i think roadkills may be a good proxy for mammal abundance if there are chances to collect such data. Of course, collection of such data will take time.

You might consider incorporating a citizen science aspect (like eBird.org) for some species - with enough participation you might be able to rise above the potential observation errors. Otherwise, road kills definitely could serve as a proxy. Depending on the species you are trying to estimate, you might also be able to use genetic mark-sight analyses from fecal surveys (certainly helpful for species composition....but you need a rigorous mark-resight design to get abundance). Bottom line, capitalize on the human resources and edges rather than seeing them as obstacles.  

Linear transect census can give you relatively good data for large mammals. Considering you will have time and money to do an appropriate effort and implement a minimal infrastructure (trails) in each forest fragment. Please, see some papers I have published recently on this method (Hilário et al. 2012, Folia Primatol 83(1):47-55; Mourthé 2013, Zoologia 30(1):119-124). There you can find several other references on linear transect method as well. Other possibility is working with occupancy instead of abundance (e.g. using the software Presence), which implicate that you will need to do repeated censuses in each trail considering the occurrence of the species instead of their abundances. This method can be more satisfactory if you guess it is not plausible invest large efforts in linear transects to get direct records (i.e. visualizations) of mammals. Yet, other clues such as vocalizations or presence of feces can indicate you that a mammal species occur in a fragment even if you cannot effectively see it, what is mandatory in the linear transect census. A final tip is camera trap! Do you can get abundances from camera trapping as well but I guess (not sure) that only in those cases you are pretty capable to identify individuals in the pictures (e.g. large felids). I hope I have helped you! Good luck!

Check out this examples of citizen science that you could adapt for your project:

http://www.inaturalist.org/guides/951

http://scistarter.com/project/38-Project%20Squirrel

Good work!

I agree that using citizen science effectively can help you get a very good idea of what's happening. You can even do a side project comparing the data you gained through citizen science, versus your more traditional data collection protocols (e.g., transects and camera traps). Keeping with the theme that road kill could be a very valuable data source, here is an initiative that appears to work well in South Africa, which allows people to either use their smartphones (more common) or websites (not as accessible to many people here):

http://endangeredwildlifetrust.wordpress.com/category/wildlife-and-transport-programme/

Adapting or using an existing app could really get people in the area involved. Good luck.

Depending on the forest, you can also try using a drone. It was used for census of different species, as well as other conservation activities in several projects. You can get some good ideas at http://conservationdrones.org/

For medium and bigger mammals (hares and larger) you could use camera traps: Rowcliffe et al. 2008. Small mammals: traps. Birds: Sound and visual observation. Line transects, pellet counts and anything else mentioned above!