Improved Pose Estimation Network Based on Spatial Registration Model
Zexiang Liu, Ziao Dong, Xin Liang Yin, Yanbing Liang · Journal of engineering system. · 2024
Position and attitude estimation refer to estimating the distance and attitude between the object to be measured and the sensor device from the input information captured by the sensor device. The traditional computer vision processing method to optimize the feature vector extracted by the feature extraction algorithm often requires a large number of resources to optimize the model. The introduction of deep learning provides a new solution. In this paper, we use the depth learning Mask-R-CNN framework to recognize and segment the object, and obtain the size and contour features of the object; Then VGG-16 algorithm is used to extract RGB image features; Then, the mask information extracted by Mask R-CNN is fused for convolution and pooling; Finally, the convolved feature map is collected into two fully connected layers to predict the translation matrix and rotation matrix respectively. At the same time, after the fusion, the feature map is sampled up and convolved, and finally the feature map with the same size as the original input image is obtained. The output is expected to be consistent with the input mask. By establishing the sensor space registration model, the least square method and the generalized least square method are used to estimate the error parameters and compensate them to the sensor system, so as to reduce the error impact caused by the equipment observation data and obtain more accurate position and attitude information. The experimental comparison shows that the average proportion of the traditional Pose CNN to predict the object's position and attitude error is 64.6 within 8 cm. The average accuracy of the position and attitude estimation network studied in this paper is 81.5, and the position and attitude estimation network have better generalization ability.