Detection of a specific ship with known ship 3d model

Hello community,

I want to detect a specific ship in sea from a low flying aerial vehicle, equipped with a camera (e.g., a drone). Suppose the ship is this. In general, the ship should remain “constant” in terms of appearance, apart from the numbering, that might change appearance and figures, depending on the specific ship. The ship might be in different sea states (meaning various sea textures) and with different shores and coastlines in the background.

The drone is approaching the ship always from the back, hence I need to detect only the back of the ship.

I was thinking to approach the problem similar to a template matching problem:

  • generate template of the ship’s back: get a wireframe-like representation of the back of the ship from the 3D model (I have access to it), a sort of binary image with only the silhouette edges and some other edges visible from the back of the ship;
  • find template in image from drone’s camera:
    • convert to greyscale
    • preprocessing to remove some noise (e.g., gaussian filter)
    • edge detector
    • chamfer matching between the template and the image

Last info: I cannot use deep learning based methods.

Do you think it is a resonable approach? Can you suggest other options?

why no DL?


Due to the inability to certify DL for critical applications

cognition becomes “complex” at “complex” levels. if complexity is a hurdle to certification, you’ve got a real roadblock.

you want to detect a ship from nothing but a color (grayscale?) picture, that requires machine learning at minimum.

if you can’t do DL, then at least look at ML that isn’t deep, but then it won’t perform well or at all.

I am perfectly aware on DL and ML-based methods, their potentialities and use cases, i use them in my day to say job routines. I am telling you that in this scope it is not possible to rely on these techniques. By the way, template matching, silhouette matching and other non-learning-based techniques are used in similar context and requirements, hence the question has a context and makes sense.
If you do not have any experience of any other technique than ML-based in safety critical problems with similar requirements and environment, that is fine. But I am looking for some help here.