Abstract: In order to improve the multi-scale detection ability of anchor-free object detectors and achieve the best compromise between detection speed and accuracy, a new anchor-free object detector with anti-aliasing ability is proposed. In the process of down-sampling, the grouped low-pass filters were used to prevent aliasing in the network, so as to improve the robustness of the model. The atrous convolutions with different atrous rates were used to fuse the multi-scale information of the objects and expand the receptive field of the neurons. The network parameters were prevented from being destroyed in the process of model training, and a better training model was obtained by using smooth L1 Loss function. The experimental results show that on Pascal VOC datasets, the mAP reaches 82.1%, and FPS reaches 32. Compared with the CenterNet model, the accuracy and speed of the model are improved. The mAP and FPS are improved by 4.3% and 18.5% compared with the original CenterNet-ResNet101 model.