<p>3D point clouds detection plays a critical role in autonomous driving, robotic control, intelligent manufacturing et al. Existing methods with single-sampling strategies may focus on long-range context, but lose fine-grained local details during feature aggregation. Moreover, the inherent sparsity of point cloud data also limits effective multi-scale feature learning, leading to degraded feature representations and inaccurate object location, particularly in complex scenarios. Thus, a multi-window parallel voxel transformer with center optimization strategy network is proposed in this paper, which integrates multiple sampling strategies and center optimization strategy into the transformer-based architectures to enhance multi-scale feature representation. Specifically, a multi-scale cross-attention module is introduced to fuse voxel features across different scales through parallel self-attention and multi-window sampling, enabling effective aggregation of both local and global information. Additionally, positional encoding is incorporated to enhance spatial awareness, improving the effectiveness of hybrid-scale feature aggregation. Furthermore, a novel center optimization strategy to mitigate localization errors caused by imprecise feature aggregation. Instead of relying on single voxel-based center, multiple candidate object centers are generated from high-confidence keypoint features and refined using confidence-weighted averaging. This confidence-aware refinement produces a more accurate and robust object center estimation, improving bounding box localization performance in sparse, occluded, or boundary-obscured scenarios.</p>