In the field of 3D detection, mainstream 3D feature extraction methods usually follow the paradigm of voxelization and downsampling to BEV (Bird’s Eye View). The feature extraction process is crucial for the quality of BEV during dimension reduction. However, mainstream methods encounter information disconnection issue when dealing with excessively sparse voxels, preventing the extraction of sufficient geometric features for large-scale object detection before dimension reduction. In this paper, we propose a novel convolutional module, termed Asterisk Sparse Convolution (Asterisk Conv), to address the aforementioned issue. Additionally, we have devised a lightweight feature extraction network to extract and balance features across various scales, enhancing overall detection accuracy. Our method achieves competitive accuracy on large targets such as trucks and buses, and significantly improves overall accuracy compared to baseline on the nuScenes benchmark.