Latches are very lightweight, short-term synchronization objects protecting actions that need not be locked for the life of a transaction. They are primarily used to protect a row while it is being read for a connection.
When the relational engine is processing a query, each time a row is needed from a base table or index, the relational engine uses the OLE DB API to request that the storage engine return the row. While the storage engine is actively transferring the row to the relational engine, the storage engine must ensure that no other task modifies either the contents of the row or certain page structures such as the page offset table entry locating the row being read. The storage engine does this by acquiring a latch, transferring the row in memory to the relational engine, and then releasing the latch.
SQL Server Performance Monitor has a Latches object that indicates how many times latches could not be granted immediately and the amount of time threads spent waiting for latches to be granted.