For the problem of occlusion interference in robot grasping of occluded targets, an improved YOLO-CA-SD and semantic segmentation occluded target detection model and grasping method are proposed to complete the grasping when multiple targets and non-target objects occlude and interfere with each other. Firstly, the model adds a coordinate attention to YOLOv5l, considers the problem of detection frame matching direction based on the loss function, adds angle information between frames, and detects the original model decoupling is partially performed to reduce information loss caused by coupling. Secondly, an improved DeeplabV3+ target segmentation model was proposed. The original DeeplabV3+ backbone network was replaced by MobileNetV2 to reduce the model complexity. A CA module was added to the Atrous Spatial Pyramid Pooling structure to fuse pixel coordinate information to improve segmentation accuracy and solve the occlusion interference problem. Finally, the end rotation angle of the target poses relative to the template pose and the optimal grasping point are obtained by point cloud registration. The performance test is carried out on the self-built 2 750 commonly used tool occlusion data set. The experimental results show that the improved model improves the detection accuracy by 0.052%, 0.968%, 6.000%, and 7.400% on mAP@0.5, mAP@0.5:0.95, 60% target object occlusion rate and 60% non-target object occlusion rate datasets. The improved semantic segmentation model on this basis improves the segmentation speed and MIOU by 33.45% and 0.625%, and the ABB IRB1200 robotic arm is used to realize the experiments on the grasping of obscured targets, which verified the feasibility and practicality of the method.