Method for generating a motor vehicle repair quotation

Abstract

This method describes an automated process for generating quotations for the repair of damage to motor vehicles, wherein a first AI model is trained with historical repair data from different motor vehicle repair companies, including information regarding repair procedures, costs and required time. The first AI model is further optimized with specific data of a single repair company, including old quotations. In addition, a second AI model is deployed to characterize instances of damage to the vehicle based on received images. Based on these analyses, the first AI model determines suitable repair processes and generates accurate quotations. This approach offers an efficient solution for the traditional, time-consuming quotation process, by automatically drawing up quotations based on visual damage information and AI technology.

Description

[0001] METHOD FOR GENERATING A MOTOR VEHICLE REPAIR QUOTATION

[0002] TECHNICAL FIELD

[0003] The present invention relates to a method for automatically generating quotations for the repair of damage to motor vehicles using Al models.

[0004] PRIOR ART

[0005] US 2010 / 0211421 Al describes a method for automatically generating a quotation for a custom construction project. The method comprises receiving data corresponding to construction components selected by the customer, such as cabinets, from a kiosk system for a custom construction project. This data is converted into an automatically generated cost quotation for the custom construction project, wherein the cost quotation contains a series of cost values for a dealer and a series of cost values for a competitor of the dealer. The quotation is sent to the kiosk system for presentation by the kiosk system. Additionally, extra data is received corresponding to a construction layout specified by the customer for the custom construction project.

[0006] CN115879966A describes a method and apparatus for the automatic generation of price lists, as well as equipment and a storage medium. The method comprises the following steps: obtaining a quotation work object created by a user and associating the quotation work object with a business work object and a user story interface; acquiring customer information in the business work object and obtaining customer information in the user story data via the user story interface; determining whether the customer information in the user story data is the same as the customer information in the business work object; and, if the user story data is the same as the user data, reading a predefined quotation template and populating the quotation template with the user story data in order to generate a quotation. According to the embodiment of the invention, the quotation work object is linked to the business work item and the user story data, such that the price list for each business work item corresponding to the customer request to be quoted is generated, which improves the efficiency of the generation of price lists and saves labor costs and time costs.

[0007] The documents cited in the search report, including DI to D4, relate exclusively to the detection, analysis and documentation of damage to motor vehicles, for example for the purposes of damage inspection, damage reporting or initial quotation determination.These systems analyze static images or short image sequences of a damaged vehicle and are not designed to track, interpret or structure repair actions during the execution thereof.

[0008] In particular, D1-D4 do not disclose systems that monitor repair activities in real time, do not perform recognition of workers based on visual identification means, do not detect tools or actions during execution, and do not perform automatic time recording per repair task or per vehicle zone.

[0009] Nor does the prior art describe a system wherein visual sensor data, personnel identification and vehicle identification are combined via sensor fusion into a structured repair process log, nor a system that functions autonomously and locally ("edge computing") within the workshop.

[0010] The invention aims to facilitate and accelerate the drawing up of a quotation for the repair of damage to motor vehicles.

[0011] There is therefore a need for a technical system that not only analyzes damage, but monitors the entire repair process objectively, automatically and reproducibly, without manual time recording or subjective reporting by employees.

[0012] More particularly, there is a need for a system capable of detecting in real time which repair actions are actually being performed, by whom, with which tool, on which vehicle zone and during which time duration, and which structures this information into reliable process data suitable for automatic final calculation, quality control and process optimization.

[0013] SUMMARY OF THE INVENTION

[0014] A method for generating a quotation for the repair of damage to a motor vehicle, comprising the steps of: training a first Al model with repair data originating from different motor vehicle repair companies, wherein the repair data comprises historical information relating, inter alia, to repair procedures, costs, and time required for repairing different types of damage to motor vehicles, preferably at least partly derived from previous quotations of the different motor vehicle repair companies; further training the first Al model with repair data originating from a specific motor vehicle repair company, wherein the repair data comprises historical information relating, interalia, to repair procedures, costs, and time required for repairing different types of damage to motor vehicles, preferably in the form of previous quotations prepared by the motor vehicle repair company; receiving images of a damaged motor vehicle; characterizing one or more instances of damage on the damaged motor vehicle on the basis of the received images of the damaged motor vehicle by means of a second Al model trained with data of undamaged and damaged motor vehicles, wherein the second Al model, inter alia, performs the following steps: recognizing the model of the motor vehicle from the images, comparing the images of the damaged motor vehicle with images of undamaged motor vehicles of the same model from a database and identifying and locating instances of damage on the basis of the comparison, identifying and locating different motor vehicle components, and associating the instances of damage and the motor vehicle components on the basis of said identifications and locations; determining one or more repair processes for the motor vehicle repair company by means of the first Al model, wherein the determination is based on the characterized one or more instances of damage and the repair data originating from the motor vehicle repair company, and wherein the determined repair process is suitable for repairing the one or more characterized instances of damage; generating one or more quotations for the motor vehicle repair company with the first Al model on the basis of the one or more determined repair processes.

[0015] The described method offers a solution for the time-consuming and error-prone traditional quotation process by automatically drawing up a quotation for motor vehicle repair companies on the basis of images of a damaged motor vehicle. For this purpose, Al models are deployed to characterize the damage and to predict a suitable repair method, on the basis of which an accurate quotation is subsequently drawn up.

[0016] In one embodiment, the system comprises a final calculation module which is configured to automatically draw up a final repair calculation on the basis of the data collected by the time recording engine and sensor fusion module.

[0017] The final calculation is based on:

[0018] - the actually performed repair actions;

[0019] - the recorded duration per action;

[0020] - the involved employee(s);

[0021] - the tools used;

[0022] - and the specific vehicle zones on which the actions have been performed.

[0023] The final calculation can be automatically enriched with supporting evidence, such as image fragments, still images and process logs, whereby an objective and verifiable file is created.In one embodiment, the final calculation is automatically presented to an expert, claims handler or insurance company, wherein the calculation serves as substantiation of the final repair cost.

[0024] DETAILED DESCRIPTION

[0025] Unless defined otherwise, all terms used in the description of the invention, including technical and scientific terms, have the meaning as commonly understood by a person skilled in the technical field of the invention. For a better assessment of the description of the invention, the following terms are explicitly explained.

[0026] As used in this document, the articles "a", "an" and "the" refer to both the singular and the plural, unless the context clearly dictates otherwise. For example, "a segment" means one or more than one segment.

[0027] The terms "comprise", "comprising", "consist of", "consisting of", "provided with", "have", "having", "include", "including", "contain", "containing" are synonyms and are inclusive or open terms that indicate the presence of what follows, and which do not exclude or prevent the presence of other components, characteristics, elements, members, steps, as known from or disclosed in the prior art.

[0028] The term "user" refers to the persons who may make use of the present invention. These are mainly employees of motor vehicle repair companies, but can possibly also be external experts or owners of a damaged motor vehicle.

[0029] The term "motor vehicle" refers to a vehicle driven by a motor that is primarily designed for transport over the road, such as cars, trucks, buses and motorcycles, but can also refer to agricultural vehicles, ATVs or camper vans.

[0030] The invention relates to a method according to claim 1 for generating a quotation for the repair of damage to a motor vehicle.

[0031] The process of drawing up a quotation for the repair of damage to a motor vehicle is traditionally a time-consuming and complex process for motor vehicle repair companies. The quotation procedure requires many labor-intensive resources, such as a thorough physical inspection by an experienced mechanic to determine the damage, the manual documenting of findings, and the calculating of costs based on parts, labor hours andthe complexity of the repairs. The variations in damage extent, motor vehicle models and the required repair methods make it difficult and time-consuming to draw up consistent and accurate quotations. This process often requires extensive experience to make an accurate estimation of both the extent of the damage and the costs and time which are necessary for the repairs. Small errors in the estimation can lead to financial losses or customer dissatisfaction.

[0032] The described method offers a solution for the time-consuming and error-prone traditional quotation process by automatically drawing up a quotation for motor vehicle repair companies on the basis of images of a damaged motor vehicle. For this purpose, Al models are deployed to characterize the damage and to predict a suitable repair method, on the basis of which an accurate quotation is subsequently drawn up.

