| Summary: | Visual place recognition aims at making unmanned vehicles recognize a revisit place of their exact location and returning reasonable query information. Most researchers regard this kind of problem as an image retrieval task. There are mainly two categories in this task: the hand-crafted feature extraction method and the learning-based feature representation method. This dissertation will focus on the latter. In this dissertation, with the core of Vector of Locally Aggregated Descriptor (VLAD) part in NetVLAD, the features are represented as vectors. To use sequential information embedded in image series, a temporal convolution part is added to get a layer of sequential descriptors, which generate top K candi- dates for further similarity check with single descriptors of the same sequences of images. In analyzing phase, the dissertation compares different backbones of VGG-16 and AlexNet to select a satisfying CNN-based feature extractor. Also, a comparison of single and sequential descriptors is performed through Oxford, Nordland and Pittsburgh 250k that have different characteristics. The result shows that the sequential model with the hierarchical structure possesses greater performance facing changing lights and view angles, the best recall@20 is over 0.96. At last, several possible future works are listed like the efficiency of algorithms, the improved structure of models and variance-based robustness improvements. Master of Science (Computer Control and Automation)
|
|---|