In the robot grasping task in unstructured environment, it is important to acquire stable and reliable grasp pose of the object. In this paper, a dynamic multitarget 3D grasp pose detection approach based on deep convolutional network is proposed. Firstly, the Faster RCNN is utilized to conduct dynamic multitarget detection, and a stabilization detection filter is proposed to reject the noise and jitter in real time detection. Then, based on proposing depth target adapter, the GGCNN model is used to estimate the 2D grasp pose. Furthermore, the target detection result, 2D grasp pose and object depth information are fused to reconstruct the point cloud of the object, and calculate the 3D grasp pose. Finally, a robot grasping platform was established. The experiment results show that the statistical grasping success rate reaches 956%, which not only verifies the feasibility and effectiveness of the proposed approach, but also overcomes the defect of fixed and single result for 2D grasp pose.