关键词: 交通运输工程；姿态辅助；驾驶行为检测；轻量化；胶囊网络

Abstract: Distracted driving is one of the primary causes of traffic accidents. To improve the speed and accuracy of detecting distracted driving behaviors, a lightweight driving behavior detection network based on pose assistance was proposed. To address the issue of poor feature extraction quality, an efficient large-kernel self-attention mechanism was designed to enhance the ability to capture both local and global features, thereby extracting rich low-level features. Meanwhile, grouped convolution was combined with capsule network to extract semantic features of driving behavior, which reduced the amount of model parameters while ensuring high accuracy. Furthermore, the detection accuracy of network in complex backgrounds was further improved by introducing pose estimation as an auxiliary. Experiment results show that the proposed method achieves accuracy of 99.71% and 95.38% on the SFD and AUC benchmark datasets, respectively; compared to current advanced models, while maintaining the same accuracy, the parameter count is only 0.29M (a reduction of 61.8%), and the inference speed in a server with a throughput of 801 images/s is about 2.57ms. The proposed lightweight driving behavior detection network based on attitude assistance achieves relatively high accuracy and its parameter amount satisfies the demands of embedded devices, providing support for safe driving.

Key words: traffic and transportation engineering; attitude assistance; driver behaviour detection; lightweight; capsule network