The short answer is: use your optical flow algorithm to get the best-fit warping of some number of previous images into your current image. Then assume that your object is moving on a line through the images, and do some kind of robust best-fit to that line. When the object reappears it is more likely to be near the predicted projection of the line than not.
You didn't give enough specifics in your question to really say more than that.
I assume by "LK" you mean "the Lucas-Kanade optical flow algorithm?" But what are you really doing? Are you taking image pyramids of each image and then doing multiple iterations of warping at each level of the pyramid? Are you really doing the simplest form of Lucas-Kanade or are you actually doing one of the varieties of Kanade-Lucas-Tomasi where you are concentrating on "corners"?
What is your description of an "object"? Is it really just a "feature" (a corner point) or a small set of features?
You may need to use a more accurate and robust optical flow algorithm. For example, some variation that does dense estimation (rather than the sparse estimation done by Lucas-Kanade) and that uses more robust distance metrics than least squares.
For example:
Secrets of optical flow estimation and their principles
Sun, D., Roth, S., and Black, M. J.,
IEEE Conf. on Computer Vision and Pattern Recog., CVPR, June 2010.