Abstract: As the most prominent problem in the field of object detection, occlusion and multi-scale seriously affect the recall and precision of the algorithm. To resolve the problems mentioned above, this paper starts from the receptive field, proposing an object detector based on the atrous convolution embedded feature pyramid network (ACFPN). Firstly, the atrous convolutional layers of different sizes are introduced into the feature pyramid to construct a hybrid receptive field module (HRFM), which aims to obtain more global feature information by increasing the receptive field with the number of parameters staging roughly the same, thereby solving the problem of occlusion; secondly, by improving the structure of the feature pyramid, we design a lower embedding feature pyramid module (LEFPM) to enhance the model’s scale adaptability, which combines shallow feature’s detail information and high-level feature’s semantic information to improve the richness and representation ability of feature maps; in particular, targing at the problem of missed detection, the Anchor Free mechanism of the fully convolutional one-stage (FCOS) algorithm is introduced to reduce the redundancy of candidate frames and further improve the positioning accuracy. The algorithm is tested on the public VOC dataset , and has shown a great improvement on detection accuracy.