A Comparative Study of Deep Learning Approaches for Human Action Recognition

Gülsüm Yiğit

Article Details

A Comparative Study of Deep Learning Approaches for Human Action Recognition

Authors:

Gülsüm Yiğit

Publisher:

Murat YAKAR

Main Topics:

Software Engineering (Other)

Journal:

Turkish Journal of Engineering Vol.9 Issue 2, pp.281-289

Document Type:

Article

Publication Date:

6/2025

Document Language:

English

Document ID (Org No.):

2001339346

pISSN:

2587-1366

Keywords:

Abstract:

 Human Action Recognition HAR plays a crucial role in understanding and categorizing human activities from visual data, with applications ranging from surveillance, healthcare to human-computer interaction. However, accurately recognizing a diverse range of actions remains challenging due to variations in appearance, occlusions, and complex motion patterns. This study investigates the effectiveness of various deep learning model architectures on HAR performance across a dataset encompassing 15 distinct action classes. Our evaluation examines three primary architectural approaches: baseline EfficientNet models, EfficientNet models augmented with Squeeze-and-Excitation SE blocks, and models combining SE blocks with Residual Networks. Our findings demonstrate that incorporating SE blocks consistently enhances classification accuracy across all tested models, underscoring the utility of channel attention mechanisms in refining feature representation for HAR tasks. Notably, the model architecture combining SE blocks with Residual Networks achieved the highest accuracy, increasing performance from 69.68 in baseline EfficientNet to 76.75, marking a significant improvement. Additionally, alternative models, such as EfficientNet integrated with Support Vector Machines EfficientNet-SVM and ZeroShot Learning models, exhibit promising results, highlighting the adaptability and potential of diverse methodological approaches for addressing the complexities of HAR. These findings provide a foundation for future research in optimizing HAR systems, with implications for enhancing robustness and accuracy in action recognition applications.

 Human Action Recognition HAR plays a crucial role in understanding and categorizing human activities from visual data, with applications ranging from surveillance, healthcare to human-computer interaction. However, accurately recognizing a diverse range of actions remains challenging due to variations in appearance, occlusions, and complex motion patterns. This study investigates the effectiveness of various deep learning model architectures on HAR performance across a dataset encompassing 15 distinct action classes. Our evaluation examines three primary architectural approaches: baseline EfficientNet models, EfficientNet models augmented with Squeeze-and-Excitation SE blocks, and models combining SE blocks with Residual Networks. Our findings demonstrate that incorporating SE blocks consistently enhances classification accuracy across all tested models, underscoring the utility of channel attention mechanisms in refining feature representation for HAR tasks. Notably, the model architecture combining SE blocks with Residual Networks achieved the highest accuracy, increasing performance from 69.68 in baseline EfficientNet to 76.75, marking a significant improvement. Additionally, alternative models, such as EfficientNet integrated with Support Vector Machines EfficientNet-SVM and ZeroShot Learning models, exhibit promising results, highlighting the adaptability and potential of diverse methodological approaches for addressing the complexities of HAR. These findings provide a foundation for future research in optimizing HAR systems, with implications for enhancing robustness and accuracy in action recognition applications.

 Human Action Recognition HAR plays a crucial role in understanding and categorizing human activities from visual data, with applications ranging from surveillance, healthcare to human-computer interaction. However, accurately recognizing a diverse range of actions remains challenging due to variations in appearance, occlusions, and complex motion patterns. This study investigates the effectiveness of various deep learning model architectures on HAR performance across a dataset encompassing 15 distinct action classes. Our evaluation examines three primary architectural approaches: baseline EfficientNet models, EfficientNet models augmented with Squeeze-and-Excitation SE blocks, and models combining SE blocks with Residual Networks. Our findings demonstrate that incorporating SE blocks consistently enhances classification accuracy across all tested models, underscoring the utility of channel attention mechanisms in refining feature representation for HAR tasks. Notably, the model architecture combining SE blocks with Residual Networks achieved the highest accuracy, increasing performance from 69.68 in baseline EfficientNet to 76.75, marking a significant improvement. Additionally, alternative models, such as EfficientNet integrated with Support Vector Machines EfficientNet-SVM and ZeroShot Learning models, exhibit promising results, highlighting the adaptability and potential of diverse methodological approaches for addressing the complexities of HAR. These findings provide a foundation for future research in optimizing HAR systems, with implications for enhancing robustness and accuracy in action recognition applications.

 Human Action Recognition HAR plays a crucial role in understanding and categorizing human activities from visual data, with applications ranging from surveillance, healthcare to human-computer interaction. However, accurately recognizing a diverse range of actions remains challenging due to variations in appearance, occlusions, and complex motion patterns. This study investigates the effectiveness of various deep learning model architectures on HAR performance across a dataset encompassing 15 distinct action classes. Our evaluation examines three primary architectural approaches: baseline EfficientNet models, EfficientNet models augmented with Squeeze-and-Excitation SE blocks, and models combining SE blocks with Residual Networks. Our findings demonstrate that incorporating SE blocks consistently enhances classification accuracy across all tested models, underscoring the utility of channel attention mechanisms in refining feature representation for HAR tasks. Notably, the model architecture combining SE blocks with Residual Networks achieved the highest accuracy, increasing performance from 69.68 in baseline EfficientNet to 76.75, marking a significant improvement. Additionally, alternative models, such as EfficientNet integrated with Support Vector Machines EfficientNet-SVM and ZeroShot Learning models, exhibit promising results, highlighting the adaptability and potential of diverse methodological approaches for addressing the complexities of HAR. These findings provide a foundation for future research in optimizing HAR systems, with implications for enhancing robustness and accuracy in action recognition applications.

Article Details

A Comparative Study of Deep Learning Approaches for Human Action Recognition

Authors:

Gülsüm Yiğit

Publisher:

Murat YAKAR

Main Topics:

Software Engineering (Other)

Journal:

Turkish Journal of Engineering Vol.9 Issue 2, pp.281-289

Document Type:

Article

Publication Date:

6/2025

Document Language:

English

Document ID (Org No.):

2001339346

pISSN:

2587-1366

Deep Learning

Squeeze-and-Excitation Block

Residual Block

Zero-shot Learning

Human Action Recognition

Issues

Article Details

A Comparative Study of Deep Learning Approaches for Human Action Recognition

Authors: Gülsüm Yiğit

Publisher: Murat YAKAR

Main Topics: Software Engineering (Other)

Journal: Turkish Journal of Engineering Vol.9 Issue 2, pp.281-289

Document Type: Article

Publication Date: 6/2025

Document Language: English

Document ID (Org No.): 2001339346

pISSN: 2587-1366

Deep Learning

Squeeze-and-Excitation Block

Residual Block

Zero-shot Learning

Human Action Recognition

Issues

Authors:

Gülsüm Yiğit

Publisher:

Murat YAKAR

Main Topics:

Software Engineering (Other)

Journal:

Turkish Journal of Engineering Vol.9 Issue 2, pp.281-289

Document Type:

Article

Publication Date:

6/2025

Document Language:

English

Document ID (Org No.):

2001339346

pISSN:

2587-1366