[0033] A first Al model is trained with historical repair data from various motor vehicle repair companies, wherein the repair data comprises information, among other things, about repair procedures, costs, and required time for repairing various types of damage to motor vehicles, preferably in the form of old quotations drawn up by motor vehicle repair companies. A quotation comprises detailed information about the individual cost items. For example, a quotation for the repair of a broken headlight and a scratch on the front bumper of a car comprises information regarding the price for the replacement of the headlight, the costs for painting the bumper, and the number of labor hours that was required for both repairs. In addition, the quotation states which materials have been used, such as the paint and the headlight unit. By processing old quotations in the Al model, the model can learn which steps are necessary for various instances of damage, how costs are structured, and how much time such a repair usually requires. This information enables the Al model to predict, in future cases, an accurate repair method and the associated quotation on the basis of a characterized damage.

[0034] The repair process is, however, strongly dependent on the specific methods and techniques employed by each motor vehicle repair company. Different companies may use unique techniques, tools and materials, as well as variations in the efficiency and experience of their employees. To make an accurate estimate of a quotation which is specifically tailored to a particular motor vehicle repair company, the first Al model is additionally trained with historical repair data originating from that specific company. This data comprises, among other things, information about previous repairs, costs, used materials and the duration of various repair methods. By feeding the first Al model with this company-specific data, it learns which processes and costs are characteristic of the relevant motor vehicle repair company. This results in a customized quotationthat takes into account the unique circumstances and practices of the company, which significantly increases the accuracy of the quotation and thereby contributes to improved customer satisfaction.

[0035] To characterize the damage, a second Al model is trained with images of both damaged and undamaged motor vehicles, whereby the second Al model learns how different instances of damage can visually appear on an image. On the basis of this information, the Al model can recognize the type of motor vehicle by comparing images of the damaged motor vehicle with images of other damaged and undamaged vehicles of the same type, which are stored in a database and labeled with the corresponding information about the present components and instances of damage. Through further comparisons with these images, the second Al model can identify the damaged components and accurately characterize the damage. The model learns, for example, to recognize the depth of a scratch or dent on the basis of reflections and contrasts in the image, and to detect the internal damage on the basis of visible deviations on the exterior of the motor vehicle.

[0036] On the basis of the damage characterized by the second Al model, the first Al model can automatically determine one or more suitable repair processes and draw up the associated quotation, wherein the specific circumstances of the motor vehicle repair company are taken into account.

[0037] In one embodiment, the said images comprise close-up photos or videos of specific instances of damage, as well as various photos or videos which show a broader view of the motor vehicle, preferably from different angles.

[0038] In one embodiment, for the step of comparing the images of the damaged motor vehicle with images of undamaged motor vehicles of the same model from a database, 'template matching' is used.

[0039] In one embodiment, the second Al model applies semantic segmentation to determine the exact locations of the instances of damage. This entails that the model classifies each pixel in the images as belonging to a specific component and / or damage area.

[0040] In one embodiment, Convolutional Neural Networks (CNNs) are used for the second Al model. CNNs are highly suitable for recognizing visual patterns in complex image data.In one embodiment, Random Forests are used for the first Al model. Random Forests are suitable for processing structured data and recognizing patterns.

[0041] In one embodiment, Support Vector Machines (SVMs) are used for the second Al model. SVMs are suitable for separating data into different classes and the second Al model is applied to classify different damage.

[0042] In one embodiment, Gradient Boosting is used for the first Al model. Gradient Boosting works by iteratively correcting errors from previous predictions, wherein each subsequent model focuses on the remaining errors of the preceding one. This continuous optimization ensures that the Al model becomes increasingly accurate as it processes more data, making it extremely suitable for fulfilling the function of the first Al model.

[0043] In a preferred embodiment, the first and / or the second Al model uses one or more of the following techniques: vision language model; advanced computer vision; expert systems (knowledge systems); hybrid approaches. It goes without saying that future techniques can also be implemented.

[0044] In a preferred embodiment, the first and / or the second Al model uses one or more of the following architectures for processing of the data: real-time processing systems; batch processing systems; cloud-based system; edge computing systems; distributed processing system. It goes without saying that future architectures can also be implemented.

[0045] In one embodiment, the method comprises the step of displaying information on a graphical user interface (GUI), wherein the information comprises at least one, or a combination of the following elements: the characterized one or more instances of damage, at least one generated quotation, a duration of the determined repair process and / or a duration of separate sub-processes within the determined repair process. Through the GUI, users can easily gain access to relevant data. This makes the system not only more accessible, but also more interactive, whereby the repair process is monitored more efficiently and can be controlled. Showing the characterized instances of damage enables users to verify what damage has been detected, which is an important step for the accuracy of the quotation. In a further embodiment, the GUI offers insight into the estimated duration of the repair, both for the entire process and for separate sub-processes, whereby motor vehicle repair companies can better manage their expectations and planning. This increases the transparency of the repair process,reduces possible misunderstandings and offers users the possibility to intervene proactively if necessary.

[0046] In one embodiment, adjustments in a determined repair process and / or generated quotation are displayed on the GUI, and the method comprises the steps of asking for a permission for said adjustments via the graphical user interface, and automatically adjusting the quotation on the basis of the granted or refused permission.

[0047] By proactively involving the user in the decision-making process regarding possible adjustments to a specific repair process or a generated quotation, transparency is increased and it is prevented that adjustments are carried out without the user being aware thereof. Requesting permission via the GUI ensures a direct and efficient communication with the motor vehicle repair company. This makes it possible to make decisions quickly without delays, which contributes to a more streamlined and efficient process. When a user agrees to certain adjustments, or refuses certain adjustments, the quotation is automatically adjusted on the basis of the granted or refused permission. This ensures that the quotation is always up-to-date and corresponds fully with the actual scope of the repairs and the expectations of the user.

[0048] In one embodiment, the method comprises the steps of repairing the damaged motor vehicle according to a determined repair process, obtaining image data of the repairing, automatically recognizing objects in the image data and automatically updating the generated quotation on the basis of the recognized objects and / or a time at which the objects are recognized, the recognized objects comprising one of, or a combination of, the following elements: employees, tools, motor vehicles, motor vehicle components.

[0049] During the repair of the damaged motor vehicle, the repair process is continuously monitored using image data that is obtained in real-time from the workshop.

[0050] The image data is analyzed by one or more Al models which are configured to detect repair actions on the basis of interactions between employees, motor vehicles, motor vehicle components and tools. This can include recognizing whether an employee is performing tasks such as disassembling, assembling, sanding, dent repair, filling, painting, drying, or inspecting.

[0051] By combining object detection, pose estimation and behavior recognition, the system can determine not only which action is being performed, but also when it starts and ends, by whom it is performed, and to which vehicle zone it relates.In a further embodiment, the image data is obtained using one or more cameras which are installed above the workplaces. In an alternative embodiment, the image data is obtained using portable cameras, worn by the employees of motor vehicle repair companies.

[0052] On the basis of this information, it can be determined which activities of the repair have been completed, whereby the system can dynamically update how much work still needs to be done and whether extra costs must be included. If unexpected complications occur, such as extra required labor or components, this can be processed immediately in the quotation. This leads to a realistic and accurate cost estimation. In a further embodiment, the motor vehicle repair company receives direct updates about the progress of the repairs based on the recognized objects. This approach results in a more transparent repair process, whereby the employees of the motor vehicle repair company can make better informed decisions and thereby can work more efficiently.

[0053] In another or further embodiment, the first Al model is used for analyzing the image data. Because the first Al model is already trained with different repair processes, it can recognize the activities faster. Thereby, the first Al model can immediately apply the obtained information when adjusting the generated quotation.

[0054] In another or further embodiment, object detection is applied to analyze the activities of employees in the workshop in real-time. The system identifies employees and tracks their movements, which gives insight into which work is being carried out and whether this is in line with the determined repair process. This information enables the system to further refine the quotation because the progress of the repairs is correctly tracked. Thereby, it can be recorded which employees perform which tasks, and how fast they perform them. On the basis of this, the necessary time for the different activities can be predicted better, whereby a more accurate estimation of the quotation can be made.

