Driver behavior detection via adaptive spatial attention mechanism. (April 2021)

• Record Type: Journal Article
• Authors: Zhao, Lei; Yang, Fei; Bu, Lingguo; Han, Su; Zhang, Guoxin; Luo, Ying
• Is Part Of: Advanced engineering informatics. Volume 48(2021)
• Publication Date: 2021-04
• Subjects: Driver behavior detection -- Adaptive spatial attention mechanism -- Class activation map -- Multiscale feature fusion; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2021.101280
• Languages: English
• ISSNs: 1474-0346

Abstract

Abstract: Drivers still play an important role in driving safety despite the presence of driverless vehicles. Over the last few years, millions of deaths are due to traffic accidents, and more than half of these accidents worldwide are caused by distracted driving. Therefore, driver behavior detection during driving is crucial. A novel driver behavior detection system based on the adaptive spatial attention mechanism is proposed in this study. This system realizes the extraction of adaptive discriminative spatial regions of driver images by cascading multiple attention-based convolution neural networks. Feature representation in each subnetwork is extracted from the output layer, and the discriminative region of the input image is cropped using class activation maps. The obtained region is then fed into the next subnetwork to highlight important region for improving the system performance. The model starts from full images and iteratively crops the region adaptively from coarse to fine to extract the feature representation at multiscales. Finally, the k-nearest neighbor classifier is applied to classify the cascaded multiscale features and obtain the category of driver behavior. The systems are evaluated on a driver behavior recognition database captured in actual driving environments. Experimental results indicate that our systems can achieve superior recognition performance to other state-of-the-art methods and can run in real-time with simplified structure and model in our platform.