ABSTRACT Considering the upcoming OWFA, we use simulations of the foregrounds and the z = 3.35 H i 21-cm intensity mapping signal to identify the (kā„, kā„) modes where the expected 21-cm power spectrum P(kā„, kā„) is substantially larger than the predicted foreground contribution. Only these uncontaminated k modes are used for measuring P(kā„, kā„) in the āForeground Avoidanceā technique. Though the foregrounds are largely localized within a wedge. we find that the small leakage beyond the wedge surpasses the 21-cm signal across a significant part of the (kā„, kā„) plane. The extent of foreground leakage is extremely sensitive to the frequency window function used to estimate P(kā„, kā„). It is possible to reduce the leakage by making the window function narrower; however, this comes at the expense of losing a larger fraction of the 21-cm signal. It is necessary to balance these competing effects to identify an optimal window function. Considering a broad class of cosine window functions, we identify a six term window function as optimal for 21-cm power spectrum estimation with OWFA. Considering only the k modes where the expected 21-cm power spectrum exceeds the predicted foregrounds by a factor of 100 or larger, a $5\, \sigma$ detection of the binned power spectrum is possible in the k-ranges $0.18 \le k \le 0.3 \, {\rm Mpc}^{-1}$ and $0.18 \le k \le 0.8 \, {\rm Mpc}^{-1}$ with 1000ā2000 and 104 h of observation, respectively.