Star tracker using vector-based deep learning for enhanced performance
Star tracker systems and methods are provided. The star tracker incorporates deep learning processes in combination with relatively low cost hardware components to provide moderate (e.g. ˜1 arc second attitude uncertainty) accuracy. The neural network implementing the deep learning processes can include a Hinton's capsule network or a coordinate convolution layer to maintain spatial relationships between features in images encompassing a plurality of features. The hardware components can be configured to collect a blurred or defocused image in which point sources of light appear as blurs, and in which the blurs create points of intersection. Alternatively or in addition, a blurred or defocused image can be created using processes implemented as part of application programming. The processing of collected images by a neural network to provide an attitude determination can include analyzing a plurality of blurs and blur intersections across an entire frame of image data.
1. A method, comprising: obtaining a plurality of frames of training image data, wherein each frame of training image data is associated with an attitude and contains image information associated with a plurality of stars, wherein at least some of the frames of training image data include blurred image data, and wherein the blurred image data includes blurs corresponding to at least some of the stars in the plurality of stars;
for each of the frames of training image data, including the frames including blurred image data, storing information related to relative locations of image features in the frame of training image data, wherein the image features include centroids of the blurs corresponding to the at least some of the stars and points of intersection between different blurs corresponding to the at least some of the stars;
training a vector based deep learning model using the information related to relative locations of the image features in the frames of training image data, including the frames including blurred image data, to associate different relative locations of image features in the frames of training image data with different attitudes;
operating a star tracker to obtain a frame of star tracker image data, wherein the star tracker image data is blurred;
processing the frame of blurred star tracker image data using the trained vector based deep learning model to determine a relative location of each of a plurality of image features present in the frame of blurred star tracker image data; and
applying the determined relative location of at least some of the plurality of image features present in the frame of blurred star tracker image data to determine an attitude of the star tracker, wherein the image features applied to determining the attitude of the star tracker include centroids of the blurs corresponding to at least some of the stars, and points of intersection between different blurs corresponding to the at least some of the stars.
2. The method of claim 1 , wherein the star tracker image data is blurred at an imaging plane of a detector of the star tracker.
3. The method of claim 2 , wherein the star tracker image data is blurred by a defocus module implemented by the star tracker.
4. The method of claim 1 , wherein the star tracker image data is blurred by a defocus module implemented by the star tracker.
5. The method of claim 1 , wherein the training image data is defocused.
6. The method of claim 1 , wherein the vector based deep learning model includes a Hinton's capsule network.
7. The method of claim 1 , wherein the vector based deep learning model includes a coordinate convolution layer.
8. The method of claim 1 , wherein the star tracker image data includes blurs associated with at least one hundred stars.
9. The method of claim 8 , wherein the star tracker image data is obtained using a detector having thousands of pixels.
10. The method of claim 1 , wherein the star tracker is carried by a platform, and wherein the star tracker image data is obtained while the platform is deployed in space.
11. The method of claim 1 , further comprising storing information related to an attitude associated with each of the frames of training image data.