AI-Assisted RSA 21 Inspection Drives in Practice

AI-assisted inspection drives under Germany’s RSA 21 framework combine professional work-zone inspections with GPS, mobile documentation, and automated analysis of images or operational data. AI can prioritize anomalies, identify recurring defects, and support dispatching. Responsibility for inspection, assessment, and corrective action nevertheless remains with qualified personnel and the roles established by the traffic order and contractual requirements.

Why are RSA 21 inspection drives becoming a digital operations process?

A work zone does not remain in the condition in which the traffic safety crew originally installed it. Traffic, construction activity, weather, deliveries, pedestrians, cyclists, third parties, and the progress of the project continuously affect temporary traffic control.

A delineator may be moved. A warning light may lose power. A temporary sign can rotate. A temporary marking can become difficult to recognize. Construction vehicles may change access conditions, and a traffic signal may develop a fault between two scheduled visits.

For a German traffic safety contractor managing many active sites, the operational question is therefore not simply whether an employee drove past a site. The company needs to know which work zone was inspected, at what time, which parts were actually checked, what the employee observed, whether a defect was found, what corrective action followed, and whether another inspection became necessary.

Traditional paper records can document many of these facts, but they create additional work. Photos often remain on a phone, handwritten notes need to be transferred, inspection times must be reviewed manually, and unresolved defects have to be communicated to dispatchers through calls, messaging apps, or separate task lists.

A digital inspection process changes this architecture. The important step is not replacing a sheet of paper with a screen. It is connecting the work order, traffic control plan, inspection route, GPS evidence, photos, defects, corrective measures, and follow-up inspection as related operational records.

That distinction becomes even more important when artificial intelligence is added. AI is most useful when it receives structured operational context rather than thousands of unrelated images and text notes.

Which rules actually govern RSA 21 inspection drives in Germany?

The term “RSA 21 inspection drive” is widely understandable in the German work-zone industry, but the regulatory structure is more nuanced.

RSA 21 stands for the German Richtlinien für die verkehrsrechtliche Sicherung von Arbeitsstellen an Straßen, the guidelines governing traffic-related safety measures at road work sites. They replaced RSA 95 and were formally announced by the German Federal Ministry of Transport through General Circular Road Construction No. 24/2021.

Section 45(6) of Germany’s Road Traffic Regulations, the StVO, requires contractors to obtain the relevant authority’s traffic order before beginning work that affects road traffic and to comply with that order.

Inspection and maintenance requirements also involve ZTV-SA, the additional technical contract conditions and guidelines for safety work at road work sites. The FGSV currently continues to list ZTV-SA 97 as a complementary document to RSA 21. Certain provisions dealing with portable traffic signals have been replaced by the 2023 rules for portable traffic signals, while a broader new edition of ZTV-SA remains under development.

This matters when software is designed.

For longer-duration work zones, German administrative practice can require inspections at least twice per day, including around daybreak and after darkness begins. On non-working days, at least one inspection per day may apply, with additional immediate inspections after severe weather or storms. The Hessian government service portal explicitly states these requirements with reference to Section 7 of ZTV-SA 97. The specific traffic order, contract, location, client requirements, and risk situation remain relevant.

A professional software platform therefore should not hard-code a universal schedule called “RSA 21 frequency.” Inspection rules should be configurable for each project and order.

Where can AI provide useful support during an inspection drive?

The strongest use case is not an AI system that announces that a work zone is legally compliant.

A much more practical model is an AI assistant that identifies changes, structures evidence, and tells a qualified employee where human attention may be needed.

Consider a driver photographing a line of temporary delineators. Computer vision could detect traffic signs, barriers, warning lights, or other objects and compare the current image with the previous inspection. If one delineator appears to have shifted significantly, the system could flag the location.

The message should not read, “RSA 21 violation detected.”

A more appropriate message would be: “Position differs from previous inspection. Verify on site.”

That difference is fundamental because a photograph does not contain the entire regulatory context.

The technical components already exist in adjacent infrastructure applications. The Fraunhofer Institute for Industrial Mathematics ITWM describes its AEROS project as automatically detecting and locating relevant road-space objects such as traffic signs, traffic lights, pavement markings, and guardrails. Inspection vehicles collect imagery, while GPS and visual distance measurement help determine object locations.

Applied to temporary traffic control, the principle becomes valuable: the human performs the inspection and makes the professional decision, while the system remembers previous conditions, analyzes differences, and organizes evidence.

How could an AI-assisted inspection drive work from dispatch to completion?

The process should start before the vehicle leaves the depot.

Dispatch knows which work zones are active, when the last inspection occurred, which inspections are due, whether defects remain open, whether severe weather has created additional inspection requirements, and which employee is qualified and available.

The system can use this information to generate a proposed route.

