We estimate on a compact interval densities with isolated irregularities, such as discontinuities or discontinuities in some derivatives. From independent and identically distributed observations we construct a kernel estimator with non-constant bandwidth, in particular in the vicinity of irregularities. It attains faster rates, for the risk $L_{p}, p\geq 1$, than usual estimators with a fixed global bandwidth. Optimality of the rate found is established by a lower bound result. We then propose an adaptive method inspired by Lepski's method for automatically selecting the variable bandwidth, without any knowledge of the regularity of the density nor of the points where the regularity breaks down. The procedure is illustrated numerically on examples.