Quantifying imperfect camera-trap detection probabilities: implications for density modelling