All sounds sensible to me but depending on what your overall aim is you might want to split long segments into shorter ones of equal length, as links in road network can be anywhere from 2 metres to 10 miles long so your spatial resolution will vary enormously (unless it is the link itself you are interested in, which for a pollution problem seems unlikely). I think there is an Arc tool for this though can't find it right now.
As to mean or total, that depends on your end use again. Do you have enough traces to reliably estimate traffic levels per link? Then you could sum. If not then take mean though bear in mind this will introduce heteroscedasticity.
As to whether 'near' is reliable, the only pitfall I see is erroneous GPS points that either (1) don't fall near any link, (2) fall close to an intersection so it's not possible to tell which link they are on, or (3) are recorded in built up areas with poor signal and appear to fall on a neighbouring link instead. (1) and (2) can easily be discarded if you can afford to lose the data. (3) is more insidious and can't be solved without far more advanced data cleaning, but you may just decide it's not a problem.