AI Inspection of Roads, Trees and Street Furniture: Where Computer Vision Helps

What computer vision already handles reliably

Three urban asset classes are today practical targets for automated visual inspection: the road and footpath surface, the street trees along the corridor, and the small infrastructure objects around them such as signs, bollards, lights and barriers. What unites them is that their most important health signals appear as detectable changes in appearance or geometry: a crack or pothole in pavement, a dead branch or diseased leaf in a canopy, a faded or damaged signpost. That makes them well suited to object detection and semantic segmentation models running over standard camera footage and LiDAR point clouds.

The economics have shifted because capture has become cheap and repeatable. Vehicles already travel every maintained street, and adding a high-resolution 360-degree camera, a LiDAR unit and a laptop is far less costly than sending surveyors to inspect each asset by hand. Detection frameworks from the YOLO family and segmentation networks that once required specialist hardware now run near real time on modest compute, and public training data such as the RDD2022 road-damage set lowers the barrier for a first model. The realistic position is not that algorithms replace inspectors, but that they turn sparse, subjective and slow surveys into dense, geolocated and repeatable observations that staff can verify.

Roads and footpaths: from subjective surveys to repeatable ratings

A concrete example of what is possible comes from Dún Laoghaire-Rathdown in Ireland, which in late 2025 became the first local authority there to trial a vehicle-based road and asset survey. High-definition 360-degree cameras mounted on vehicles capture imagery across an 823 km network, and AI analysis delivers consistent condition ratings, identifies worn road markings early and geo-locates street furniture such as signs and bollards. The council contrasts this with traditional Pavement Surface Condition Index surveys, which it describes as manual and subjective, and notes the system adds footpath condition ratings and inventory mapping its older methods could not provide.

The value proposition is managerial as much as technical. The imagery feeds lifecycle modelling and scheme identification, letting the council prioritise works that present the highest risk, target public money more precisely and plan footpath upgrades that support accessibility and active travel. The council also expects the data to reduce public liability risk by giving it a defensible, up-to-date picture of defects, and links the programme to a climate action target of cutting emissions 51% by 2030 through more targeted, sustainable interventions. In short, the survey's worth is measured in how it changes the maintenance queue and the budget conversation, not in detection accuracy alone.

Similar concepts appear at the project level. Engineering students at Novosibirsk State Technical University in Russia have built a computer-vision prototype that flags potholes and cracks, broken road signs, non-working traffic lights and worn pedestrian crossings directly on a video stream, trained on a labelled dataset of more than 6,000 images, with the intent to push geolocated detections onto a city map for municipal road services. Such efforts underline that the detection component is now the least exotic part of the pipeline; deployment, verification and integration with existing road-maintenance workflows are where the real work lies.

Street furniture: cameras as an asset inventory

The same drive-by imagery that rates pavement can also catalogue the small objects a city owns along its streets. Madrid's city digital twin programme runs an artificial-vision system over cameras installed on vehicles that travel public space; it automatically identifies and catalogues traffic signs, street furniture, road-infrastructure elements and other points of interest. That inventory is then combined with 2D and 3D cartography, high-resolution point clouds and, in some cases, live traffic feeds to support asset management, urban planning and citizen services.

The benefit of such an automated catalogue is that an asset register stops being a snapshot that is out of date almost as soon as it is printed. When every sign, barrier and bollard carries a geolocation and a condition assessment, a council can answer practical questions: which bollards are missing or damaged, which signs have faded below readability, where street furniture blocks a sightline, and which assets sit in corridors scheduled for resurfacing. Enriching the vision layer with real-time traffic information adds a use dimension, so a fixed asset can be understood in the context of how the street is actually used.

For maintenance teams the practical effect is a shift from reactive complaints to planned work. Instead of waiting for a resident to report a damaged sign or barrier, the city can produce a ranked list of objects by age, location and condition and fold it into the same GIS and work-order system that handles pavement defects. The main caution is that furniture detection depends on the same image quality, angle and occlusion constraints as any vision task: objects hidden behind parked cars or foliage will be missed, so the automated inventory should be treated as an initial pass that periodic recapture keeps current.

