Aiming at the low precision and efficiency of the fruit identification and harvesting motion of existing harvesting robots, the hand-eye coordination planning with deep visual servo for harvesting robot was carried out. A small lifting harvesting robot with deep visual servo of RealSense-in-hand was developed, which was composed of the autonomous vehicle, lifting bin, electric fork lifter, grip-cut integrated end-effector, and 3-freedom manipulator. The workspace and posture analysis of fruit picking and placing was performed, and the coordinate transformation model of hand-eye coordination was established for the eye-in-hand mode. Based on the depth visual servo of RealSense-in-hand, the far-to-close hand-eye coordination strategy was proposed for the harvesting robot. The canopy detection from a distance, sub-region division and location, close-range fruit accurate identification and positioning were effectively combined, so that the step-by-step visual guidance of the manipulator was realized. According to the parameters of RealSense and the manipulator, the segmented motion planning between far-to-close key points based on depth vision was completed. It was shown in the hand-eye coordinated harvesting test that, the average positioning accuracy of the end-effector in the X, Y and Z directions was 3.51mm, 2.79mm and 3.35mm, respectively. The average time consuming was 19.24s, which included the motion time of the manipulator from the initial position to picking position (12.04s), the fruit recognition and computing time (3.82s), and the fruit placing time (7.2s). The time of manipulator motion accounted for 80.2% of the whole cycle. Both the robot structure and the hand-eye coordination strategy based on depth visual servo of RealSense-in-hand can meet the needs of fruit harvesting operation.