On the mobile device, the inspection employee receives more than an address. The job record can contain the project number, client, work-zone section, traffic control plan, traffic order, responsible contact, previous defects, current phase of work, and the required inspection scope.

Once the employee reaches the location, GPS can associate the inspection with the appropriate site. The application can then guide the employee through project-specific checkpoints rather than one generic checklist.

Possible checkpoints include the advance warning area, beginning of the temporary traffic control arrangement, longitudinal channelization, transverse closure, pedestrian routing, cycling route, site access, temporary traffic signal, temporary barrier, and downstream termination.

Photos are assigned directly to the checkpoint where they were captured. Time, location, employee, and work order do not have to be retyped.

The employee can concentrate on exceptions.

If a delineator has been struck, the employee can dictate: “Second delineator after the intersection knocked down. Replaced during inspection. Completion photo recorded.”

Speech recognition and language models can turn this into structured data such as defect category, location, corrective action, status, and follow-up requirement.

The operational advantage appears immediately. Dispatch does not have to wait for a report to return to the office. A serious defect can be visible to the operations team while the employee is still on site.

Which inspection items are especially suitable for digital support?

Many elements of an inspection do not require AI at all.

A well-designed mobile workflow can already improve checks of temporary traffic signs, warning lights, delineators, barriers, portable signals, temporary markings, pedestrian and cycling routes, work-zone entrances, protective devices, visibility conditions, and changes caused by construction progress.

AI becomes more valuable when historical observations are available.

A single photograph provides limited context. A sequence of inspections from the same location creates a time series. The system can compare positions, appearance, configuration, recurring defects, and repeated maintenance events.

This turns a documentation system into a source of operational intelligence.

For example, management may discover that the same section of a closure repeatedly experiences displaced delineators. Another site may require unusually frequent battery changes. A third project may generate repeated corrective visits after weekends.

Those patterns can influence equipment choice, inspection planning, staffing, and discussions with clients.

Research in broader construction monitoring is already exploring similar concepts. Fraunhofer IDMT’s ConstructionsAIt project describes ten use cases for multimodal AI-based construction monitoring, including audiovisual traffic monitoring, worker safety applications, logistics monitoring, and automated detection of events. It is not an RSA 21 compliance system, but it illustrates the technical direction of machine-supported site monitoring.

How do paper records, inspection apps, and AI-assisted processes compare?

The final row is the most important one.

AI should not remove the responsible employee from the decision loop. Its economic value comes from reducing repetitive administrative work, assembling relevant information, and identifying records that deserve attention.

How can AI improve routing for inspection and maintenance vehicles?

Route planning becomes an important issue once a contractor manages a larger portfolio of active work zones.

The shortest route is not necessarily the best route.

One site may require an inspection after darkness begins. Another may have an unresolved warning-light defect. A third may be located along the same corridor, while a fourth needs an additional visit because of severe weather.

A modern planning engine can combine several operational constraints: required inspection windows, travel times, work-zone priority, outstanding defects, employee assignments, vehicle availability, stock carried in the service vehicle, and known site conditions.

AI and optimization methods can then produce a sequence that minimizes unnecessary driving without compromising operational deadlines.

The same model can integrate maintenance.

Suppose the morning inspection reported a warning lamp with a weak battery. Before generating the evening route, the system knows that replacement batteries should be loaded. If a damaged delineator was reported, the route may be assigned to a vehicle carrying replacement equipment.

Inspection, dispatching, and maintenance become parts of the same workflow rather than separate activities.

What commonly goes wrong when companies digitize inspection drives?

One of the most common mistakes is reproducing the paper checklist exactly on a smartphone.

The employee still completes dozens of fields at every location, except now the process involves scrolling, tapping, and typing outdoors. The company has technically digitized the form but has not improved the operating model.

A second problem is excessive photography.

More evidence is not automatically better evidence. Hundreds of photos per shift become difficult to use if the system does not identify the project, location, checkpoint, time, and reason for each image.

The same caution applies to GPS tracking.

GPS can provide strong supporting evidence that a device or vehicle was present at a location, but a route trace alone does not demonstrate that every traffic control element was inspected. Checkpoints, structured observations, photo records, and defect handling provide the operational context.

Another problem appears when AI output is stored as if it were a verified fact.

Computer vision models can produce false positives. Vehicles may obscure a traffic sign, darkness may affect the image, rain may alter reflections, or a camera angle may differ significantly from the previous inspection. AI observations therefore need a confirmation workflow.

The employee should be able to accept, reject, or correct a suggested defect.

That feedback has an additional benefit: over time, it produces higher-quality training and evaluation data for the company’s specific environment.

How could a real-world mid-market use case look?

Imagine a German traffic safety contractor managing work zones for utility construction, fiber deployment, municipal road projects, and civil engineering contractors.

During the afternoon, the operations center prepares the second inspection round.

