3D object detection has important applications in fields such as autonomous driving. However, data labeling is expensive, and training a high-performance 3D detector requires an expensive and time-intensive labeled dataset. Aiming at the problems of misleading training with fixed threshold and edge blurring caused by traditional point cloud enhancement in the existing semi-supervised methods, an Adaptive supervision and Feature Guidance (Adaptive Supervision and Feature Guidance, ASFG) method is proposed. This method innovatively uses global information to dynamically adjust the threshold, combines with the teacher network to verify the reliability of the pseudo-label, and maintains the integrity of the target through the feature guidance method. Experiments on KITTI and Waymo datasets show that when using only 1% labeled data, ASFG improves the detection performance by 19.2% compared with the baseline, and the pedestrian detection indicators even surpass the fully supervised model. Especially, the performance of pedestrian and cyclist detection is improved by 24.1% and 25.9% respectively compared with the baseline.