Your 2D points correspond to a ray in 3-space that passes through the 3D feature and the optical centre of the camera. You say you know the intrinsic parameters of the camera. So if you invert the camera's intrinsic calibration matrix, then you can place the image points on a 3D plane in the same Euclidean space as your point cloud. You then need to devise a way of finding the 3D point that's on (well, near actually) the ray emanating from the camera centre and passing through the image point on the image plane. Depending on the object in question this might be easy or quite difficult.
Hope that's useful in a vague sort of way. If you can be more specific, then we may be able to be of more help!
