Abstract: To solve the problem of object detector that negative samples are much more than positive samples, a dynamic sample balancing algorithm is proposed. And based on this algorithm, an anchor free object detection model was constructed. Based on the position of feature points and object boxes in the feature map, the model selected the feature points responsible for predicting the object, calculated their cost matrix, and dynamically assigned positive and negative samples to the objects in the image. Therefore, the model did not require manual setting of anchor boxes, and it also improved the number of detectable targets and detection accuracy. In the VOC dataset, the model achieved an average accuracy of 91.3% and a detection speed of 32.1 FPS. Applying the model to airport apron support operations could achieve an average accuracy of 89.37%, and the detection accuracy of key targets was above 90%.