Loop　closure　detection　(LCD)　is　an　important　technique　for　maintaining　global　consistency　in　mobile　robot　simultaneous　localization　and　mapping　(SLAM)　systems.　However,　the　LCD　problem　remains　a　highly　challenging　under　the　interference　of　changes　in　perspective　and　appearance.　To　address　this　problem,　a　fast　online　LCD　algorithm　based　on　lightweight　convolutional　neural　network　MobileNetV3　and　local　feature　aggregation　layer　NetVLAD　is　proposed.　First,　the　cropped　MobileNetV3　is　used　for　image　feature　encoding.　Second,　the　obtained　feature　is　passed　into　the　NetVLAD　Layer　and　output　the　global　features　of　the　original　image.　Then,　an　approximate　nearest　neighbor　method　based　hierarchical　navigable　small　world　(HNSW)　is　adopted　to　reduce　feature　matching　time.　Evaluation　results　on　multiple　publicly　available　datasets　confirm　that　the　proposed　method　significantly　enhances　computational　efficiency　while　maintaining　high　accuracy.　©　2024　ACM.