For ease of explanation, let me explain with human beings as example.
Humans are diploid, we have two chromosomes and at one loci we have alleles (one from each chr). Lets say the alleles are G and T.
When we examine a person on that loci if he has GG, then he is said to be G homozygous, if he has TT then he is T homozygous and if the observation is GT he is heterozygous in that loci.
When you examine 100 person and the frequency of alleles are like this
G - 80 % and T - 20 % then your minor allele is T in that particular loci.
In NGS context, the probable read distribution for each allele is 50 % in heterozygous individual. The MAF is more useful in downstream filtering for associating the variants with phenotype. Please read more on "common disease common variant hypothesis".