[0055] In another or further embodiment, object detection is applied for recognizing tools and / or components that are used for the repair. This increases the accuracy with which the activities of the employees can be determined and thereby improves the prediction of the quotation.

[0056] In one embodiment, after the repair of the damaged motor vehicle, a final calculation is made on the basis of which a definitive quotation is drawn up, wherein the final calculation is based on the objects recognized in the obtained image data and the timecorresponding to the recognized objects, from which the duration of the different repair actions and the used means is derived. During the repair process, the model automatically records the time spent on the vehicle. This system offers a real-time overview of the exact labor duration per motor vehicle. This enables the motor vehicle repair company to generate a detailed and accurate final quotation.

[0057] The final calculation is automatically drawn up on the basis of the recorded repair actions and associated time data, wherein the calculation is not merely based on previously estimated values, but on the actually performed activities.

[0058] In one embodiment, the final calculation comprises a breakdown by repair action and by vehicle zone.

[0059] - the performed action;

[0060] - the effective duration;

[0061] - the tools used;

[0062] - and, if relevant, the involved staff member.

[0063] The final calculation can be automatically supplemented with visual evidence, such as images or image fragments originating from the repair process, and with automatically generated process logs. This creates a transparent and verifiable file that can be submitted to loss adjusters or insurance companies.

[0064] In one embodiment, the system automatically records the duration of each recognized repair action. The time recording takes place without manual input from employees and is based on detecting start and end times of specific actions in the image data. Detecting start and end times can be automated (e.g., based on image data, such as movements of the employee, moving away of the employee relative to the vehicle, absence of previously detected damage, etc.) and / or manual (e.g., by an employee recording this).

[0065] The recorded time data is linked to:

[0066] - the recognized employee;

[0067] - the tools used;

[0068] - the specific motor vehicle zone;

[0069] - and the type of repair action.

[0070] This creates a detailed and objective time profile of the complete repair process.

[0071] In a further or alternative embodiment, the first Al model processes purchase invoices of the damaged motor vehicle, wherein references such as the registration plate, filenumber and / or chassis number are automatically recognized, and based on which the quotation is generated. This automated approach minimizes administrative errors and increases efficiency, while the company can offer a more transparent cost overview to customers.

[0072] In another or further embodiment, the method comprises the step of displaying information on a graphical user interface (GUI), wherein the information comprises at least one of, or a combination of, the following elements: the progress of the repairs in real time, adjustments in the determined repair process, changes in a generated quotation; and wherein the information is based on the recognized objects. As a result, users can be directly informed when changes take place in the repair process and what consequences this has for the quotation. This ensures quick communication within the motor vehicle repair company, which increases the efficiency of the work and reduces the risk of errors. In addition, this offers the possibility to keep the owner of the motor vehicle that is being repaired directly informed of the progress of the repairs. The owner remains better informed of the repair process as a result, which leads to better customer satisfaction.

[0073] In another or further embodiment, the information of the progress of the repairs comprises photographs of the motor vehicle that is being repaired. This provides visual evidence of the performed activities, whereby the transparency and reliability of the process increase. Photos provide a more complete picture of the progress of the repairs so that any errors can be observed more easily and the users have a clearer picture of the progress of the repairs.

[0074] In a further embodiment, said photos are used to train the second Al model. These images help the model to capture variations in damage patterns and to recognize subtle deviations, such as hidden internal problems that might otherwise be overlooked. Moreover, feedback from employees, who indicate when the model has not correctly recognized certain damage, can provide valuable information for further optimization of the model. As a result, the Al model remains relevant and it can adapt to the changing damage repair processes, which leads to a more efficient process and increased customer satisfaction.

[0075] In one embodiment, the method comprises the step of training the first Al model on the basis of the recognized objects and / or the time at which the objects are recognized. The recognized objects in the image data, such as for example which employees are working on specific tasks, which tool is used and which materials are involved therein, providevaluable information for training the first Al model because they provide insight into the operational processes of the motor vehicle repair company. This data makes it possible for the model to recognize patterns and to evaluate the progress of different working methods, which leads to more accurate cost estimates and time estimates which are more specifically tailored to the motor vehicle repair company. Moreover, the model can learn which combinations of employees, tools and materials are the most efficient, whereby it can optimize the performance of the repair process. This enables the Al model not only to draw up more accurate quotations, but also to give recommendations for the most efficient repair methods.

[0076] In one embodiment, the method comprises the step of providing an upload function integrated into a graphical user interface (GUI), wherein the upload function is configured to upload data relating to one or more instances of damage, and wherein, during the step of receiving data from a damaged motor vehicle, said data is received via said upload function, and wherein the graphical user interface (GUI) is suitable for implementation on a mobile device, preferably a mobile phone and / or tablet. This enables users to document the damage at the location of the motor vehicle, whereby a quotation can be generated immediately at the location of the damaged motor vehicle. Employees of the motor vehicle repair company and the owner of the damaged motor vehicle can thereby, during inspecting the damage, directly upload a photo of the damage and obtain an accurate quotation. Both parties thereby directly have a realistic idea of the situation, which accelerates the entire repair process.

[0077] In one embodiment, the method comprises the step of receiving a description of the one or more instances of damage, and wherein the step of characterizing the one or more instances of damage is performed on the basis of said description using a Natural Language Processing (NLP) model. Basing the characterization of damage on a description improves the accuracy, because it may contain explanations regarding aspects that are not visually visible, such as hidden structural damage or other internal problems. By using Natural Language Processing (NLP), the description can be automatically accurately analyzed so that relevant information is extracted. This step enables a more complete understanding of the damage and leads to a more accurate characterization of instances of damage, which makes the generated quotation more accurate.

[0078] In one embodiment, the second Al model and Natural Language Processing (NLP) are combined in one vision-language model. This increases the efficiency because image and language are processed simultaneously, reduces the complexity of the architecture,and improves the interpretation of multimodal relationships, whereby more accurate and contextually more relevant results are achieved.

[0079] In one embodiment, the first Al model and the second Al model are integrated into a single Al model. Combining the first and second Al model into a single model ensures that the information from both processes is shared and is accessible within the same system. This allows the Al model, when generating quotations, to take into account not only the full information of the detected damage, but also directly the specific repair processes and costs. This leads to a more complete and accurate analysis. Because the data from both processes is integrated, the system can better determine which repairs are most suitable for the recognized damage, which results in more accurate quotations.

[0080] In one embodiment, the first Al model and / or the second Al model are designed as composite systems comprising multiple sub-AI models, each configured to perform a specific function within the overall framework of the Al model. These sub-AI models can work independently or in a coordinated manner to achieve the desired functionality. In the first Al model, individual sub-AI models can, for example, specialize in tasks such as analysis of historical data, cost prediction and pattern recognition, wherein each model focuses on a separate aspect of the repair estimation process. In the same way, the second Al model can contain sub-AI models that are dedicated to image segmentation, object detection and feature classification, which cooperate to extensively analyze and characterize vehicle damage. In certain configurations, these sub-AI models can function as separate entities, which communicate via well-defined interfaces to exchange data and insights. Alternatively, they can be integrated into a uniform architecture, wherein shared resources and interconnected layers are utilized to perform all tasks cohesively as a single Al model.

[0081] In one embodiment, the method comprises the step of comparing the generated quotation, or a part thereof, with a quotation, or a part thereof, drawn up by the motor vehicle repair company, wherein the compared quotations, or parts thereof, relate to similar repair processes or similar sub-processes within repair processes, and adjusting the first Al model and / or approving or adjusting the generated quotation, or part thereof, based on the comparison. Through this comparison, deviations between the quotation generated by the Al and the quotation of the repair company can be identified and corrected. This ensures that the Al quotation is increasingly better aligned with the actual repair processes and the associated costs. Furthermore, any unrealistic estimates of the quotation are hereby prevented. Furthermore, this promotes a continuous learning process for the first Al model. The quotations drawn up by the motor vehiclerepair company reflect the actual repair process and the actual costs. Through the comparison of the generated quotations with quotations that have been drawn up by the motor vehicle repair company, the model can adjust and optimize its algorithms based on the observed deviations, so that the generated quotations correspond more accurately to the actual costs of the repairs.

