A Driver Distraction Behavior Recognition Method Integrating Upper-Body Skeleton Reconstruction and an Improved PoseC3D Model
Abstract
To address the challenges of small motion amplitudes, high similarity among upper-limb postures, and substantial variations in the duration of different driver distraction behaviors, this paper proposes a driver distraction behavior recognition method based on pose estimation and an improved PoseC3D model. First, YOLO11s is employed to localize the driver’s body region, and RTMPose is then used to extract two-dimensional human skeletal keypoints. Considering that the lower limbs are frequently occluded in in-vehicle scenarios and that distraction-related movements are primarily concentrated in the head and upper-body regions, an Upper-11 upper-body skeleton representation is constructed. Shoulder-center normalization, shoulder-width-based scale normalization, and geometric relational feature construction are then introduced to enhance the skeleton representation. On this basis, to address the limited capability of PoseC3D in modeling actions with different temporal durations and its insufficient utilization of cross-level features, a Factorized Multi-Scale Temporal Feature Fusion (F-MSTFF) module and a Residual Hierarchical Spatio-Temporal Fusion (R-HSTF) module are specifically proposed. F-MSTFF jointly models short-term and sustained action patterns through parallel branches with different temporal receptive fields, whereas R-HSTF enhances information complementarity across different network levels through progressive residual fusion of shallow-, intermediate-, and deep-level features. Experimental results show that the test-set accuracy is improved from 78.74% to 82.34%, and ultimately reaches 83.83%, demonstrating that the proposed method can effectively improve the recognition performance of driver distraction behaviors.