Fig. 1. Flowchart of the proposed algorithm

Fig. 2. Typical ResNet18 structure

Fig. 3. Designed ResNet structure

Fig. 4. Direct output form of designed network

Fig. 5. Typical images in different datasets. (a) City Centre; (b) New College

Fig. 6. GPS data. (a) City Centre; (b) New College

Fig. 7. Comparison of P-R curves of proposed algorithm, different layers in ResNet18, and other algorithms. (a) City Centre; (b) New College

Fig. 8. Form of output combination

Fig. 9. Influence of different sequence lengths on the results. (a) City Centre; (b) New College

Fig. 10. Influence of different m on the results when the sequence length is 15 in City Centre dataset

Fig. 11. Similarity curves before and after combination of 15 length sequences. (a) Similarity between sequence combination features and other frames; (b) similarity of sequence without combination