TY  - JOUR
AU  - Wang, Fafa
AU  - Yi, Shenglun
PY  - 2024
DA  - 2024/10/25
TI  - Spatio-temporal Feature Soft Correlation Concatenation Aggregation Structure for Video Action Recognition Networks
JO  - ICCK Transactions on Sensing, Communication, and Control
T2  - ICCK Transactions on Sensing, Communication, and Control
JF  - ICCK Transactions on Sensing, Communication, and Control
VL  - 1
IS  - 1
SP  - 60
EP  - 71
DO  - 10.62762/TSCC.2024.212751
UR  - https://www.icck.org/article/abs/TSCC.2024.212751
KW  - video action recognition
KW  - soft correlation
KW  - spatio-temporal feature extraction
KW  - concatenation aggregation structure
KW  - bidirectional LSTM
AB  - The efficient extraction and fusion of video features to accurately identify complex and similar actions has consistently remained a significant research endeavor in the field of video action recognition. While adept at feature extraction, prevailing methodologies for video action recognition frequently exhibit suboptimal performance in the context of complex scenes and similar actions. This shortcoming arises primarily from their reliance on uni-dimensional feature extraction, thereby overlooking the interrelations among features and the significance of multi-dimensional fusion. To address this issue, this paper introduces an innovative framework predicated upon a soft correlation strategy aimed at augmenting the representational capacity of features by implementing multi-level, multi-dimensional feature aggregation and concatenating the temporal features produced by the network. Our end-to-end multi-feature encoding soft correlation concatenation aggregation layer, situated at the temporal feature output terminal of the Video Action Recognition network, proficiently aggregates and integrates the output temporal features. This approach culminates in producing a composite feature that cohesively unifies multi-dimensional information, markedly enhancing the network's competency in differentiating analogous video actions. Empirical findings demonstrate that the approach delineated in this paper bolsters the efficacy of video action recognition networks, achieving a more comprehensive characterization of spatiotemporal features, and yielding superior accuracy and robustness.
SN  - 3068-9287
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Wang2024Spatiotemp,
  author = {Fafa Wang and Shenglun Yi},
  title = {Spatio-temporal Feature Soft Correlation Concatenation Aggregation Structure for Video Action Recognition Networks},
  journal = {ICCK Transactions on Sensing, Communication, and Control},
  year = {2024},
  volume = {1},
  number = {1},
  pages = {60-71},
  doi = {10.62762/TSCC.2024.212751},
  url = {https://www.icck.org/article/abs/TSCC.2024.212751},
  abstract = {The efficient extraction and fusion of video features to accurately identify complex and similar actions has consistently remained a significant research endeavor in the field of video action recognition. While adept at feature extraction, prevailing methodologies for video action recognition frequently exhibit suboptimal performance in the context of complex scenes and similar actions. This shortcoming arises primarily from their reliance on uni-dimensional feature extraction, thereby overlooking the interrelations among features and the significance of multi-dimensional fusion. To address this issue, this paper introduces an innovative framework predicated upon a soft correlation strategy aimed at augmenting the representational capacity of features by implementing multi-level, multi-dimensional feature aggregation and concatenating the temporal features produced by the network. Our end-to-end multi-feature encoding soft correlation concatenation aggregation layer, situated at the temporal feature output terminal of the Video Action Recognition network, proficiently aggregates and integrates the output temporal features. This approach culminates in producing a composite feature that cohesively unifies multi-dimensional information, markedly enhancing the network's competency in differentiating analogous video actions. Empirical findings demonstrate that the approach delineated in this paper bolsters the efficacy of video action recognition networks, achieving a more comprehensive characterization of spatiotemporal features, and yielding superior accuracy and robustness.},
  keywords = {video action recognition, soft correlation, spatio-temporal feature extraction, concatenation aggregation structure, bidirectional LSTM},
  issn = {3068-9287},
  publisher = {Institute of Central Computation and Knowledge}
}