Visual loop closure detection in robotics is defined as the ability of recognizing previously seen places given the current image captured by the robot. The Bag-of-Words image representation has been widely used for these kinds of tasks. However, in this paper, an appearance-based approach for loop closure detection using local invariant features is proposed. Images are described using SIFT features and, for avoiding image-to-image comparisons, a set of randomized KD-trees are employed for feature matching. Further, a discrete Bayes filter is used for predicting loop closure candidates, whose likelihood is based on these KD-trees. The approach has been validated using monocular image sequences from different environments.