Intelligent grasping robot can complete high intensity work instead of human beings. In order to realize the accurate positioning of objects and improve the success rate of robot grasping, the robot grasping system based on region of interest was studied. Firstly, the system calibrated the depth camera, then, the loss function of the deep convolution neural network was improved, and the focus function was used instead of the traditional cross entropy function to train the model to get the category of the target and the pixel coordinate value and depth value of the target in the two-dimensional envelope box. Secondly, the coordinates of the depth sensor were transformed into the manipulator coordinate system by hand-eye calibration method, and the positioning of the object was completed according to the imaging principle of the camera. Finally, the joint angle was solved by inverse kinematics of the robot, which drove the robot to grasp. The experimental process was analyzed and verified on the aubo_i5 manipulator. The experimental results show that the target recognition and positioning error is small, the average accuracy is increased by 2.36%, the average success rate of grasping is 93.4%, which is 13.4% higher than that before improvement, which can meet the needs of robot grasping.