Urban trees: structure, health and risk from LiDAR and imagery

Trees sit at the intersection of civil safety and green-asset value, and they are where a fusion of LiDAR and ordinary imagery pays off. Busan in South Korea began a pilot in 2025 covering 8,000 street trees along a central corridor, combining vehicle-mounted LiDAR, high-resolution cameras and AI analysis to automatically capture tree location, species, height, canopy width and the presence of dead branches across roughly 25,000 square metres. The stated ambition goes beyond surveying: the city wants a pre-emptive system that analyses pest outbreaks, stem decay and fall risk in advance and links to lifecycle management, including prompts for watering, fertilising, pest control and records of pruning and replanting.

Camera-only methods add a leaf-level health signal. Researchers at Waseda and Ryukoku universities developed a system, described in the journal Measurement, that uses YOLOv8 and DeepSORT to detect and track individual leaves across video frames, keeping only the clearest views, and then a DeepLabV3Plus segmentation network to quantify damaged areas. It can automatically identify diseased regions on individual leaves caused by bacteria, pests and fungi, and was validated on footage of urban plants in Tokyo. Because the cameras can be mounted on drones or on ordinary city maintenance vehicles such as refuse trucks, routine services double as data collection without damaging the trees by sampling.

The realistic division of labour is by scale. Aerial LiDAR excels at broad coverage and at extracting height and canopy metrics across a whole district, which feeds carbon, pollen and canopy indicators. But research from Bologna, comparing two individual-tree segmentation algorithms on airborne LiDAR, notes that aerial data can struggle to detect understory vegetation and precisely capture crown morphology in dense areas, and that species identification and detailed health status need terrestrial or mobile LiDAR and spectral sensing closer to the trees. Mature programmes therefore combine a wide but shallow aerial pass with a narrow, detailed ground pass.

Accuracy is per class, not per system

The strongest lesson from operational pilots is that accuracy must be reported per object class, because it varies enormously. In San José, California, a pilot funded and evaluated by the US Department of Transportation ran from December 2023 to July 2024, mounting cameras on city vehicles and testing four vendors' AI against footage that city staff reviewed by hand. Pothole detection achieved a 97% true-positive rate, and residents strongly supported using AI to speed pothole repairs. Trash and debris detection reached 88% accuracy, though staff raised concerns about misidentifying encampments.

The failures are just as instructive. Recreational-vehicle detection was only 70% accurate, lived-in-vehicle detection dropped to 12.5%, and the model could not reliably track vehicles that moved slightly, making 72-hour parking-violation detection unusable. Graffiti and encampment detection were removed from the pilot entirely due to capacity limits and privacy concerns, and all personally identifiable information in the imagery was blurred during collection. The takeaway is that appearance-based tasks such as finding a pothole generalize well, while tasks that require legal classification or tracking motion over time remain firmly human territory.

A practical consequence is that a council should set expectations per defect type when it buys or builds a system. Asking a vendor for an overall accuracy number is less informative than asking for a confusion matrix per class on local footage, ideally on the same streets and in the same lighting and weather the system will actually run in. Where the stakes are high, staff should review detections against the original imagery before any enforcement action or high-value work order, mirroring the verification process San José built into its pilot.

Ground truth, data hygiene and the human in the loop

Automated inspection is only as trustworthy as the reference data used to validate it, and municipal records are frequently the weak link. In Bologna, researchers comparing two tree-segmentation approaches against the city's public tree catalogue found that roughly half the trees detected in LiDAR data were absent from the catalogue, that several catalogued positions corresponded to places with no visible tree, and that most height records were about two decades old. They concluded the register could not serve as ground truth for training or evaluating tree-extraction models, and recommended regular aerial acquisitions to keep it current.

Similar data problems affect roads. The developers of one multi-model road-damage pipeline trained on the RDD2022 dataset highlight persistent challenges: a lack of appropriate labelled data, large data volumes, unfavourable photographic angles, poor resolution and differing surface types. Each of these degrades the transfer of a model from the laboratory dataset to a specific city's asphalt and climate. Because conditions differ by region and season, models that work in one place cannot be assumed to work elsewhere without local retesting and, ideally, fine-tuning on locally collected imagery.

This is why responsible deployment keeps a person in the loop and defines the refresh cadence up front. Seasonal leaf-on imagery is needed to assess tree crowns, a new capture pass is required after resurfacing or major storms, and privacy screening must blur faces, licence plates and other identifying information before footage is stored or shared. The pattern that emerges across roads, trees and furniture is consistent: cameras and models excel at finding candidate problems and measuring them repeatably, but the decision to act, and the legal and safety judgement behind it, stays with trained people who confirm each finding before it becomes a work order.