Research On Multi-type Cooperative Targets Detection And Pose Estimation Approach

The urgent demands on the three-dimensional(3D)geometric parameters precise measurement of large components such as satellite antennas,aircrafts and oil tanks have been placed by the rapid development of large-scale equipment and high-end manufacturing industry.However,the existing measuring instruments have the problems like the insufficient range,the single measurement mode,the low measuring accuracy,and particularly the low aiming accuracy of the cooperative targets due to the large size and complex structure of the components measured,the high measuring accuracy,and the changeable measure environment.Therefore,it is of great significance both in theory and practice to research the multi-type cooperative targets detection and pose estimation approach for improving the aiming accuracy of the cooperation targets and providing the foundation for aiming the multi-type cooperative targets on the application of the 3D geometric parameters precise measurement of large components.Based on the detailed analysis of existing methods,this thesis proposes a Dense Connection and Spatial Pyramid Pooling Based YOLO(DC-SPP-YOLO)object detection method.In this method,the generative adversarial network is used for data augmentation;based on YOLOv2,an improved dense connected convolutional layer is employed to replace the original convolutional layers of the backbone network,a new spatial pyramid pooling is introduced to extract the multi-scale image features,a new loss function using the cross entropy to represent the loss of classification is constructed,and the DC-SPP-YOLO model is trained for object detection.Compared with YOLOv2,DC-SPP-YOLO ameliorates the accuracy of object detection by strengthening the propagation of feature information flow in the network and utilizing the richer image features fully.On this basis,the camera is calibrated,the different geometric features of multi-type cooperative targets detected are extracted,and the pose estimation method based on nonlinear optimized EPnP is proposed.In this pose estimation method,pose parameters of the cooperative target estimated by EPnP are taken as the initial values;the LM algorithm,of which the constrained optimization problems is transformed into the unconstrained optimization problems for simplifying the solution process by the lie algebra,is used for improving the accuracy of pose estimation while maintain a fast speed by a few iterations.The experimental results show that,the object detection accuracy of DC-SPP-YOLO proposed is 1.20%-2.25%higher than that of YOLOv2 by using the PSACAL VOC datasets and UA-DETRAC datasets to test,and is 5.89%higher than that of YOLOv2 by using the multi-type cooperative targets dataset to test;the detection speed of DC-SPP-YOLO is up to 58.9 fps.Using the images of the cooperative targets detected to test,the pose estimation error of the nonlinear optimized EPnP method is reduced by 0.024-0.395 pixels compared with that of the EPnP method while the pose estimation speed is close to that of the EPnP method.Generally,the approach proposed in this thesis improves the accuracy of object detection and pose estimation while maintaining a fast speed,which achieves the high-precision and real-time aiming of cooperation targets in the 3D geometric parameters precise measurement of large components.