To overcome the limited generalization and insufficient real-time performance of existing bucket fill factor measurement methods, this paper proposes a measurement approach that integrates an enhanced RandLA-Net with Poisson surface reconstruction. A geometric feature enhancement (GFE) module and an illumination-aware module (IAM) are introduced to extract discriminative geometric features between the bucket and the material and to improve robustness under varying lighting conditions, respectively. In addition, a cross-feature fusion (CFF) module is developed to jointly integrate neighborhood spatial features, local differential features, and illumination-aware features, thereby improving semantic segmentation accuracy. Based on the point cloud segmentation results, the material point cloud and the material-covered bucket region are extracted, after which Poisson reconstruction is employed for fill-factor computation. Experimental results show that, on the validation set, the IoU of material, bucket, ground, and other classes increases by 9.78%, 8.84%, 6.71%, and 8.30% over the baseline RandLA-Net. The proposed method also generalizes well to independent test scenarios. The mean relative error of the fill-factor measurement is 3.02%, and the single processing time is approximately 1.43 s, demonstrating an effective balance between accuracy and real-time performance.