One project had a damaged warning light during the morning inspection. Another site has an earlier required inspection window. A third lies directly along the service vehicle’s route. The software calculates the sequence accordingly.

At the first site, the driver follows the same documented checkpoints used during the previous inspection. The application compares new imagery with earlier photos and identifies a traffic sign whose orientation appears significantly different.

The system asks the employee to verify the sign.

The employee confirms that the sign has rotated, corrects its position, and captures a completion photo. The record now contains the initial observation, employee confirmation, corrective action, completion evidence, timestamp, and location.

At the next site, the employee notices a knocked-over delineator without any AI prompt. The employee records the issue by voice, replaces the device, and photographs the restored condition.

Dispatch can see that both issues have been closed before the vehicle reaches the next job.

That is a more realistic vision of AI-supported traffic safety than fully autonomous site approval. Most of the value results from many small improvements connected in one workflow.

Why should the PDF report be an output rather than the primary record?

Companies often focus heavily on the final inspection PDF because it resembles the traditional signed inspection sheet.

The more scalable approach is to treat the PDF as one representation of structured operational data.

The underlying system can store the project, inspection event, employee, checkpoints, timestamps, GPS information, observations, media, defects, corrective actions, follow-up tasks, and audit history as separate linked records.

This creates far more value later.

A manager can analyze which projects produce the most corrective work, which equipment fails repeatedly, which locations generate frequent third-party interference, and which clients or work-zone configurations result in additional inspection mileage.

The data can also feed dashboards and client reporting without employees creating separate spreadsheets.

An audit trail is equally important. If a record is changed after the inspection, the system should retain information about what changed, when it changed, and who performed the change.

Why should AI not make the final RSA 21 compliance decision?

A camera sees only what is inside its field of view.

Professional assessment may require the approved traffic control plan, the traffic order, local road geometry, current construction phase, temporary lane arrangement, pedestrian requirements, cycling facilities, client-specific provisions, and conditions imposed by the authority.

Some requirements are spatial. Others are operational. Still others depend on information that is not visually observable.

An AI model can therefore support the professional decision without becoming the decision maker.

It can detect an object, compare current and historical images, highlight a missing photo, identify an unusually long interval between inspections, extract information from voice notes, or prioritize a defect.

The qualified person remains responsible for interpreting the actual traffic situation and determining what action is appropriate.

This human-in-the-loop model also makes system performance easier to manage. Employees can understand why a record was flagged and provide corrective feedback when the model is wrong.

Why does work-zone inspection remain relevant despite established rules?

The issue is not merely administrative.

Germany’s Federal Statistical Office recorded 109 cases in 2024 in which a missing or inadequately secured road work site was listed among the general causes of crashes involving personal injury. The statistic does not establish that the crashes resulted from missed inspection drives, and it should not be interpreted that way. It does demonstrate that inadequate work-zone protection still appears in actual crash statistics.

This is an important distinction for digitalization projects.

The purpose should not be to produce a report faster simply because the report is mandatory. The more valuable objective is to identify deterioration earlier, organize corrective work, document exceptional inspections, and provide operations staff with an accurate current picture of the work zones they are responsible for.

How should a mid-sized German traffic safety company introduce this technology?

The first stage does not need computer vision.

The company should first establish digital master data for projects, work zones, responsible employees, inspection schedules, checkpoints, defect categories, and documentation requirements.

The second stage connects mobile inspection, GPS, photos, defect workflows, follow-up tasks, and automated reports.

At this point, the company already has a useful production system.

The third stage can introduce AI where enough structured data exists. Voice notes can be converted into structured inspection records. Image comparison can identify changes. Historical defect patterns can support prioritization. Route optimization can combine inspection schedules and maintenance requirements.

A later stage could compare detected objects with structured traffic control plan data, but this requires significantly more reliable source data and careful professional validation.

This gradual approach is usually more attractive to mid-market contractors than starting with a large computer-vision project. Each stage produces operational value while creating the data foundation required for the next one.

It also avoids a frequent technology problem: attempting advanced AI before basic project, asset, and inspection data are consistently managed.

For traffic safety companies, the practical benchmark is therefore straightforward. Employees should spend less time entering information twice, dispatchers should receive relevant defects earlier, service trips should be easier to coordinate, and every important event should remain traceable from inspection through corrective action.

Which sources support the figures used in this article?