[0082] In a further embodiment, the step of training the first Al model comprises automatically parsing historical quotations and / or invoices, preferably associated with the same company, into a structured set of repair components, wherein each quotation is split into:

[0083] - separate repair actions;

[0084] assigned vehicle components and vehicle zones;

[0085] - associated labor duration per action;

[0086] used materials and components;

[0087] - and associated cost parameters.

[0088] To this end, use is made of one or more extraction modules which recognize and normalize textual, numerical and semantic elements from unstructured or semistructured invoice documents.

[0089] In a further embodiment, the repair components extracted from historical invoices are coupled to damage representations which have been generated by the second Al model, wherein the damage is encoded as a multidimensional damage vector which comprises at least:

[0090] - type of damage (for example scratch, dent, breakage);

[0091] severity level;

[0092] - affected vehicle component;

[0093] and spatial location on the vehicle.

[0094] The first Al model learns on the basis of this coupling which combinations of damage vectors have historically led to specific repair actions, time expenditures and cost structures.

[0095] In a further embodiment, historical invoices are used as ground truth data ("ground truth") for validating and adjusting the first Al model, wherein predicted quotations are compared with actually invoiced repair costs for similar instances of damage.

[0096] Deviations between predicted and historical costs are automatically analyzed, wherein the Al model learns which parameters (such as damage severity, vehicle type or chosen repair method) are responsible for systematic overestimation or underestimation.In a further embodiment, the processing of historical invoices comprises a normalization step, wherein cost parameters are corrected for company-specific, regional or temporal factors, such as:

[0097] hourly rates;

[0098] regional wage differences;

[0099] inflation or price indexation;

[0100] and availability of components.

[0101] As a result, the first Al model can distinguish between structural differences in repair processes and purely economic variations, so that the predicted quotation remains technically based on the repair process itself.

[0102] In a further embodiment, the first Al model is configured to handle incomplete, inconsistent or differently structured historical invoices, wherein missing parameters are derived based on statistical correlations with comparable repair actions in the dataset. If, for example, no explicit labor duration is stated, the model can reconstruct this based on historical averages for the same repair action and vehicle zone. This ensures that the system not only assigns a price to a detected instance of damage, but reconstructs a complete repair process based on actually performed repairs from the past, whereby the generated quotation is reproducible, verifiable and company-specific.

[0103] In one embodiment, the method comprises the step of adapting a determined repair process based on information provided by one or more employees of the motor vehicle repair company, wherein the information relates to a change in the determined repair process and / or a change in the characterized one or more instances of damage. This promotes the accuracy of the determined repair process because employees can share valuable insights and practical experiences that influence the execution of the work. The employees have valuable experience whereby they can recognize deviations of the characterized damage or the determined repair process from reality. This information is therefore essential to estimate a realistic quotation. In addition, it increases the flexibility of the repair process. When employees report changes, such as the discovery of hidden damage during the repair, the first Al model can immediately update the quotation so that all users are aware of this new information.

[0104] In a further embodiment, the method comprises the steps of repairing the damaged motor vehicle according to a determined repair process, and training the first Al model based on the information provided by one or more employees of the motor vehicle repair companies, wherein the provided information relates to the repairing. The experience of the employees of the motor vehicle repair company is valuable information becausethis concerns specific damage and can therefore not be obtained anywhere else. This information can relate to specific actions that must be performed and were not yet foreseen by the first Al model, or to incomplete characterization of the damage by the second Al model. This can also concern photos of newly encountered damage or the progress of the repairs. By using this information for training the first Al model, the accuracy of the generated quotation can be significantly increased because the repair process can be characterized more accurately by the first Al model. In addition, this information provides valuable insights into the method of the different employees of the motor vehicle repair company, whereby the quotation can be tailored more accurately to methods specific to the motor vehicle repair company.

[0105] In a preferred embodiment, the analyzes of the image data are performed locally via an edge computing architecture which is installed in or near the workshop.

[0106] By performing the processing locally, large amounts of raw video data is not forwarded externally, which contributes to data protection, lower latency and higher reliability. Only structured metadata, such as recognized actions, duration and process information, is stored or shared externally. This architecture makes the system suitable for autonomous operation per repair location, wherein Al models can be adjusted locally based on specific methods of the concerned motor vehicle repair company.

[0107] In an alternative embodiment, the system is combined with one or more autonomous image acquisition units, implemented as physical passages or photo tunnels, which can be placed at strategic locations. These photo tunnels can also be considered separately as an invention, ideally in combination with one or more aspects of the remainder of this document, in particular then the manner of detecting damage.

[0108] When a motor vehicle is located in the image acquisition unit, preferably when a motor vehicle drives through such an image acquisition unit, images of the vehicle are automatically captured from multiple angles. These images are used for vehicle identification and damage analysis, after which the system can automatically propose one or more repair options. Preferably, images are taken of the underside, sides, front, rear and / or top of the vehicle.

[0109] In one embodiment, said autonomous image acquisition units are embodied as publicly accessible physical passages or photo tunnels which are arranged at public or semipublic locations, such as for example gas stations, supermarkets, parking facilities, rental companies, leasing companies or logistics hubs.In one embodiment, the owner of the vehicle receives multiple proposals from suitable repairers, as well as indicative repair dates. Because the image acquisition units are not managed by a specific repairer, an insurance company can determine objectively which repairer is technically most suitable.

[0110] In one embodiment, based on the image data obtained in the tunnel, the damage is automatically identified using the second Al model, and subsequently, using the first Al model, separate cost estimates and / or quotations are generated for multiple different motor vehicle repair companies. This information is subsequently provided to an involved party such as for example the owner of the vehicle and / or the insurance company that insures damage to the vehicle.

[0111] In a further embodiment, the generated information comprises, in addition to the cost estimate, also one or more additional selection parameters, including: an estimated availability or waiting time of the motor vehicle repair company; a geographical distance or travel time between the motor vehicle and the repair company; and / or a quality score or preference score assigned by an insurance company or claims handler. These parameters are combined to propose one or more recommended repairers to the owner of the motor vehicle, wherein the recommendation is based on an optimization between cost, time, distance and quality.

[0112] In a further embodiment, the image acquisition unit is also suitable for analyzing the tire condition of the motor vehicle, wherein the second Al model is applied to detect features such as tire wear, tread depth, uneven wear and visible damage.

[0113] In a further or alternative embodiment, a conformity analysis is performed based on the images obtained with the image acquisition unit, wherein it is determined whether the motor vehicle complies with predefined technical and / or administrative requirements, such as statutory safety standards, inspection requirements or maintenance regulations.

[0114] In a further or alternative embodiment, a safety assessment is performed based on the images obtained with the image acquisition unit, wherein it is assessed whether essential safety components of the motor vehicle, such as lighting, mirrors, tires and bodywork parts, comply with established safety criteria.

[0115] The result of this safety assessment can be used to identify and prioritize urgent repairs, and can if desired be shared with an insurance company or maintenance provider.In a further embodiment, after obtaining image data with the image acquisition unit, an integrated report is automatically provided electronically to the owner of the motor vehicle, for example via email, a mobile application or a web portal. This report preferably comprises one or more elements of the following list: the identified damage; one or more proposed repair options; cost estimates per motor vehicle repair company; results of the tire, conformity and safety analysis; an indication of the maintenance status of the motor vehicle. In a further embodiment, costs for which the owner of the vehicle is insured are not included in the report, but shared with the corresponding insurance company. For example: in the case of damage, which is covered by the insurance, the report states exclusively the nature of the damage and the available options for repair and the repair company. If the damage is not covered by the insurance, the generated cost estimate or quotation is provided to the owner of the vehicle.

[0116] In a further embodiment, the system determines, on the basis of the image data, historical vehicle data and / or maintenance intervals, an indication of a next maintenance service or periodic inspection.

[0117] In doing so, the system may automatically generate proposals for different maintenance companies, including: estimated maintenance costs; availability; and geographical proximity, wherein the cost estimates are personalized again per maintenance company in an analogous manner as described for repair quotations.

[0118] This embodiment may function independently of the repair monitoring system in the workshop, but may be combined therewith within one integrated solution.

