In the study's most direct planner comparison, a method that drew a sidewalk centreline directly on the camera image had much lower reported centering error and runtime than the best bird's-eye-view (BEV) method. Across 32 hand-annotated frames, its lateral error was 14.3 pixels and its runtime was 2.2 milliseconds, compared with 65.0 pixels and 926.8 milliseconds for BEV. Yet the two paths were nearly equally aligned with the annotated sidewalk area, at 98.5% and 98.6%.
That result comes from an offline methods evaluation, not a demonstration of closed-loop scooter navigation. The pipeline used video from a forward-facing monocular camera mounted on an electric scooter, and the evaluation did not close the loop on physical scooter control or explicitly model dynamic obstacles.
The comparison stayed close to the camera view
The system turned each frame into a binary sidewalk mask and compared five ways to draw a centreline trajectory through it: three planners in BEV space and two in image space. Its training used pseudo-labels from a high-capacity OneFormer Swin-L teacher and 1,419 image-mask pairs, made up of 699 earlier-sequence frames and 720 newly extracted frames from four additional campus videos.
The main benchmark contained 32 hand-annotated frames. A broader replay covered six campus sequences and 22,679 frames, allowing the authors to examine segmentation quality and frame-to-frame stability beyond the smaller annotated set.
The segmentation gain was clearer on labeled frames
On the hand-annotated benchmark, the candidate mask overlapped the reference mask more closely than the baseline: intersection over union (IoU), a measure of overlap, was 0.946 versus 0.758. Other reported mask scores—precision, recall and F1—were also higher for the candidate, at 0.983, 0.962 and 0.972, compared with 0.910, 0.835 and 0.851. Reported latency was 11.7 milliseconds versus 18.9 milliseconds.
In the full replay, the gap was smaller but still favored the candidate on several measures. Mean segmentation IoU was 0.925 versus 0.909; the unstable rate was 0.33% versus 1.46%; template success was 79.3% versus 73.7%; and fallback was used on 14.3% of frames versus 19.0%. Both systems reported a path on 100% of replay frames.
The BEV result came with a major reliability caveat
In a separate 4,407-frame BEV reliability profile, a valid path appeared on 10 frames, or 0.2%, while a held path appeared on 20 frames, or 0.5%. The remaining 4,377 frames—99.3%—had no path, and mean BEV mask occupancy was 0.0002.
That 99.3% figure came from one profiled run, so it should not be treated as a general BEV failure rate.
The runtime comparison pointed in the same direction. At 640 by 360 pixels on a CPU-only setup, the image-space stack totaled 16.9 milliseconds per frame, equivalent to 59.2 frames per second, versus 416.0 milliseconds and 2.4 frames per second for the BEV skeleton stack. Runtime depends on hardware and measurement conventions, so the figures are specific to this setup.
When midpoint extraction failed, the designated fallback was image-space distance-transform planning, which had the highest inside-GT ratio—the share of path pixels inside the reference mask—at 99.4%. It took 108.1 milliseconds and was reported as 8.6 times faster than BEV distance-transform planning.
Useful evidence, but still a narrow test
An oracle-mask diagnostic, using ground-truth masks rather than predicted ones, suggested that planner choice remained important. Image-space midpoint had a 15.2-pixel center error and 98.4% inside-GT ratio, compared with 69.2 pixels and 98.0% for BEV distance-transform planning; image-space distance-transform planning recorded 62.5 pixels and 98.9%.
The supplied analysis reports no confidence intervals, p-values or formal hypothesis tests. The differences therefore describe performance in these recorded subsets rather than establishing that one approach will consistently outperform another.
The evaluation fixed the BEV homography to one camera pose, used campus video from the tested camera configuration, and compared planners on only 32 hand-annotated frames. It did not include closed-loop scooter control or explicit dynamic-obstacle modeling, leaving real-world navigation performance unestablished.
Paper data and sources
Original title: Lightweight Machine Learning-Driven Monocular Sidewalk Path Extraction for Embedded Micromobility Navigation
Authors: Lkhanaajav Mijiddorj, Yang Yan, Tyler Beringer et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text