| 研究課題/領域番号 |
19K20335
|
|---|
| 研究機関 | 国立研究開発法人産業技術総合研究所 |
|---|
研究代表者 |
Gatto Bernardo 国立研究開発法人産業技術総合研究所, 情報・人間工学領域, 産総研特別研究員 (10826267)
|
|---|
| 研究期間 (年度) |
2019-04-01 – 2023-03-31
|
| キーワード | elderly surveillance / subspace representation / image recognition / deep learning |
|---|
| 研究実績の概要 |
Inspired by applications of subspace analysis, two new collections of methods were presented in this project: (1) Shallow networks for image classification; and (2) Subspaces for tensor representation and classification. New representations are proposed with the aim of preserving the spatial structure and maintaining a fast processing time. We also introduce a technique to keep temporal structure, even using the principal component analysis, which classically does not model sequences. These well-grounded learning algorithms were evaluated over problems involving person detection, action and gesture representation, and classification. More precisely, we focused on the fusion of visual and acoustic data to support the safe life of the elderly living alone. We employed information from visual and acoustic sensors (e.g., cameras and microphones) and recognized events, such as domestic activities or abnormal events.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
The research has progressed more rapidly than initially planned. For instance, we developed a subspace method based on singular spectrum analysis, where we could represent acoustic data, which opened the range of application possibilities of the technique. We also developed a comprehensive formulation using the proposed neural networks and tensor decomposition methods for falling action recognition.
|
| 今後の研究の推進方策 |
We aim to improve the proposed neural networks and tensor analysis methods for future work. We understand that the recent research direction guides a data fusion scheme, where tensor data from videos and audio can be regarded as multimodal distributions. A tensor fusion analysis is required. This approach is unexplored and may reveal new limits of subspace learning. Investigation of the proposed neural networks' theoretical limits and tensor analysis methods is also unexplored. In the particular case of the elderly surveillance system, interpretability is a requirement that cannot be bypassed. Here we list some interesting research directions: (1) Investigate new subspace representations to express tables, trees, and graphs, to name a few. These variants can be readily applied to the subspaces framework and employ their benefits. (2) Develop new shallow networks using Lie groups' theory, which presents a simple model for continuous symmetry, such as rotational symmetry found in three dimensions. (3) Introduce a deterministic neural network initialization by applying the convolutional kernels produced by FKT or GDS as an alternative to the random initialization process.
|
| 次年度使用額が生じた理由 |
Not Applicable.
|
