Consider using different edit distances (cyclic and rotational variant to counter translation for location aware hashes)

[KilianB]

Most of the non frequency based hashes encode a subset of the image into a single bit (the hash maps 1 bit of information to specific location). For motion tracking we use sliding windows to compare hashes. The hamming distance might be off due to simply being translated a bit. Try to find a new edit distance accounting for bigger swaps in the binary string.

Maybe:

1. naive levenshtein distance
2. N-Grams (https://www.researchgate.net/publication/237107253_Using_q-grams_in_a_DBMS_for_Approximate_String_Processing) ?
3. cosine similarity
4. https://ieeexplore.ieee.org/document/4708948 Edit distance with block swapping

[KilianB]

http://www.academia.edu/4400094/A_Better_Edit_Distance_Measure_Allowing_for_Block_Swaps