Robotic grasping in multi-object stacking scenes is an important aspect of robot intelligence. In this paper, an object recognition method based on 3D point cloud is proposed, which can be used for robot to recognize the object in the stacking scenes and complete the subsequent grasping operation. Firstly, the classical PointNet++ model is improved to make it more suitable for stacking objects in terms of feature sampling quantity and accuracy. Then, based on the transformation relationship between CAD model and point cloud, a dataset containing 200 sets of stacking scene point cloud was build. Using the dataset to train the improved PointNet ++ model, we get 96.97% training accuracy and 87.47% testing accuracy, which are 0.37% and 1.08% higher than the classical PointNet ++ model respectively. Finally, 30 sets of point cloud data of stacking objects were captured with a 3D camera in real scenes to validate the segmentation performance of the model. Experiments show that the average segmentation accuracy of this method is 84.74%, and the highest segmentation accuracy can reach 90.75%, which satisfies the requirements of industrial applications.