Modern live cell fluorescence microscopy imaging systems, used abundantly for studying intra-cellular processes in vivo, generate vast amounts of noisy image data that cannot be processed efficiently and accurately by means of manual or current computerized techniques. We propose an improved tracking method, built within a Bayesian probabilistic framework, which better exploits temporal information and prior knowledge. Experiments on simulated and real fluorescence microscopy image data acquired for microtubule dynamics studies show that the technique is more robust to noise, photobleaching, and object interaction than common tracking methods and yields results that are in good agreement with expert cell biologists.