Whenever the impact happens, that portion of audio has to be detected( if you want to automate it) or else you can identify those sections manually.
To automate it you need to identify characteristics of the audio during the impact, for ex: the frequency contents, or time spread. A good approach to combine these two characteristics is to either do a wavelet decomposition (morlet wavelet) or do a short time Fourier transform.
Once you have a method to identify these regions then you would need a reference (may be using an already available audio meter) to map these signals to audio (in DB), this would depend on how you quantize the analog audio, Sampling rate, FFT size etc. From this identified reference you can then map other audio impacts to suitable audio decibels.