Forest 3D object classification in large-scale Terrestrial Laser Scanning(TLS) data is a challenging work, since the large quantity and varying local density of points. In order to adapt to the characteristics of forest TLS data, this paper improves RandLA-Net neural network by including relative elevation and change of curvature. We also compare the classification performance of traditional machine learning and deep learning. Experimental results indicate that the RandLA-Net neural network has better performance than the Random Forest(RF) classifier, and the proposed method provides an effective result for points classification, achieving M_IoU of 93.3% and overall accuracy (OA) of97.8%, respectively.