[0119] In the following, the invention is described by means of non-limiting examples illustrating the invention, and which are not intended or to be interpreted to limit the scope of the invention.

[0120] DESCRIPTION OF FIGURES

[0121] Figure 1 shows a rear view of a photo tunnel (1) which is arranged for the automatic capturing of images of a vehicle (3). In the photo tunnel (1), a plurality of cameras (2) are arranged, wherein these cameras are positioned laterally, at the underside and at the top side of the tunnel. This configuration makes it possible to record the vehicle (3) from different angles and perspectives. During use, the vehicle (3) moves through thephoto tunnel (1), wherein the cameras (2) record images of the vehicle (3) at predetermined moments.

[0122] Figure 2 shows a side view of the photo tunnel (1) as shown in Figure 1, wherein the positioning of the cameras (2) along the length of the tunnel is visible.

[0123] EXAMPLES

[0124] EXAMPLE 1

[0125] The first Al model is trained by making use of historical repair data which originates from different motor vehicle repair companies. This comprises different types of data, such as old quotations, detailed cost overviews for components and labor, the duration of various repairs and information about materials used. Possible sources of this information are internal databases of the repair companies, such as software for damage management and invoicing, in which previous repairs and the associated costs are recorded. In addition, data may originate from documentation, such as work orders and reports, which have been drawn up by employees after the completion of repairs. Data from customer feedback and service history, or market prices for components and labor may also be used for training the first Al model, to ensure that the Al model obtains a more complete picture of the variables which influence the repair costs and time. By combining these diverse information sources, the first Al model can identify patterns and correlations which help in accurately predicting costs and time for future repairs on the basis of the instances of damage of the motor vehicle.

[0126] The training of the first Al model begins with collecting and preparing this data, wherein incomplete or inconsistent information is corrected. Subsequently, machine learning algorithms, such as Random Forests or Gradient Boosting, are applied to the dataset to identify patterns and correlations.

[0127] During the training process, the model learns, for example, which repair methods are typically applied for specific instances of damage and how these are related to the associated costs and time. After the training, the Al model is validated, wherein it is tested with a separate dataset, possibly consisting of 10% to 20% of the collected data, but this can also be between 5% and 10%, or between 20% and 40%. This dataset contains examples of previous repairs for which the actual costs are already known. The model predicts the costs for these repairs on the basis of the input variables that it has learned during the training. Subsequently, these predictions are compared with the actual costs to determine how accurate the model is. This process makes use of various statistical metrics, such as the mean absolute error (MAE) and the mean absolutepercentage error (MAPE), to obtain a clear picture of the model's performance. The results of the validation are carefully analyzed. If the accuracy of the predictions does not meet the expectations, adjustments are made to the Al model. This can entail that hyperparameters are tuned, such as the number of trees in a Random Forest or the depth of the trees, to improve the generalization capacity of the model. Additionally, new data, such as more recent repair information, can be added to enrich the training of the model and to further increase the ability to make accurate predictions of the repair process and the costs thereof.

[0128] The second Al model is trained by means of a dataset which consists of images of both damaged and undamaged motor vehicles. In the first step, these images are collected from various sources, including internal databases of motor vehicle repair companies, in which previous instances of damage are recorded, and other databases which contain reference photos of various types of motor vehicles. Herein, each image is labeled with relevant information about the damage, such as the type of damage, the location and the severity thereof. Subsequently, the collected images are cleaned and prepared for analysis. This entails that the images are normalized and annotated, so that the second Al model can learn which visual features correspond to which specific damage. The annotations contain important information, such as identifying components of the vehicle and classifying the types of damage, such as dents, scratches or fractures.

[0129] The training of the second Al model takes place by means of techniques such as Convolutional Neural Networks (CNNs), which are particularly effective for image recognition. During this phase, the model learns by analyzing the annotated images and identifying patterns and features that are specific to various types of damage. The Al model makes use of a process of feedforward and backpropagation, wherein it compares its predictions with the actual labels and adjusts its parameters to improve the accuracy. After the initial training, the model undergoes a validation phase, wherein a separate dataset of images is used to test the performance of the model. The accuracy of the predictions is evaluated using statistical metrics, such as the accuracy and the Fl-score, to determine how well the model is able to correctly identify and classify damage. Based on the validation results, adjustments are made to the model in the optimization phase, such as refining hyperparameters or applying data augmentation techniques to increase the variability of the training data. This ensures that the model becomes more robust and is better able to generalize to new, unseen images.

[0130] Finally, a Natural Language Processing (NLP) model is trained with a dataset consisting of textual descriptions of instances of damage, originating from employees of motorvehicle repair companies or from customer feedback, in which customers share their experiences and comments regarding the damage and repairs. Additionally, information and reports from previous work orders can be used, in which detailed descriptions of previous instances of damage are recorded. Data from external platforms or databases, such as insurance claims in which damage is described, can also be included. Furthermore, manuals or documentation of vehicles themselves can be analyzed to integrate technical terms and definitions, which helps the model in understanding specific technical language. Finally, online forums and discussion groups where damage repair is discussed can provide valuable textual information about many different instances of damage.

[0131] In the first step, these descriptions are collected and annotated, wherein important elements such as the nature of the damage, the severity and specific circumstances of the accident are recorded. These annotations can also contain contextual information, such as previous repair experiences with similar damage. Subsequently, the textual data is cleaned and prepared for analysis. This entails that irrelevant information and stop words are removed, and that the text is normalized, for example by using techniques such as tokenization and lemmatization. These steps ensure that the model can concentrate on the essential content of the descriptions.

[0132] The training of the NLP model is possibly performed using techniques such as word embeddings (for example Word2Vec or GloVe) and recurrent neural networks (RNNs) or transformers. During this phase, the model learns to recognize patterns and semantic relationships in the text, whereby it is able to understand contextual information and extract relevant aspects of damage descriptions. The model applies a process of feedforward and backpropagation, wherein it compares its predictions with the actual annotations and adjusts its parameters to improve the accuracy.

[0133] After the initial training, the model is validated with a separate set of textual data to test the performance. The accuracy of the predictions is evaluated with statistical metrics such as precision, recall and Fl-score, to assess how well the model is able to extract the relevant information from the descriptions. Based on the validation results, optimizations are performed, such as adjusting hyperparameters and refining the training dataset to increase the variability. This leads to a more robust model that is better able to provide accurate characterizations of damage based on the textual input.

[0134] The first Al model is retrained to specifically tune to a motor vehicle repair company by making use of historical data which is representative of the unique method of that company. For this purpose, old quotations, detailed work orders and specific cost overviews which characterize the company are collected. During the retraining, the model optimizes its parameters and algorithms on the basis of this specific data, so thatit learns from the work processes, cost structures and patterns which belong to the company. This new information comprises variables such as the prices of components and labor, which can differ strongly depending on the company and the region. Additionally, the methods of employees can vary; some companies, for example, prefer to replace components instead of repairing them, while others employ repair techniques. Also the presence of advanced equipment is of importance since companies with more modern technology can arrange their processes more efficiently. The use of specific materials, such as OEM components versus aftermarket options, as well as the time which employees need to perform certain actions, are crucial for the training of the model. Further, differences in personnel capacity, customer segments and warranty and service conditions can be of influence on the operational methods of the company. By integrating all this diverse and specific information, the Al model can learn from the unique characteristics and operational needs of the repair company, whereby it is able to generate more accurate quotations which are tuned to the reality of their daily activities.

[0135] To effectively characterize a new instance of damage, image information of the damage is used. The image information can comprise photos and / or videos which give a clear representation of the external damage to the motor vehicle, such as dents in the bodywork, scratches on the paint or broken lights. This visual data is essential for identifying the extent of the damage and determining the necessary repairs. In addition, a description can also be added: "The dent in the left door was caused by a collision with another car, wherein the impact has also led to a tear in the rubber seal." This description records not only the visible damage, but also details about the cause and possible internal damage which is not visible in the images.

