Robot grasping technology is favored by scientific researchers because of its importance in the field of robotics and its prospects. The main challenge of robotic grasping is to design an effective grasping detection algorithm for accurate grasping of the robot. Recently, various robotic grasping techniques have been proposed to improve the grasping performance of the robot. However, for random objects in an unstructured environment, even if accurate target classification is performed, multiple targets may overlap, which leads to a decline in the grasping performance of many robot grasping methods and cannot meet the accuracy requirements. To solve this problem, we propose an innovative grasping detection algorithm, which we call MR-GPD. MR-GPD is based on the robot grasping gesture detection algorithm (GPD), embeds the Mask-RCNN deep learning model to classify and segment the grasped target, then combine the results of target segmentation to construct a grasping candidate screening mechanism to evaluate and filter the outputted grasping candidates, and retain the best grasping candidate plan in the planned grasping task. To meet the requirements of the robots actual grasping task, an improved Mask-RCNN neural network model is proposed, the speed and accuracy of target segmentation in the actual grasping process of the robot are improved, and the accuracy of the model reaches 97.5%. In this article, we conducted experimental evaluations through experimental tests and Baxter robots. The experimental results show that the method has a good grasp effect on randomly placed objects in unstructured environment, and the average success rate is about 91.2%, it reveals the feasibility of our proposed MR-GPD method, which is expected to provide important information for robot grasping in unstructured environment.