RDG-Three-Branch Network With GAF-Based 1-D-to-2-D Image Transformation for Pedestrian Scene Recognition

Scene recognition technology has been widely applied and plays a crucial role in various fields, including transportation, elderly care, smart homes, environmental monitoring, and autonomous driving. Wearable sensor devices have emerged as the core data source for pedestrian scene recognition technology, owing to their significant advantages such as low invasiveness and real-time data collection capabilities. However, constrained by issues like inadequate utilization of multimodal data and a narrow range of recognizable scenes, current technologies still struggle to achieve high recognition accuracy in fine-grained scenes. To address these challenges, this study proposes an RDG-Three-Branch Network for pedestrian scene recognition, which is based on a Gramian Angular Field (GAF)-based 1D-to-2D image conversion algorithm. This method overcomes the limitation of traditional approaches that can only recognize static scenes, enabling the further identification of fine-grained dynamic scenes where pedestrians are situated. Specifically, the proposed method first performs time synchronization and data filtering on six types of sensors embedded in wrist-worn and foot-worn devices to eliminate noise interference. Subsequently, the temporal sequence data is converted into 2D images via GAF to enhance the feature representation of different scenes, which are then fed into the designed RDG-Three-Branch deep learning network. Three branches within the network—the residual branch, depthwise separable branch, and global dependency branch—extract features of the 2D images from distinct perspectives, ultimately realizing the classification of pedestrian scenes. Experimental results demonstrate that the proposed method can effectively distinguish 14 types of pedestrian activity scenes both during the day and at night. Compared with two control groups using a single type of sensor device, the recognition accuracy is improved by 5.27% and 20.57%, respectively. Meanwhile, the performance of the designed model outperforms state-of-the-art (SOTA) models such as Swin-Transformer, Max-ViT, and DeiT, with accuracy improvements ranging from 1.05% to 8.41%. This work achieves fine-grained and high-accuracy multi-scene recognition for pedestrians.