To address the issues of facial small target false detection and low detection accuracy caused by vibration and background occlusion for tractor drivers in complex agricultural environments, this study proposes a facial small target detection method for tractor drivers based on improved YOLOv7, termed YOLO-SOD. Firstly, in the neck network, the improved Spatial Pyramid Pooling module AS_SPPFCSPC is utilized to replace SPPCSPC, effectively aggregating low-frequency global information with high-frequency local information to enhance the accuracy of facial localization for drivers. Secondly, the Cross-Level Partial Network module VoVGSDCSP is employed to replace the E-ELAN module in the neck network, achieving higher computational efficiency. Finally, the 20 pixel × 20 pixel large target detection layer P5 is removed, and a new 160 pixel × 160 pixel small target detection layer P2 is added to enhance the feature extraction capability for small targets. Additionally, a new detection head SC_C_detect is introduced to improve the computational efficiency of the model. Experimental results demonstrate that the improved algorithm achieves a single-image detection time of 7.8 ms, with AP0.5 at 97.29% and AP0.5:0.95 at 69.45%. Compared to the baseline model, there is an improvement of 2.49 and 6.83 percentage points respectively. Compared to current mainstream object detection networks Faster-RCNN, YOLOv5l, and YOLOv8l, the AP0.5 is increased by 6.79, 3.99, and 0.59 percentage points respectively, with model sizes reduced by 106.003, 15.956, and 11.346M. The improved facial small target detection algorithm exhibits high detection accuracy and inference speed, providing technical support for fatigue monitoring and safety warning systems for tractor drivers. Keywords: tractor; driver; facial detection; small target detection; YOLOv7.