WDDR-YOLO: a Driver Distraction Detection Algorithm.

Saved in:

Bibliographic Details
Title:	WDDR-YOLO: a Driver Distraction Detection Algorithm.
Authors:	Lu, Yan¹ wylylu@163.com, Xing, Zhongyu¹ xingzhongyu@gxjcxy.edu.cn, Li, Meng¹ 1052590135@qq.com, Liu, Guoqiang² 15693881490@163.com
Source:	Engineering Letters. May2026, Vol. 34 Issue 5, p1744-1756. 13p.
Subjects:	Distracted driving, Discrete wavelet transforms, Object recognition (Computer vision)
Abstract:	To address the challenge that deep visual models struggle to attain a well-tuned equilibrium between detection accuracy and inference efficiency in resource-limited operational contexts, this study proposes a frequency-domain dynamic reparameterization model, WDDR-YOLO. This method takes YOLOv11 as the baseline, with core innovations as follows: 1) Haar Wavelet Decomposition (HWD) is adopted to replace spatial downsampling, enabling the reorganization of features into the frequency-domain channel dimension; 2) A frequency-domain energy-aware Reparameterization Convolutional Structure (RCS) is designed, which extracts frequency-domain features through multiple branches during training and dynamically fuses them into a single convolution kernel during inference; 3) A collaborative mechanism between Multi-Local Channel Attention (MLCA) and Dynamic Upsampling (DySample) is constructed, enhancing the ability to retain high-frequency details through frequency-domain weighting and offset constraints. Experimental results for the driver distraction detection task demonstrate that WDDR-YOLO attains an mAP@50 of 87.0% on the State-Farm dataset, a metric that outperforms the baseline by a margin of 3.9 percentage points, alongside a 17.6% cut in model parameters and a rise in inference speed to 81.4 FPS. It further attains a value of 83.5% for mAP@50 on the AUC DDD dataset, which validates the model's generalization capability across heterogeneous scenarios. The proposed frequency-domain dynamic reparameterization paradigm provides an extensible new approach for the design of lightweight visual perception models. [ABSTRACT FROM AUTHOR]
	Copyright of Engineering Letters is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	To address the challenge that deep visual models struggle to attain a well-tuned equilibrium between detection accuracy and inference efficiency in resource-limited operational contexts, this study proposes a frequency-domain dynamic reparameterization model, WDDR-YOLO. This method takes YOLOv11 as the baseline, with core innovations as follows: 1) Haar Wavelet Decomposition (HWD) is adopted to replace spatial downsampling, enabling the reorganization of features into the frequency-domain channel dimension; 2) A frequency-domain energy-aware Reparameterization Convolutional Structure (RCS) is designed, which extracts frequency-domain features through multiple branches during training and dynamically fuses them into a single convolution kernel during inference; 3) A collaborative mechanism between Multi-Local Channel Attention (MLCA) and Dynamic Upsampling (DySample) is constructed, enhancing the ability to retain high-frequency details through frequency-domain weighting and offset constraints. Experimental results for the driver distraction detection task demonstrate that WDDR-YOLO attains an mAP@50 of 87.0% on the State-Farm dataset, a metric that outperforms the baseline by a margin of 3.9 percentage points, alongside a 17.6% cut in model parameters and a rise in inference speed to 81.4 FPS. It further attains a value of 83.5% for mAP@50 on the AUC DDD dataset, which validates the model's generalization capability across heterogeneous scenarios. The proposed frequency-domain dynamic reparameterization paradigm provides an extensible new approach for the design of lightweight visual perception models. [ABSTRACT FROM AUTHOR]
ISSN:	1816093X