All dataset used for distance detection training and testing are acquired from the MTA (multi camera track auto) dataset
We have explored and tested multiple methods of distance detection non-machine-learning methods, including:
- assuming Cartesian coordinates for the ground plane and calculate the euclidean distance.
- assuming Polar coordinates for the ground plane and calculate the euclidean distance
- using space transformation to transform the 3d camera plane (pre-configured using the existing camera data) to a 2d one.
The final method that we employed is the Cartesian method as it provides the best accuracy among all possibilities.