The extreme variability in the appearance of a place across the four seasons of the year is one of the most challenging problems in life-long visual topological localization for mobile robotic systems and intelligent vehicles. Traditional solutions to this problem are based on the description of images using hand-crafted features, which have been shown to offer moderate invariance against seasonal changes. In this paper, we present a new proposal focused on automatically learned descriptors, which are processed by means of a technique recently popularized in the computer vision community: Convolutional Neural Networks (CNNs). The novelty of our approach relies on fusing the image information from multiple convolutional layers at several levels and granularities. In addition, we compress the redundant data of CNN features into a tractable number of bits for efficient and robust place recognition. The final descriptor is reduced by applying simple compression and binarization techniques for fast matching using the Hamming distance. An exhaustive experimental evaluation confirms the improved performance of our proposal (CNN-VTL) with respect to state-of-the-art methods over varied long-term datasets recorded across seasons.