Visual Place Recognition by DTW-based sequence alignment

Abstract

Place recognition, also called visual localization, facilitates the autonomous navigation capabilities of the future of driverless cars. This paper proposes a new place recognition algorithm that considers the appearancebased methodology to localize the vehicle by utilizing visual route map, i.e. a sequence of images, or sets of features extracted from these images, that were recorded over different times and dates for the route environments. These reference sequences are accurately labeled and annotated using GPS tags or manually using odometry information. The dynamic time warping (DTW) algorithm is used to achieve image sequence alignment and find the best match for each frame from the test sequence. The proposed algorithm considered hand-crafted features like SIFT, HOG, and LDB. Experiments, using common challenging and benchmark datasets, i.e. “UQ St Lucia” and “Nordland”, have been conducted, and it has been observed that the proposed technique has significantly improved the performance of well-known appearancebased descriptors SIFT, HOG, and LDB as compared to its individual performance and to some of the state-of-the-art localization and mapping methods such as ABLE (Binary-appearance Loop-closure).