Hybrid generative-discriminative learning for online tracking of sperm cell

Sperm motility is an essential metric for evaluation of human semen quality. Computer-assisted sperm analysis (CASA) aims at objective assessment of sperm motility by trajectory construction and an-alysis, where online single sperm tracking is a crucial step for the intracytoplasmic sperm injection (ICSI). Most existing sperm tracking approaches are error-prone to handle the large uncertainty of sperm motion and the background distracters in optical microscopy, leading to inaccurate inference of the sperm moti- lity in CASA. To this end, we propose in this work a hybrid generative-discriminative tracker (HGDT) for online single sperm tracking. HGDT retains most of the desirable properties (explicit appearance mo-deling, for example) of generative trackers, while improving the classification performance to separate the visual sperm from the background distracters. To deal with the large motion uncertainty of the sperm cell, an energy-biased stochastic approximation Monte Carlo (EB-SAMC) sampling algorithm is proposed for more effective Bayesian tracking. Experiments results on several clinical videos demonstrate the efficacy of our method, as well as the superiority to several state-of-the-art methods in terms of tracking accuracy and computational efficiency.