Laboratory for Image and Media Understanding,Department of Advanced Information Technology, Kyushu University
Research
Video Surveillance（Past）
Object Detection under Varying Illumination
The goal is to detect moving objects under background changes. We are researching about background modeling which is effective for illumination changes and motion changes such as movements of trees and water surface. To realize low cost and high performance background model, we are also researching about a new framework for modeling scene changes and for selecting the most effective model for each scene.
Journals (Peer-reviewed)
- Satoshi Yoshinaga, Atsushi Shimada, Hajime Nagahara, Rin-ichiro Taniguchi
Statistical Local Difference Pattern for Background Modeling
IPSJ Transactions on Computer Vision and Applications, Vol.3, pp.198-210, 2011.12
BibTeX - Atsushi Shimada, Satoshi Yoshinaga, Rin-ichiro Taniguchi
Maintenance of Blind Background Model for Robust Object Detection
IPSJ Transactions on Computer Vision and Applications, Vol.3, pp.148-159, 2011.12
BibTeX
Analysis of adapter in attention of change detection Vision Transformer
Background subtraction methods using ViT have been studied in anomaly detection systems using surveillance cameras, but their accuracy in unlearned scenes is low and additional learning requires high computational complexity. For practical change detection, adaptive methods with low computational complexity are required.
In this study, we introduce MLP-based adapters to reduce computational complexity and show that knowledge about background variation can be learnt by introducing adapters to attention and residual connections.
In this study, we introduce MLP-based adapters to reduce computational complexity and show that knowledge about background variation can be learnt by introducing adapters to attention and residual connections.
International Conferences (Peer-reviewed)
- Ryunosuke HAMADA, Tsubasa MINEMATSU, Cheng TANG, Atsushi SHIMADA
Analysis of adapter in attention of change detection Vision Transformer
AI-based All-Weather Surveillance System, AWSS 2024, 2024.12
BibTeX
Wide-area Object Tracking
We are researching about object tracking in the wide area where many sensors are arranged distributedly, such as a shopping mall and a town.
The connection among the sensors is automatically estimated by using the tracking results in each sensor. As with the number of sensors, many computers are required for the wide area tracking.
Therefore, with the view of load balancing and accuracy enhancement, we are also researching about estimating automatically an optimal sensor assignment for each computer.
The connection among the sensors is automatically estimated by using the tracking results in each sensor. As with the number of sensors, many computers are required for the wide area tracking.
Therefore, with the view of load balancing and accuracy enhancement, we are also researching about estimating automatically an optimal sensor assignment for each computer.
International Conferences (Peer-reviewed)
- Shuhei Noda, Atsushi Shimada, Daisaku Arita, Rin-ichiro Taniguchi
Object Tracking across Non-overlapping Views of Multiple Sensors
Proc. of International Workshop on "Sensing Web" in conjunction with the 19th International Conference on Pattern Recognition (ICPR2008), 2008.12
BibTeX
A Human-in-the-Loop Annotation Framework with Open-set Recognition Under Surveillance Scenarios
The use of surveillance video has increased in recent years, but the high cost of labeling is a challenge. The proposed framework aims to minimize the workload while maintaining high accuracy by improving existing methods.
In this study, we measured human workload and retraining accuracy of the classification network YOLOv3. It was shown that the proposed method is effective in both workload and retraining accuracy. Looking ahead, we expect to develop a more robust background change detection model and improve a more stable clustering method to replace DBScan.
In this study, we measured human workload and retraining accuracy of the classification network YOLOv3. It was shown that the proposed method is effective in both workload and retraining accuracy. Looking ahead, we expect to develop a more robust background change detection model and improve a more stable clustering method to replace DBScan.
Sensing Web
Various kinds of sensors have been arranged on the street, in the buildings, the roads and so on. We are researching about “the disclosure of sensing information” to utilize the sensing information for a number of applications.
Each sensor is used independently for each purpose. If such sensors are linked each other through the network, we can create new values for people by combining the information of the sensors.
Each sensor is used independently for each purpose. If such sensors are linked each other through the network, we can create new values for people by combining the information of the sensors.
International Conferences (Peer-reviewed)
- Rin-ichiro Taniguchi, Atsushi Shimada, Yuji Kawaguchi, Yousuke Miyata, Satoshi Yoshinaga
Structuring and Presenting the Distributed Sensory Information in the Sensing Web
Information Processing and Management of Uncertainty in Knowledge-Based Systems. Applications, Vol.8, pp.643-652, 2010.06
BibTeX
Analysis of change detection methods robust to background variation using the Attention mechanism
The demand for security camera surveillance in densely populated areas such as urban centers and large facilities is increasing, and highly accurate anomaly detection is required.
In this study, we analyze the feature representation of each structure and the role of the attention mechanism in the background subtraction algorithm TransCD using the Attention mechanism. we show that the attention mechanism corrects background features and that the residual connection corrects the features of the background grid, but the foreground grid is retained. This allows foreground detection without detecting changes in comparisons between different backgrounds.
In this study, we analyze the feature representation of each structure and the role of the attention mechanism in the background subtraction algorithm TransCD using the Attention mechanism. we show that the attention mechanism corrects background features and that the residual connection corrects the features of the background grid, but the foreground grid is retained. This allows foreground detection without detecting changes in comparisons between different backgrounds.
