Video-Based Abnormal Driving Behavior Detection via Deep Learning Fusions

Huang, Wei; Xi, Liu; Luo, Mingyuan; Zhang, Peng; Wang, Wei; Wang, Jin · 2019 · OpenAlex-citations

DOI: 10.1109/access.2019.2917213

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of automatically detecting abnormal driving behaviors from video footage, a critical component for ensuring driver safety and advancing autonomous driving systems. Abnormal driving is defined by the International Organization for Standardization as impaired driving ability due to focus on unrelated activities, categorized into physical comfort needs (e.g., eating), distraction needs (e.g., using mobile phones), and environmental distractions. The authors identify limitations in existing detection methods, such as physiological signal monitoring (high variance between individuals) and facial detail analysis (susceptibility to head movement and false positives). To overcome these issues, the study proposes three novel deep learning fusion models inspired by Densely Connected Convolutional Networks (DenseNet), aiming to improve detection accuracy and efficiency using high-resolution video clips captured by a single visible-light camera. The methodology introduces three specific architectures: the Wide Group Densely (WGD) network, the Wide Group Residual Densely (WGRD) network, and the Alternative Wide Group Residual Densely (AWGRD) network. These models integrate key deep learning concepts—depth, width, and cardinality—into the DenseNet framework. WGD replaces conventional convolutions with group and wide convolutions to enhance generalization without significantly increasing parameters. WGRD and AWGRD further incorporate residual network ideas, utilizing superpositions of previous layers to strengthen feature learning. The study compares these proposed models against five conventional deep learning architectures: Convolutional Neural Networks (CNN), Wide CNN, Group CNN, Deep Residual Networks (ResNet), and standard DenseNet. Experiments were conducted on a large abnormal driving database, treating the detection task as a multi-class classification problem. The results demonstrate that the proposed fusion models outperform existing popular deep learning models in video-based abnormal driving behavior detection. The extensive experiments substantiate the superiority of WGD, WGRD, and AWGRD through rigorous statistical comparisons. The integration of width and cardinality in WGD, along with the sophisticated residual connections in WGRD and AWGRD, effectively addresses the challenges of gradient vanishing and training efficiency inherent in deep networks. The findings indicate that these novel architectures provide better generalization capabilities and detection accuracy compared to traditional narrow or purely residual structures. The significance of this work lies in its first attempt to incorporate DenseNet-inspired fusion techniques into the domain of abnormal driving detection. By leveraging the strengths of dense connectivity, wide architectures, and residual learning, the study offers a robust solution for real-time monitoring of driver status. This approach avoids the high variance associated with sensor-based physiological detection and the instability of eye-tracking methods. The proposed models contribute to the field of intelligent transportation systems by providing a reliable, data-driven method for identifying distracted driving, thereby supporting the development of safer autonomous driving technologies and advanced driver assistance systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms