Non-cooperative target pose measurement method, system and terminal equipment based on deep neural network
A deep neural network and non-cooperative target technology, applied in the field of close-range non-cooperative target pose measurement, to meet real-time requirements, improve algorithm speed and matching accuracy, and achieve high-precision effects
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0028] The PointNet network is aimed at point cloud classification and segmentation tasks in the field of stereo vision, and has achieved good results in various mainstream databases. The starting point of the algorithm design is to solve the disorder problem of point cloud data. In the initial stage, each point is treated identically and independently, and in the basic setup, each point consists of its three coordinates. The key to this method is to use max pooling as a symmetric function, so that the extracted feature vector can ignore the disorder of point cloud data.
[0029]Due to the out-of-order problem of point cloud data, it is difficult for the network to learn a consistent mapping from input to output. In order to make the neural network model not be affected by the sorting of input data, there are three solutions: the first is to sort the input data in a canonical order, which sounds simple, but in high-dimensional space, there is actually no one Stable point per...
Embodiment 2
[0054] Such as figure 2 As shown, the present invention also provides a non-cooperative target pose measurement system based on a deep neural network, including:
[0055] The first unit: down-sampling the data point cloud P and the model point cloud Q of different angles of the non-cooperative target to obtain the point cloud P', Q';
[0056] The second unit: use the trained PointNet network model to perform feature extraction on the point cloud P', Q' to obtain the feature matrix A containing the feature vector of each point n×1024 and the global eigenvector B 1×1024 ;
[0057] The third unit: filter the feature points of the point cloud P', Q' according to the preset feature point detection and screening threshold, and according to the global feature vector B 1×1024With the point cloud P', Q' the eigenmatrix A containing the eigenvectors of each point n×1024 Perform feature point matching to obtain feature point sets P", Q";
[0058] The fourth unit: perform point clou...
Embodiment 3
[0068] This embodiment is a simulation verification of a non-cooperative target pose measurement method based on a deep neural network. The performance of the above algorithm is evaluated by using the Bunny 3D point cloud data in the Stanford University Graphic Laboratory dataset. In the experiment, the results of feature point detection are visualized first, and then the algorithm is compared with the traditional geometric feature descriptor FPFH to verify the real-time performance and accuracy of the algorithm.
[0069] Such as image 3 Shown is a schematic diagram of the original data before pose estimation in this embodiment. by right image 3 The original data in is down-sampled to get a new point cloud.
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More