In the next generation of wireless systems, Massive MIMO offers high angular resolution for localization. By virtue of large number of antennas, the Angle of Arrival (AoA) of User Terminals (UTs) can be estimated with high accuracy. According to Dense Multipath Component (DMC) channel model, local scatters around UTs can create different multipath signals for each antenna at the Base Station (BS). We obtain a deterministic form for the Cramer-Rao Lower Bound (CRLB) in a multi-user scenario when the contribution of the multipath signals is considered. We do this when the multipath signals are independent and identically distributed (i.i.d) with arbitrary distribution. Then, we redefine a localization efficiency function for a multi-user scenario and numerically optimize it with respect to (w.r.t) the number of antennas. We prove when only a subset of the available antennas is used, CRLB can be minimized w.r.t which set of antennas is used. Then, an antenna selection strategy that minimizes CRLB is proposed. As a benchmark, we apply the proposed antenna selection scheme to the MUltiple SIgnal Classification (MUSIC) algorithm and study its efficiency. Numerical results validate the accuracy of our analysis and show significant improvement in efficiency when the proposed antenna selection strategy is employed.