[0136] The images and description of the damage can be easily uploaded via a graphical user interface (GUI) implemented on a mobile device such as a smartphone or tablet. Employees or owners of damaged motor vehicles can take photos of the damage to the vehicle with their mobile phone or tablet and send these directly via the upload function in the GUI. The interface offers user-friendly buttons to select or take photos, and a text box wherein a description of the damage can be entered, for example details about how the damage has occurred or possible internal problems which are not visible on the photos. The GUI is designed to work quickly and efficiently on mobile devices, so that users can document damage even on site and share this information real-time with the motor vehicle repair company. As a result, the damage repair process can be started faster.The second Al model analyzes the received images of an instance of damage in a structured process to accurately characterize the damage. First, the second Al model optimizes the image for further analysis. This can for example entail that the image is normalized, wherein the brightness and the contrast are adjusted to improve the visibility of damage. Subsequently, the Al model performs object detection, wherein specific components of the vehicle, such as doors, bumpers and windows, are identified. This can be realized with the aid of techniques such as Convolutional Neural Networks (CNNs), which are trained to recognize the contours and features of different vehicle components. Alternative methods such as YOLO (You Only Look Once) and SSD (Single Shot Multibox Detector) can also be applied to detect objects quickly and efficiently. For example, the model can detect and mark a door in an image to determine where damage has occurred. Subsequently, the Al model evaluates the specific damage to the identified components by capturing different aspects of the damage. This comprises analyzing the intensity, color and texture of the pixels in the relevant areas, whereby the model can determine whether the damage is superficial, deep or structural. For example, the model can determine that the right door has a dent 5 cm wide, but also that the paint is damaged, which may indicate a need for both dent repair and repainting. Additionally, the model can also identify internal damage, such as torn cabling or damaged lighting systems, by looking at unusual deviations or changes in the areas surrounding the damage.

[0137] The first Al model then establishes a repair process by analyzing the characterized damage from the second Al model and linking this to the historical repair data of the specific motor vehicle repair company. Suppose that the damage consists of a crack in the windscreen, deformation of the window frame and a slight displacement of the window seal. The model recognizes on the basis of the damage classification that the windscreen must be replaced, the frame must be aligned and the seal possibly must be reapplied.

[0138] Based on this detailed evaluation, the first Al model selects repair procedures which best match the available resources and working methods of the motor vehicle repair company. For example, if the company specializes in glass repairs with advanced equipment for accurately aligning window frames, the model takes this into account and chooses a procedure wherein the frame is aligned directly in the workshop. Additionally, the model calculates the required time for each of these steps on the basis of the work experience of the employees and previous repairs which the company has carried out. Finally, the Al model draws up a repair process with a clear step-by-step approach, including the estimated time and costs, which results in an efficient and personalized repair plan for the specific company.Possibly, in addition to an image of the damage, a description is also added. This description is analyzed by the NLP model by first tokenizing the text, wherein the words are divided into separate units such as damage categories and vehicle components. The model subsequently recognizes relevant terms and patterns, such as "dent in the front bumper" or "scratches on the right door", by making use of trained word vectors which capture the semantic relationship between words. The NLP model can also identify synonyms and technical jargon, such as "fracture" or "crack", and thereby classify the extent and severity of the damage. This analysis refines the overall characteristic of the instances of damage.

[0139] After the second Al model and the NLP model have characterized the damage, the first Al model can subsequently determine one or more repair processes on the basis thereof. There may be several methods suitable for repairing the damage, which differ from one another. For example, if there is a damaged wheel, the Al model may propose fully replacing the wheel as one option, or repairing the wheel using specialized equipment as an alternative option. Each process is adapted to the capacities and preferences of the company, wherein the costs and duration for the repair are carefully calculated. The quotations for each process are automatically generated by the Al model, wherein account is taken of the costs of components, such as the wheel or any fastening materials, and the labor costs that are necessary to carry out the specific repair process. For the replacement option, the quotation can for example turn out higher because of the price of a new wheel, while the repair option entails lower costs but possibly requires more labor. The model can also take into account factors such as the use of specialized equipment or the engagement of extra personnel.

[0140] All this information is displayed on a graphical user interface (GUI), which enables employees of the motor vehicle repair company to easily view the specific repair process and any alternative repair processes, together with the associated quotations. The GUI presents a repair process clearly, with details such as the materials that are necessary, the time the repair will take, and the estimated total costs. Employees can browse through any alternative repair processes and compare the options.

[0141] The GUI also offers the possibility to make adjustments in the proposed repair processes or quotations. If an employee decides for example that a wheel repair is feasible without the complete replacement of the wheel, they can implement this change via the GUI. Or possibly the price of necessary materials, the available materials, or the available equipment has changed without the first Al model being aware of this. It can also occur that the employee observes hidden damage during the repair as a result of which therepair process must be adapted. All these changes can be adjusted via the GUI by the employees and the first Al model subsequently adapts the quotation and the specific repair process automatically on the basis of this new information.

[0142] The first Al model compares generated quotations for verification with quotations that have been drawn up by the motor vehicle repair company, wherein the Al model recognizes any deviations. Herein parts of quotations are compared that relate to comparable damage, actions or materials. This means that the Al model analyzes specific elements, such as the costs for parts or the required labor time, in the context of previous repairs that correspond to the current damage. When a clear discrepancy is established, such as a significant deviation in the estimated costs or duration, the first Al model attempts to determine the cause of this deviation and to judge whether an error has been made. If the deviation is, for example, greater than a certain percentage or absolute difference that has been established based on historical data, such as for example 20% higher costs for labor or components compared to previous repairs of comparable damage, the model can mark this deviation as a potential error. Additionally, contextual information can also be used, such as the nature of the damage and the associated repair methods. If, for example, it encounters a situation wherein a simple scratch is marked with unrealistically high costs for repair, the first Al model can conclude that an error has likely been made. By employing such criteria, the Al model can effectively determine which deviations must be investigated further and which must be corrected as an error.

[0143] When an error has been made, the generated quotation is automatically adjusted. If, however, it is not immediately clear in the case of a significant deviation whether an error has been made, this discrepancy is displayed on the graphical user interface (GUI), so that an employee of the motor vehicle repair company can check it.

[0144] During the repair of the damaged motor vehicle, the determined repair process and the associated quotation can be automatically adjusted based on image data that is obtained in real time from the workshop of the motor vehicle repair company. This image data can be captured with cameras that are installed above the workplaces or with portable cameras, such as bodycams that are worn by employees. These cameras continuously register the course of the repairs and send the images to the Al model for analysis. An Al model, which is trained with an extensive dataset of labeled images of various repair scenarios, analyzes the obtained images to determine the progress of the repairs. The model can, for example, identify tools via object detection that are used by employees, such as screwdrivers, dent removal tools or paint sprayers, and can also recognize vehicle components such as doors, bumpers or wheels. By detecting whichtool is used and which component of the vehicle is being worked on, the Al model can determine which specific repair activities are being performed, such as removing a damaged door or removing dents from a bumper.

[0145] Additionally, the Al model can, by making use of pose estimation, follow and recognize the actions of employees, such as removing components, remounting components or preparing a vehicle for spraying paint. If the model detects, for example, that a bumper has been replaced instead of repaired, the quotation can be automatically adjusted to include the extra costs for the new component.

[0146] Further, the Al model can also recognize specific employees. The quotation can subsequently be further adjusted by taking into account the specific pace at which individual employees work. The model can, for example, detect that an experienced employee works faster than a less experienced colleague, which has an influence on the expected time for completing certain repairs. By linking the recognized employee to historical data regarding their performance, the Al model can make an accurate estimate of the required time and costs for the repair. If an employee who usually needs longer for certain actions is recognized, the quotation can be automatically adjusted to reflect extra time and labor costs.

[0147] Based on the information that is extracted from the image data by means of object detection, the first Al model can be further trained to be tuned even more accurately to the specific motor vehicle repair company. The first Al model thereby learns how different objects and actions have an influence on the repair processes and can use this data to make future quotations more accurate. If the Al model for example recognizes that a particular employee uses a replacement component instead of a repair method, it can learn that this method generally entails higher material costs and fewer labor hours. Additionally, the model can, by analyzing how different employees perform specific repairs, learn which actions require more time or costs, depending on the skills and experience of the involved employees. By adding this information to the existing database of the first Al model, estimates for time and costs can be further improved. Furthermore, the first Al model can be trained with data regarding unexpected complications which emerge from the image data, such as internal damage which only becomes visible during the repairs. The model can then learn to anticipate such situations in future predictions.