Hessian Government Service Portal – Restriction of Public Road Space: Permit Information
The portal describes at least two daily inspections for longer-duration work zones, at least one daily inspection on non-working days, and additional inspection requirements following severe weather or storms.
https://verwaltungsportal.hessen.de/leistung?leistung_id=L100001_8964666&regschl=066310006006

Federal Statistical Office of Germany – General Causes of Personal-Injury Road Crashes
The 2024 data reports 109 cases involving a missing or inadequately secured work site as a general accident cause.
https://www.destatis.de/DE/Themen/Gesellschaft-Umwelt/Verkehrsunfaelle/Tabellen/allgemeine-unfallursachen-von-unfaellen-mit-personenschaden.html

Fraunhofer IDMT – ConstructionsAIt: Multimodal AI-Based Technologies for Automated Construction-Site Monitoring
The project white paper describes ten use cases for multimodal AI-supported construction monitoring.
https://www.idmt.fraunhofer.de/de/institute/projects-products/projects/constructionsait.html

Which sources are useful for further reading?

German Federal Ministry of Transport – General Circular Road Construction No. 24/2021: RSA 21
https://www.bmv.de/SharedDocs/DE/Anlage/StB/ars-aktuell/allgemeines-rundschreiben-strassenbau-2021-24.html

FGSV Publishing – RSA 21: Guidelines for Traffic-Related Safety at Road Work Sites
https://www.fgsv-verlag.de/rsa-21

Fraunhofer ITWM – AEROS: Automatic Recognition of Traffic Scene Objects
https://www.itwm.fraunhofer.de/en/departments/bv/ai-solutions-digitalization-sustainability/aeros.html

What are AI-assisted RSA 21 inspection drives?

AI-assisted inspection drives combine professional inspection of temporary traffic control with digital tools such as GPS, mobile checklists, photo documentation, speech processing, and automated data analysis. AI can identify changes between inspections, categorize observations, or prioritize routes. It does not replace the authority’s traffic order, applicable contractual requirements, or the professional assessment performed by the responsible qualified employee.

Does RSA 21 itself specify how many inspection drives are required?

The commonly referenced inspection intervals are particularly associated with ZTV-SA 97 and requirements incorporated into contracts or traffic orders. Longer-duration sites may require inspections around daybreak and after darkness begins, while non-working days may require at least one inspection. The actual obligation must always be determined from the applicable order, contract, local conditions, and project-specific requirements.

Can AI automatically determine whether a work zone complies with RSA 21?

A fully automated compliance determination would be unreliable because an image represents only part of the situation. The system may not know every requirement in the approved traffic control plan or the authority’s traffic order. AI is more suitable for detecting objects, identifying changes, finding missing documentation, and directing a qualified employee toward conditions that deserve professional review.

What information should a digital inspection drive record?

A useful record includes the project, work order, employee, inspection time, defined checkpoints, location data, inspection results, photos, defects, corrective measures, and any unresolved follow-up tasks. An audit history should also record later changes. Photos and observations should be associated directly with their corresponding project and checkpoint so employees do not need to reconstruct the evidence afterward.

Is GPS by itself sufficient evidence that an inspection was performed?

GPS can support evidence that a device or vehicle was present in a particular area at a certain time, but it does not demonstrate that every required traffic control element was examined. A stronger operational record combines GPS with defined inspection checkpoints, timestamps, photographs, observations, defect records, corrective actions, and a direct relationship to the work order and applicable traffic-control documentation.

What types of defects could computer vision detect during inspection drives?

Depending on training data, image quality, and camera position, computer vision can identify traffic signs, delineators, markings, warning lights, barriers, and other roadside objects. It can then compare observations with earlier inspections. Occlusion by vehicles, darkness, rain, dirt, construction traffic, and changing camera angles can reduce reliability, so an employee should verify any automatically generated defect recommendation.

Can AI optimize inspection vehicle routes as well?

Yes. Route planning can combine inspection deadlines, travel time, current defects, work-zone priority, weather events, employee assignments, and required replacement equipment. The goal is not simply to calculate the shortest route. It is to create an operational sequence that meets required inspection windows while reducing unnecessary mileage and integrating inspection activity with maintenance and corrective work.

What should companies consider when capturing road-side photos or video?

Images may contain vehicles, license plates, workers, pedestrians, or other identifiable information in addition to the traffic control equipment being inspected. Companies should therefore design access control, retention periods, purpose limitation, and data minimization into the system. Targeted photographs of relevant equipment may often be sufficient, while continuous video should only be used when justified by the operational purpose and legal assessment.

What is the most common mistake when digitizing inspection drives?

A common mistake is copying the existing paper inspection form directly onto a smartphone. Employees then complete the same administrative work through a less convenient interface. A better workflow automatically provides project information and asks employees primarily to document exceptions, defects, and corrective action. AI should be added only after the underlying inspection workflow and data structure function consistently.

Which traffic safety contractors benefit most from AI-assisted inspections?

Benefits generally increase with the number of simultaneous work zones, inspection events, service vehicles, and employees involved. Companies operating recurring maintenance routes, night inspections, weekend checks, multiple client contracts, or frequently changing temporary traffic arrangements have particularly strong use cases. Integrated digital workflows can reduce administrative effort while improving dispatching, defect follow-up, maintenance coordination, and access to inspection history.