This study presents DMSA-YOLOv8, a lightweight detector designed to improve the representation of subtle facial cues in driver emotion recognition. The proposed Dynamic Multi-Scale Attention (DMSA) module combines Dynamic Channel Attention (DCA), which recalibrates emotion-relevant feature channels, with Multi-Scale Spatial Attention (MSA), which aggregates spatial context at multiple receptive fields. The standard CIoU bounding-box loss is replaced by Minimum Point Distance Intersection-over-Union (MPDIoU) to provide an additional corner distance-based localization signal. The complete training objective jointly optimizes MPDIoU regression, emotion classification, and distribution focal loss. Under the standard assumptions used in non-convex stochastic optimization, the conventional convergence guarantee for stochastic gradient descent remains applicable; a complementary complexity analysis quantifies the additional cost of DMSA. In the reported experiments, DMSA-YOLOv8 achieves a mAP@0.5 of 94.0% with 3.12 M parameters and 8.3GFLOPs, improving mAP@0.5 by 3.3 percentage points over the YOLOv8 baseline. These results indicate a favorable accuracy–complexity trade-off on the evaluated data, while broader cross-subject evaluation remains necessary.