Dynamic Time Warping of Deep Features for Place Recognition in Visually Varying Conditions

This paper presents a new visual place recognition (VPR) method based on dynamic time warping (DTW) and deep convolutional neural network. The proposal considers visual place recognition in environments that exhibit changes in several visual conditions like appearance and viewpoint changes. The proposed VPR method belongs to the sequence matching category, i.e., it utilizes the sequence-to-sequence image matching to recognize the best matching to the current test image. This approach extracts the image’s features from a deep CNN, where different layers of a two selected CNNs are investigated and the best performing layer along with the DTW is identified. Also, the performance of the deep features is compared to the one of classical features (handcrafted features like SIFT, HOG and LDB). Our experiments also compare the performance with other state-of-the-art visual place recognition algorithms, Holistic , Only look once , NetVLAD and SeqSLAM in particular.