<p>Supervoxels serve as a more natural and compact representation of 3D point clouds, allowing segmentation operations to be performed on regions rather than scattered points. Most current supervoxel segmentation methods rely on random sampling of representative seed points to generate supervoxels with fixed resolution. However, when seed points fall on the object boundary, erroneous growth of supervoxels may occur, leading to over-segmentation or under-segmentation issues. Additionally, the fixed resolution cannot adapt to different scenes. To address the issues in the concave-convex segmentation algorithm for supervoxels, this paper proposes a multi-feature fusion segmentation method based on an energy function. First, ideal seed points are selected based on the mean curvature of the local neighborhood. Then, an energy function incorporating normal and color information is employed to generate supervoxels and exchange boundary points. Finally, an entropy function based on the dimensional feature information of the local neighborhood is constructed to assist in computing the normal vectors of the supervoxels. The experimental results show that the generated supervoxels align more closely with object boundaries and exhibit high robustness across different scenes. Among the three test datasets, the segmentation method proposed in this paper outperforms other region-growing methods in terms of precision, recall, and mean intersection over union.</p>