This thesis presents the search for narrow resonances decaying into a pair of vector bosons (W W , W Z or ZZ) in proton-proton collision data recorded with the ATLAS detector, as well as the development of identification algorithms for this search. Di-boson systems are a prime probe for new physics: the Standard Model has precise expectation values for the amplitudes of di-boson interactions and any observed deviation, either in the form of a resonance (predicted by many BSM theories) or in a tail excess, would indicate the presence of new physics. This thesis focuses on the fully hadronic decay mode of the di-boson system: an excess is searched for in the di-boson mass distribution of events in which the two vector bosons decay into two quarks each. Each vector boson is reconstructed as a single large-radius jet that contains the two quarks produced in the decay of the parent vector boson, and their hadronisation products. To increase the sensitivity of the search, jets are built from particle-flow inputs, such as Track- CaloClusters (TCCs) and United Flow Objects (UFOs), that combine information from both the tracker and calorimeter portions of the ATLAS detector. Both of these procedures are introduced and motivated. In the full search performed over data collected between 2015 and 2018 the use of TCCs resulted in a significant improvement in jet substructure performance at high pT and allowed for the development and optimisation of a new boson identification (tagging) algorithm. The combined use of TCC inputs and the new tagger led to a two times improvement in our ability to identify vector bosons and reject QCD background with respect to the reference configuration. The search didn't reveal any deviation from the background expectations, and 95% confidence level exclusion limits were set for the existence of new particles predicted by three BSM theories. The development of UFOs in 2018, resulting in a significant improvement in jet substructure per- formance over the entire pT spectrum, justified the exploration of new tagging techniques involving machine learning. The use of a combined approach that exploits both the traditional jet substructure variables and the raw jet constituent information was developed and resulted in up to 4 times improvement in our ability to identify vector bosons and reject QCD background with respect to the TCC jet tagger used in the full search. In addition, this novel approach allows for the identification of the nature of vector bosons (W or Z) while also letting us explore the possibility of identifying their polarisation (longitudinal and transverse).