LFGNet: Low-level feature-guided network for drivers’ calling behavior detection

Drivers’ calling behavior probably leads to traffic accidents. Thus, there is a significant need to detect such distracted driving behavior. Mobile phones are easily overlooked in detection due to their small size and limited visibility. This paper proposes a low-level feature-guided network (LFGNet) for the drivers’ calling behavior detection, which integrates spatial and semantic information. Specifically, a cross-feature extraction (CFE) module is designed to extract low-level and high-level features by utilizing filters in dual directions. It leverages cross-attention with multi-branch convolutions (CMC) to learn and establish spatial representations between the extracted features from contextual information. As the network deepens, an excessive amount of low-level information can impede the network’s capacity for feature representation. Consequently, an irrelevant feature filter (IFF) module is introduced to selectively filter out irrelevant feature information. Two remote sensing datasets are employed to explore the effectiveness of the proposed method in detecting small objects in general. The public distracted driving datasets State Farm and SynDD1 are used for validation. Furthermore, the LFGNet is evaluated on a private driver’s calling behavior (DCB) dataset. The experimental results demonstrate the effectiveness of the proposed method in drivers’ calling behavior detection.