Impact is a common source of in-service damage that compromises the safety and performance of engineering structures. Impact monitoring has been extensively studied by several researchers and it has been shown that damage extent can be correlated with the impact magnitude. For this reason, online impact detection systems are essential and require automatic and intelligent techniques providing a probabilistic interpretation of their diagnostics. On that account, the present study proposes an automatic approach based on Bayesian regression with Gaussian processes for impact magnitude estimation and localization. Firstly, structural dynamic responses captured by PZT transducers due to impact events are recorded from simple impact experiments. In a second step, the discrete wavelet transform and nonlinear principal component analysis are evaluated for signal filtration, feature extraction and data compression of the time histories respectively. In order to improve the impact localization, a chirplet atomic decomposition is developed and used in conjunction with a time-frequency energy distribution in order to accurately separate wave packets and extract frequency information. Additionally, the signals power spectral density and time difference of arrivals are estimated and used in conjunction with the extracted features to provide the training inputs to the Gaussian Processes. At the end, the effectiveness of proposed methodology is demonstrated experimentally.