[0148] This continuous feedback loop, wherein the model continuously collects new data and learns from the images of the repair process, ensures that the first Al model is increasingly accurately tuned to the unique processes, personnel and preferences of the specific motor vehicle repair company. As a result, it can generate increasingly realistic quotations which not only take into account standard repair processes, but also the specific characteristics and dynamics of the company.All information regarding the progress of the repair, obtained from the image data of the activities, is displayed on a graphical user interface (GUI) so that employees of the motor vehicle repair company can easily remain informed of the status of the repair. This information can comprise, inter alia, which specific repairs are currently being performed, the time which has been spent on each task, and which components have been used or replaced. For example, the GUI can indicate that the right door is currently being undented and that the process is expected to take another 30 minutes, while it also gives a warning if extra damage is detected. This transparency promotes a better collaboration among employees, because they have direct insight into each other's activities and can respond faster if problems occur or if there is miscommunication. As a result, the overall repair efficiency can be improved, which leads to a faster completion of the work and a higher customer satisfaction.

[0149] During the repairs, various photos of the motor vehicle can be taken during the process to record the progress of the activities. These photos can be taken at important moments, such as before the repair begins, during important steps such as the removal of damaged components, and after the completion of the repairs. This visual documentation provides a clear overview of the progress and can bring any complications or changes in the damage to light.

[0150] The photos taken can be easily uploaded via the graphical user interface (GUI). As a result, employees of the repair company can immediately share the images with external experts, such as insurance advisors or loss adjusters, or with the owner of the motor vehicle itself. This makes it possible to maintain transparent communication about the status of the repair and any necessary adjustments which result from new findings. For example, if during the repair it appears that additional damage has been detected, photos can be uploaded to visually substantiate the situation, whereby both the customer and external parties remain well informed about the progress and any changes in the repair process.

[0151] The first Al model can improve the accuracy of the quotation on the basis of these photos. By analyzing the images, the Al model can collect additional data regarding the performed repairs and the used materials, and integrate this information into its estimates. This ensures that the quotation is better aligned with the reality of the repair.

[0152] Additionally, these photos can also be used to further train the second Al model and / or the first Al model. By labeling and annotating the images with information about the performed repairs and the associated damage, the Al models can characterize damage even more accurately and determine a repair process, because they continuously learnmore unique instances of damage and unique operational characteristics of the motor vehicle repair company.

[0153] EXAMPLE 2

[0154] The present invention relates to a method for accurately estimating the costs of repairing a damaged vehicle and generating a detailed invoice. The method starts with retrieving current damage information via direct communication with external systems via Application Programming Interface (API) endpoints. The API query system is configured to integrate with databases to obtain relevant data about the damaged car, including historical maintenance data, technical specifications and previous repair details. This streamlined query system ensures that the most accurate and current information about the damaged vehicle is retrieved, whereby the subsequent steps can be aligned with the specific context of the damage.

[0155] Once the data has been retrieved, analytical approaches with advanced computer vision capabilities are applied to segment the vehicle components in the provided images of the damaged vehicle. The system isolates individual car components, such as doors, bumpers or windows, for a detailed analysis of the specific damage to each component. Segmentation enables the system to identify the precise location and extent of the damage, whereby the granularity of the assessment is improved.

[0156] The segmented images are subsequently prepared for further analysis in subsequent steps. The segmented car components are analyzed using a combination of vision language models and expert systems, which perform a comparative analysis between the current images and historical data of similar vehicles. These analytical approaches use input from the segmented images and the historical data to assess the damage to individual components. By comparing current damage patterns with reference data from the extensive database, the hybrid analytical system generates a detailed assessment of the type, the severity and the possible repair methods for each damaged component.

[0157] To support the comparative analysis, the system relies on an extensive database with data of both damaged and undamaged vehicles, repair histories and cost estimates. This database is organized to facilitate rapid access to relevant historical data, which is used to compare the current damage with previous repair scenarios. The database ensures that the system has access to accurate reference points for cost estimates, wherein previous repair results are utilized to improve the reliability of the predictions.The results of the analytical approaches are input into a processing architecture with real-time and cloud-based implementations for cost estimations. This architecture processes input features derived from the hybrid analysis and historical comparisons, wherein advanced price calculation algorithms are applied to determine a detailed cost breakdown. The model takes into account labor, parts and extra service costs and produces an accurate estimate tailored to the specific damage and repair requirements. The output comprises a complete cost analysis, which offers transparency and accuracy for both repair companies and customers.

[0158] The final step comprises generating a structured invoice using expert systems trained for professional documentation. Based on the output of the distributed processing options, the system compiles an itemized invoice with a specification of the repair costs, details of the damaged components and a summary of the estimated total price. The invoice is formatted to comply with industry standards and can be adapted with additional information, such as warranty conditions or service conditions. This document serves as a professional, customer-oriented report of the repair estimate.

[0159] Throughout the entire process, the various components of the system can optionally be integrated into a uniform workflow. Data retrieved via the API query system flows seamlessly to the segmentation and analysis phases, while results from the VLM and machine learning models are dynamically incorporated into the invoice generation step. This integration ensures efficiency, accuracy and a streamlined user experience for repair companies and their customers. The combination of real-time data retrieval, component-specific damage assessment, historical data integration, machine learningbased cost estimation and automated invoice generation of the invention creates a robust framework for modernizing vehicle repair processes. By automating and optimizing these steps, the system improves the precision, speed and transparency of repair cost estimations and documentation.

[0160] EXAMPLE 3

[0161] The present invention relates to a method and system for implementing multiple analytical approaches and processing architectures within a defined operational environment, specifically tailored to analyzing activities and interactions relating to employees and vehicles. This invention integrates various analytical methodologies and distributed processing options to enable robust and efficient system performance, even in environments with limited resources.The system starts with implementing real-time processing capabilities for zone monitoring via analytical approaches, which define operational zones and their boundaries. This analytical framework lays the foundation for subsequent analysis, enabling the system to focus on specific areas within the environment. Within these zones, expert systems and hybrid approaches identify persons present in the operational regions, while advanced computer vision techniques analyze the positions of the detected employees to classify their activities and monitor their movement.

[0162] Simultaneously, vision language models and advanced computer vision approaches identify vehicles within the monitored zones, which forms the basis for further analysis or segmentation. Vehicle identification is performed using hybrid approaches comprising multiple analytical methodologies, including advanced pattern recognition systems, to extract alphanumeric characters for vehicle identification. The system also comprises distributed processing options for numerical feature recognition and training functionality, thereby allowing it to recognize and process specific numerical features, such as vehicle IDs or specific markings, with high accuracy.

[0163] For detailed analysis, identification of vehicle components is performed using a combination of vision language models and expert systems. This allows the system to isolate and identify specific vehicle components, whereby the accuracy of subsequent evaluations, such as damage assessments, is improved. Furthermore, hybrid analytical approaches with multiple processing architectures further refine the analysis by interpreting the key points and positions of detected entities, thereby facilitating the classification of movements and activities within the zones.

[0164] To maintain continuity in the analysis, the system implements cloud-based implementations and edge computing solutions which monitor the movement of both persons and vehicles in the analyzed zones. This ensures a comprehensive understanding of interactions and activities over time. To optimize the system for implementation in resource-constrained environments, the processing architectures are adapted into efficient implementations, including both batch processing systems and distributed processing options. These implementations enable efficient performance, while the accuracy and functionality of the comprehensive analytical approaches are maintained. This makes the system suitable for diverse operational scenarios.

[0165] EXAMPLE 4The user drives his motor vehicle onto the site of a publicly accessible photo tunnel, for example in the car park of a supermarket or next to the entrance of a gas station. At the entrance to the tunnel stands a clearly recognizable portal with a marking on the road surface indicating where and at what speed to enter. As soon as the vehicle approaches the zone, the user is informed via a screen or a short notification in a mobile application that an inspection will take place automatically and that the results will be delivered as a report. The user drives through slowly, for example at walking pace, wherein the tunnel is designed such that the vehicle is automatically guided into the correct position. At the moment that the vehicle is located in the tunnel, sensors activate the image acquisition. Around and above the vehicle, images are captured simultaneously from multiple angles, so that the front, rear, sides, roofline, underside and wheel zones are visible. Depending on the embodiment, detail images are also captured of critical zones such as bumpers, headlights, mirrors and wheel arches, wherein the tunnel guarantees consistent lighting and camera angle due to its fixed setup.

