Sequence-based place recognition under appearance change and with low-resolution imagery.
SeqSLAM recognises previously visited places by matching image sequences. Introduced at ICRA 2012, it combines low-resolution image comparison, local contrast enhancement and temporal consistency to tolerate substantial appearance change. The RSS 2012 and IJRR 2013 studies examined recognition with severely reduced visual information. [S1][S3][S7]
Appearance-change recognition and the visual information required for route recognition.
SeqSLAM: Visual route-based navigation for sunny summer days and stormy winter nights
The original SeqSLAM paper introduced sequence-based route recognition under substantial changes in illumination and weather. Local comparison of image differences and temporal matching enabled recognition when individual images were ambiguous. [S1]
ICRA 2012 paper ↗Visual Route Recognition with a Handful of Bits
Vision-based place recognition: how low can you go?
The RSS study varied image resolution, bit depth, field of view and sequence length. Experiments included four-pixel images and two light sensors. The IJRR paper expanded the investigation to seven datasets, including compression, motion blur and failure cases. [S3][S7]
Two single-pixel light sensors provided intensity sequences for indoor localisation. The figure shows the sensor apparatus and route across three building levels. [S7]
Image comparisons from the original SeqSLAM research.
The Alderley dataset contains repeated traversals under daytime and rainy night-time conditions. It evaluates recognition under changes in illumination, visibility and surface reflections. [S2]
Compare single-image and sequence matching under noise and repeated scenery.
This synthetic difference matrix contains a route, noisy observations and occasional misleading matches. Change the sequence length and inspect the computed match. The white dashed line is known ground truth; the gold path is the selected candidate. Synthetic teaching example; the archival images above show the real visual challenge.
The demo searches four straight-line speed hypotheses. Noise is deterministic so settings can be compared. It illustrates one part of the idea; it is not the complete SeqSLAM pipeline.
Image reduction, local normalisation, pairwise comparison and sequence search.
A MATLAB toolbox for inspecting preprocessing, matching costs and sequence estimates.
From low-resolution route recognition to adaptive, learned and application-specific extensions.
16 milestones shown
The same suburban route was recorded in daylight and during a rainy night. The dataset provides image frames and manually established correspondences for evaluating recognition across changes in illumination and weather. [S2]
Experiments varied image resolution, field of view, bit depth and sequence length. Reported results included recognition with four-pixel images and indoor localisation using two single-pixel light sensors. [S3]
The journal study examined seven datasets, varying resolution, pixel depth, field of view, compression, motion blur and sequence length. It also documented failures caused by viewpoint variation and sequence aliasing. [S7]
The ICRA 2013 study evaluated localisation with low-cost cameras in dark environments. Experiments included exposures up to 10 seconds from a moving vehicle, substantial motion blur and matching between daytime and night-time imagery. [S8]
A directed graph represents image sequences and intersections. A particle filter searches possible paths and image scalings, supporting branching routes and lateral viewpoint changes. [S15]
Intra-image SeqSLAM matches sequences of patches within images. The ACRA 2016 study evaluated multimodal rock-image registration and laboratory robot-arm positioning under changing illumination, motivated by Mars 2020 targeting requirements. [S30]
The temporal window is selected from the statistical significance of candidate localisation hypotheses. Tests with visual and Wi-Fi observations, including reordered traversals, reduced the maximum interval without correct localisation relative to fixed windows while retaining competitive accuracy. [S29]
This preprint develops joint visual and positional learning with a CNN and LSTM, aiming to replace an all-pairs matching matrix and handcrafted temporal filtering with learned inference. [S24]
LENS integrates an event sensor, spiking neural networks and neuromorphic processing for on-device localisation. It represents subsequent work on resource-efficient visual recognition rather than a version of SeqSLAM. [S26]
No milestones match. Try another term or select “All milestones”.
Constraints on visual distinctiveness, viewpoint and motion.
Large viewpoint changes can reduce whole-image similarity. Subsequent methods address this through image scaling, predicted depth or regional matching. [S15][S13][S22]
Longer sequences may resolve ambiguity but increase latency and sensitivity to speed changes, junctions and sequence aliasing. [S7][S3]
Performance depends on viewpoint, route ordering, frame spacing, matching tolerance and algorithm parameters. OpenSeqSLAM2.0 supports systematic parameter evaluation. [S9][S11][S19]
The original research and practical starting points.
The ICRA 2012 method and experiments. [S1]
Read the paper ↗Explore preprocessing, sequence matching and parameter effects in the MATLAB toolbox. [S19][S21]
Get the toolbox ↗Revisit the original day/night challenge. [S2]
Explore the dataset ↗The independent MATLAB implementation associated with the four-season Nordland evaluation. [S9][S10]
Explore the implementation ↗Sources for the methods, experiments and research connections.
Biologically inspired simultaneous localisation and mapping.
Explore RatSLAM →