The problem the researchers tackled is stark. Heavy goods vehicles, buses, and delivery trucks have mass, extended body geometry, wide turning radii, and sweeping paths that make them disproportionately lethal to vulnerable road users. Naturalistic driving studies have repeatedly identified right-turn scenarios as among the most severe conflict configurations, because a cyclist or scooter rider can travel straight alongside a truck while remaining entirely invisible to the driver. Existing countermeasures—mirrors, cameras, aftermarket sensors—leave gaps. Truck drivers sometimes fail to check blind-spot mirrors at the critical moment, and sensor layouts still miss zones near the front and nearside corners of large vehicles.

What makes this new framework distinctive is that it operates on two complementary levels simultaneously. At the macro scale, a GIS-based spatial decision-support model translates eighteen evaluation criteria into a deployment-priority map. At the micro scale, six YOLO-based object detection models were trained to recognize the specific objects—e-scooters, bicycles, e-bikes, and riders—that a warning system would need to detect in real time. The two layers are not interchangeable: the spatial model answers where a warning system matters most, while the detection models answer whether the technology can actually work.

The eighteen criteria were organized into three categories: micromobility user exposure potential, heavy-vehicle operating exposure, and blind-spot conflict and safety vulnerability. Five experts—two transportation academics, a computer engineer, an urban planner, and a heavy-vehicle driver—evaluated each criterion using linguistic terms converted into triangular fuzzy numbers. Two weighting methods, Fuzzy Simple Weight Calculation and the Fuzzy Logarithm Methodology of Additive Weights, were applied and then averaged into a hybrid weighting scheme. This dual-method approach reduced the risk that a single weighting philosophy would skew the results.

When the weights were aggregated, one finding stood out: blind-spot conflict and safety vulnerability received the largest total weight of any criterion group, at 0.40449. Within that group, the single most influential sub-criterion was high-activity blind-spot overlap potential—the places where micromobility exposure, heavy-vehicle activity, and conflict-prone geometry all coincide. Turning-conflict susceptibility and intersection-based blind-spot conflict density ranked second and third. This pattern confirms that the model was not simply mapping where scooters are popular; it was identifying where scooters and trucks are most likely to meet dangerously.

The researchers applied their framework to Erzurum, a high-altitude city in eastern Türkiye with a large student population and emerging micromobility demand. Using ArcMap, they converted point, line, polygon, and raster datasets into normalized spatial layers through Euclidean distance, kernel density, point density, and slope analysis, then combined them through weighted overlay. The resulting priority map identified central and inner urban zones as the most critical candidate areas for warning-system deployment. Notably, the two fuzzy weighting methods produced nearly identical spatial patterns, suggesting the prioritization is robust rather than an artifact of methodological choice.

On the detection side, the team compiled a dataset of 6,070 images covering five micromobility-related classes and trained six YOLO models under identical configurations. The results were striking: mean average precision at the 0.50 intersection-over-union threshold ranged from 0.9438 to 0.9694 across all models, while the stricter mAP@0.50:0.95 metric ranged from 0.8739 to 0.9079. YOLO26s achieved the highest mAP@0.50 at 0.9694, while YOLO26m produced the best mAP@0.50:0.95 at 0.9079. These figures indicate that even lightweight models can reliably identify the objects a blind-spot warning system would need to detect.

Not every class performed equally well. The person-on-scooter category—the one that matters most, because it represents an active vulnerable road user rather than a parked device—proved the most challenging to detect. This compound visual category requires the model to recognize both rider and scooter as a single functional object, and variations in posture, occlusion, viewing angle, and clothing reduce separability. The researchers note that future datasets should prioritize more diverse person-on-scooter samples under different lighting, weather, and urban backgrounds before operational deployment.

The framework’s authors are candid about its current limitations. The YOLO component was not tested in a fully embedded vehicle-mounted system, and the image-level data split means the reported metrics reflect comparative benchmarking rather than independent spatial or temporal generalization. The GIS-based priority map is also static: in a city with severe winters like Erzurum, snow cover and icy surfaces suppress micromobility use for part of the year, so the map represents structural deployment priority rather than constant year-round risk. Future versions could incorporate real-time detection counts, seasonal demand layers, and dynamic temporal weighting to produce time-dependent priority surfaces.

What the study ultimately demonstrates is a conceptual bridge that has been missing from urban safety planning. Detection studies typically report model accuracy without asking where a system should go; spatial planning studies identify priority zones without verifying that the underlying technology can perceive the relevant road users. By connecting object-level perception with location-level intervention strategy, this GeoAI framework offers cities a reproducible method for deciding where intelligent blind-spot warning technologies would deliver the greatest safety benefit first—a question that becomes more urgent as e-scooters, e-bikes, and heavy vehicles continue to share increasingly crowded streets.

Subject of Research: An integrated GeoAI framework combining YOLO-based micromobility object detection with GIS–MCDM spatial prioritization to guide deployment of heavy-vehicle blind-spot warning systems in urban areas.

Article Title: A GeoAI framework integrating YOLO-based micro mobility detection and GIS–MCDM for heavy-vehicle blind-spot warning deployment

Article References: Kaya, Ö., & Çodur, M. Y. (2026). A GeoAI framework integrating YOLO-based micro mobility detection and GIS–MCDM for heavy-vehicle blind-spot warning deployment. Frontiers in Earth Science, 14, Article 1936045. https://doi.org/10.3389/feart.2026.1936045

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1936045

Keywords: GeoAI, YOLO, micromobility, heavy vehicles, blind spot, GIS, MCDM, road safety, vulnerable road users, e-scooters, object detection, urban planning