Image-Sequence-Based Mobile Robot Localization

Considering the structural similarity of environments, an image-sequence-based mobile robot localization algorithm is proposed, which utilizes image sequences to represent different places in environments. To achieve localization of the mobile robot, the mobile robot is allowed to explore the environments in advance. The RGB image data stream acquired by the mobile robot is partitioned into different image sequences according to the moving distance of the mobile robot. Then, each of image sequences is sent into an LSTM (Long Short-Term Memory)network to learn the characteristics of each image sequence. The final cell states of the LSTM will be treated as the features of the image sequence. When the mobile robot travels in the environments once again, the newly acquired image sequence will be sent into this well-trained LSTM network to extract the features of the image sequence. By matching features among different image sequences, the mobile robot can recognize which place it may locate in. Finally, experimental comparisons with single-frame-based localization algorithms are conducted to verify the performance of the proposed algorithm from the localization accuracy.