumulating larger data more urgent. To address this shortage, we propose dataset learning named UESTC-MMEA-CL in paper. The collected using our...
The rapid advancement of wearable sensors has significantly facilitated data collection in our daily lives. Human activity recognition (HAR), a prominent research area technology, made substantial progress recent years. However, the existing efforts often overlook issue functional scalability models, making it challenging for deep models to adapt application scenarios that require continuous evolution. Furthermore, when employing conventional continual learning techniques, we have observed...
Continual learning aims to equip deep neural networks (DNNs) with the capability continuously learn new knowledge without catastrophic forgetting. Currently, there is significant attention on multimodal continual activity recognition from a egocentric perspective. However, issue of modality imbalance can lead exacerbated forgetting in learning. To address this, we propose an exemplar-free vision-sensor Attention-based Incremental Discriminability enhancement (AID) method. Firstly, employ...
The rapid development of wearable sensors promotes convenient data collection in human daily life. Human Activity Recognition (HAR), as a prominent research direction for applications, has made remarkable progress recent years. However, existing efforts mostly focus on improving recognition accuracy, paying limited attention to the model’s functional scalability, specifically its ability continual learning. This limitation greatly restricts application open-world scenarios. Moreover, due...
Continual Learning (CL) aims to enable Deep Neural Networks (DNNs) learn new data without forgetting previously learned knowledge. The key achieving this goal is avoid confusion at the feature level, i.e., avoiding within old tasks and between tasks. Previous prototype-based CL methods generate pseudo features for knowledge replay by adding Gaussian noise centroids of classes. However, distribution in space exhibits anisotropy during incremental process, which prevents from faithfully...
The application of activity recognition in the "AI + Education" field is gaining increasing attention. However, current work mainly focuses on activities manually captured videos and a limited number types, with little attention given to recognizing surveillance images from real classrooms. In classroom settings, normal teaching such as reading, account for large proportion samples, while rare non-teaching eating, continue appear. This requires model that can learn few samples without...
The application of activity recognition in the ``AI + Education" field is gaining increasing attention. However, current work mainly focuses on activities manually captured videos and a limited number types, with little attention given to recognizing surveillance images from real classrooms. Activity classroom faces multiple challenges, such as class imbalance high similarity. To address this gap, we constructed novel multimodal dataset focused image called ARIC (Activity Recognition In...
The rapid development of wearable sensors promotes convenient data collection in human daily life. Human Activity Recognition (HAR), as a prominent research direction for applications, has made remarkable progress recent years. However, existing efforts mostly focus on improving recognition accuracy, paying limited attention to the model’s functional scalability, specifically its ability continual learning. This limitation greatly restricts application open-world scenarios. Moreover, due...
With the rapid development of wearable cameras, a massive collection egocentric video for first-person visual perception becomes available. Using videos to predict activity faces many challenges, including limited field view, occlusions, and unstable motions. Observing that sensor data from devices facilitates human recognition, multi-modal recognition is attracting increasing attention. However, deficiency related dataset hinders deep learning recognition. Nowadays, in real world has led...
<p>The rapid development of wearable sensors promotes convenient data collection in human daily life. Human Activity Recognition (HAR), as a prominent research direction for applications, has made remarkable progress recent years. However, existing efforts mostly focus on improving recognition accuracy, paying limited attention to the model's functional scalability, specifically its ability continual learning. This limitation greatly restricts application open-world scenarios....