We use dynamic hard filter. We consider those regions suitable for variant call where read depth is greater than least 20%.

It depends. Some region are very hard to align with so even you get a 20X average genome coverage, you always get lower than 3X for some region.

It depends on different factors. For an eg, I wud say even one is enough if that one is uniquely aligned paired end with high base quality and mapping quality; even 5 reads in repeat region with improper pairing and low qualities is ambigous . So I would suggest to use mapping quality and read depth together.