Look for bicycle accident stats which include location details to see what percentage happen on sidewalks and the associated crosswalks. Then try to normalize for the relative number of trips which would be on sidewalks as opposed to the congruent road. The result will necessarily be an approximation.

The problem is that a researcher can't really do an experimental study "right" because it would violate ethics to have a group of people always riding on the sidewalk, and another on the roads in the same area and measure the accident rate, when you'd be deliberately requiring risky behavior. So any stats are observational and derivative.