Abstract: Considering the similarity characteristic of indoor environmental structures, a mobile robot global localization algorithm based on topological relationship between image sequences is proposed. Firstly, Gist descriptors are extracted from images, and a local extremum algorithm is applied which divides the environment into several groups of different image sequences. Then, each group image sequence is modeled by the echo state network (ESN), which is trained on bidirectional time series, to extract robust features of image sequences, and the bidirectional matching strategy in space domain is applied to matching the sequence features. Finally, the topological relationship between image sequences is modeled by hidden Markov model (HMM), and the problem of global localization of the mobile robot is transformed into the process of finding the longest path in a directed acyclic graph. By comparing with single-frame-based matching algorithms, SeqSLAM (sequence simultaneous localization and mapping) and Fast-SeqSLAM, the proposed algorithm can achieve 100% localization in both indoor corridor and office environments, and the robot can accurately locate itself by only moving 0.80m especially in indoor office environments. The experiment results show that the proposed algorithm has more strong robustness, higher localization accuracy and efficiency.