Enumeration of specific types of circulating cells is an important problem in many areas of preclinical biomedical research. This is normally achieved by drawing and analyzing blood samples, but this is limited by the small sampling volume so that very rare cells are difficult to detect. In this work, we discuss a computer vision approach to “in vivo flow cytometry”, where fluorescently labeled cells are detected and enumerated non-invasively in the bloodstream in live animals. We first developed an instrument to acquire high sensitivity, fast video sequences in wide-field (5 x 5 mm) fluorescence trans-illumination mode to allow sensing of cells over a large area and therefore with significantly improved sensitivity versus current approaches. Our goal was to allow detection of less than 104 total cells in circulation. The major challenge with our method was distinguishing moving cells from the substantial noise content of the image sequences obtained with our electron multiplied CCD camera; this was due to weak fluorescent signatures, high illumination laser power, applied imager gain, and biological tissue autofluorescence. We developed and validated a novel two-step computer vision algorithm to remove interfering background and analyze the dynamics of moving cells. We demonstrated the feasibility of this approach first in optical flow phantoms and then in mice. Our approach resulted in a positive predictive value of 90% and sensitivity of better than 85% at very low concentrations compared to the gold standard. This represents at least an order of magnitude improvement in detection sensitivity over current techniques.