There are two difficulties in the application of robot grasping in industry. How to detect the graspable object accurately and how to select the optimized grasp target among the detected multiple objects. In this paper the homoscedastic uncertainty is introduced into Keypoint RCNN to learn the weights of various losses, the attention modules are integrated into feature extractor, which composes the improved Keypoint RCNN model. A two-stage object grasp detection algorithm is proposed based on the improved Keypoint RCNN model. In the first stage, the improved model is used to predict the masks and keypoints. In the second stage, the masks and keypoints are used to compute the grasp representation and overlap rate of the object, the overlap rate represents the level of collision while grasping. According to the overlap rate, the optimized grasp target can be selected from multiple graspable objects. Comparison experiment indicates that the performances of the improved Keypoint RCNN model are improved in object detection, instance segmentation and keypoint detection compared with those of original model, and the average precisions (AP) on the self-built dataset reach 85. 15% , 79. 66% and 86. 63% , respectively. Robot grasping experiment proves that the proposed grasp detection algorithm can accurately calculate the grasp representation, select the optimized grasp and guide the robot to grasp the target with collision-free grasp.