A high-precision and highly generalized method of detecting facial small object of driver based on YOLOv7 algorithm was proposed to address the issues of falsely detecting facial small target and the low accuracy of detection caused by vibration and background occlusion for tractor drivers in complex environments of agriculture.An improved spatial pyramid pooling module AS_SPPFCSPC was used to replace SPPCSPC to effectively aggregate low-frequency global information and high-frequency local information to enhance the accuracy of facial localization for drivers.The cross-level partial network module VoVGSDCSP was used to replace the E-ELAN module in the neck network to achieve higher computational efficiency of the algorithm.The structure of detection layer was adjusted and a new detection head SC_C_detect was introduced to improve the ability to extract small target features.The results of ablation and comparative experiments showed that the improved algorithm had a single-image detection time of 7.8 ms， with mAP@0.5 at 97.29% and mAP@0.5：0.95 at 69.45%， superior to object detection algorithms including Faster RCNN， YOLOv5l， and YOLOv8l.The results of generalization experiments conducted on tractors at levels of different vibration showed that the background error and localization error of the facial small target detection model after improvement were effectively reduced.It is indicated that the algorithm proposed combines real-time and accuracy， with good generalization performance at different levels of vibration.