The identification of jets containing bottom or charm hadrons is crucial to the LHC physics program.
Top-quark decays proceed almost exclusively through a $b$-quark. The Standard Model Higgs boson decays predominantly to $b\bar{b}$ pairs. Several scenarios of new physics result in an enhanced production of fermions of the third generation, such as models with an
extended Higgs sector or scalar top and bottom quark production in Supersymmetry. These physics scenarios correspond to very different environments for performing the identification of jets with heavy flavour hadrons of very different kinematics. This justifies a special effort in the identification of $b$- and $c$-jets with the ATLAS detector over a broad kinematical range.

Flavour tagging is based on the reconstructed trajectories of particle tracks and their extrapolation to the colliding beam envelope. The introduction of the ATLAS IBL pixel layer located at a radial distance of 3.3 cm from the interaction region with a 10 $\mu$m hit spatial resolution in Run 2 resulted in an improvement of the track extrapolation resolution by a factor up to 2 at low values of the track transverse momentum.
The adoption of a stochastic model of energy deposition in Si pixels and of a more realistic description of the material in the ATLAS inner detector system in the 2017 software configuration improved the data/MC agreement for the track extrapolation resolution and the response of track-based taggers.

Improvements and innovations in physics taggers, new approaches to multivariate analysis and training samples have brought optimised and more performant flavour-tagging algorithms for the analysis of the 2017 LHC collision data with ATLAS. This contribution summarises these recent developments.