[0166] While the user continues driving, the system immediately begins with vehicle identification. It recognizes the vehicle type and model on the basis of external features and, if visible, also on the basis of license plate information. This identification is used to compare the recorded images with reference images of undamaged vehicles of the same type. Subsequently, the second Al model automatically performs a damage analysis. The system detects deviations with respect to the reference model, locates these on specific vehicle zones and links the detected instances of damage to the relevant motor vehicle components. In the same pass, without the user having to perform extra actions, the wheels and tires are also imaged. The second Al model analyzes the tire condition by examining visual characteristics of wear and possible damage, and can distinguish uneven wear patterns which may indicate incorrect tire pressure, alignment or other technical problems.

[0167] As soon as the user leaves the tunnel and can drive his vehicle at normal speed again, the first processing has often already started. In the background, the encoded damage information is transmitted to the first Al model, which generates separate cost estimates for multiple motor vehicle repair companies. Because the first Al model is trained or further trained per repairer on the basis of historical quotations and company-specific repair data, the system can produce different quotation outcomes for the same instance of damage which align with the method, material costs, labor duration estimation and price structure of each involved repair company. For the user, this results in multiple concrete repair options which contain not only a total price, but also an indication of theexpected repair time and the planning. The system subsequently combines this cost information with additional selection parameters. It determines, for example, which repairers are available in the short term, how long the waiting time is expected to be, and what the travel time or distance is between the location of the user and the repairer. If an insurance company or claims handler is involved, a quality score or preference score is furthermore taken into account which may for example be based on previous repair quality, turnaround time, customer satisfaction or contractual agreements. On the basis of this combination of cost, time, distance and quality, the user receives not only separate quotations, but also a recommendation which explains which repairer is considered most suitable in this case.

[0168] Parallel to the damage and quotation analysis, the system performs, on the basis of the same tunnel images, a conformity analysis and safety assessment. Thereby, it is assessed whether the motor vehicle complies with predefined technical or administrative requirements, such as visible conformity of lighting, mirrors and bodywork parts, and whether there are indications that a periodic inspection or maintenance interval is approaching. The safety assessment may for example signal that a headlamp unit is damaged, that a mirror is missing or that tire tread appears to be below a threshold value. When the system detects such urgent safety-relevant points, it can automatically mark these as priority, so that the user immediately understands which repairs cannot wait. Depending on the configuration, the system can also share these urgent findings with a maintenance provider or insurance company, so that for example a claims procedure can be started more quickly or so that the user receives guidance regarding a safe next step.

[0169] Shortly after the passage through the tunnel, the user receives an integrated report, for example via a notification in an app, an email or a web portal. In that report, the user sees a summary of what was established by the tunnel: the recognized vehicle identity, an overview of the detected damage with mention of the location on the vehicle, and the associated repair options with cost estimates for multiple repair companies. The user can thereby also see indicative repair dates, so that it is clear which repairer can perform the repair quickly. Additionally, the report contains the results of the tire check, any findings from the conformity analysis and a safety indication which makes clear whether immediate attention is required. If the system can deduce, on the basis of image data, historical vehicle data or maintenance intervals, that a maintenance service or periodic inspection is necessary within the foreseeable future, then this is also included in the report. In that case, the user can furthermore receive proposals for different maintenance companies with an estimated maintenance cost, the availabilityand the proximity, personalized per company in an analogous manner as with the repair quotations. The user thus ends the process without a manual inspection or telephone inquiry, but with a concrete, detailed and comparable overview of both repairs and any maintenance needs, including practical choices about where, when and on what conditions this can be carried out.

[0170] In certain situations, if the damage is covered by insurance, the report mentions no cost estimate for the user, and only, for example, the nature and location of the damage, the affected vehicle components, the proposed repair method and suitable repair companies. The cost information is in that case shared separately with the corresponding insurance company. If the damage is not covered by insurance, the generated cost estimate or quotation is included in the report to the owner of the vehicle.

[0171] In a further embodiment, the report is linked to a digital environment, such as a mobile application or a web platform, wherein the user can consult the full report by means of a unique reference code which is for example automatically generated upon passing through the photo tunnel. Via this application or website, the user can directly, without additional communication, schedule one or more appointments with the companies concerned that are mentioned in the report, such as motor vehicle repair companies, maintenance companies, inspection bodies or other service providers. When the user confirms an appointment, the relevant information from the report is automatically and securely forwarded to the selected company, including the identified damage, the proposed repair or maintenance process, the estimated time duration, the required parts and any safety or priority indications. As a result, the receiving company has all necessary data at its disposal beforehand and no additional consultation, inspection or re-evaluation needs to take place. In cases where the report contains multiple necessary actions, the user can, if desired, make separate appointments at different companies, for example one appointment for damage repair and another for maintenance or inspection, wherein each company involved automatically receives only that information which is relevant to the respective action. For the user, this results in a greatly simplified process wherein after scheduling the appointments only the vehicle still needs to be dropped off at the agreed time, while all administrative and substantive preparation has already been handled beforehand.

Claims

1. CLAIMS1. A method for generating a quotation for the repair of damage to a motor vehicle comprising the steps of:- training a first Al model with repair data originating from different motor vehicle repair companies, wherein the repair data comprises historical information relating, inter alia, to repair procedures, costs, and time required for repairing different types of damage to motor vehicles, preferably at least partly derived from previous quotations of the different motor vehicle repair companies;- further training the first Al model with repair data originating from a specific motor vehicle repair company, wherein the repair data comprises historical information, inter alia, about repair procedures, costs, and required time for repairing different types of damage to motor vehicles, preferably in the form of old quotations drawn up by the specific motor vehicle repair company; receiving images of a damaged motor vehicle;characterizing one or more instances of damage to the damaged motor vehicle on the basis of the received images of the damaged motor vehicle using a second Al model trained with data of undamaged and damaged motor vehicles, wherein the second Al model, inter alia, undertakes the following steps: recognizing the model of the motor vehicle from the images, comparing the images of the damaged motor vehicle with images of undamaged motor vehicles of the same type from a database and identifying and locating instances of damage on the basis of the comparison, identifying and locating different motor vehicle components, and associating the instances of damage and the motor vehicle components on the basis of the identifications and the determined locations;- determining at least one repair process for the specific motor vehicle repair company using the first Al model, wherein the determination is based on the characterized one or more instances of damage and the repair data originating from motor vehicle repair companies, and wherein the determined repair process is suitable for repairing the one or more characterized instances of damage;generating one or more quotations for the specific motor vehicle repair company with the first Al model based on the determined repair process.

2. The method according to any of the preceding claims, further comprising the step of displaying information on a graphical user interface (GUI), wherein the information comprises at least one, or a combination of the following elements: the characterized one or more instances of damage, at least one generated quotation, a duration of the determined repair process and / or a duration of separate sub-processes within the determined repair process.

3. The method according to claim 2, wherein the information comprises adjustments in the determined repair process and / or in the one or more generated quotations, and wherein the method further comprises the steps of asking for a permission for said adjustments via the graphical user interface, and automatically adjusting said quotations on the basis of the granted or refused permission.

4. The method according to any of the preceding claims, further comprising the steps of repairing the damaged motor vehicle according to a determined repair process, obtaining image data of the repairs, automatically recognizing objects in the image data and automatically updating the generated quotation on the basis of the recognized objects and / or a time at which the objects are recognized, the recognized objects comprising one of, or a combination of, the following elements: employees, tools, motor vehicles, motor vehicle components.

5. The method according to claim 4, further comprising the step of displaying information on a graphical user interface (GUI), wherein the information comprises at least one of, or a combination of, the following elements: the progress of the repairs in real time, adjustments in the determined repair process, changes in a generated quotation; and wherein the information is based on the recognized objects.

6. The method according to claim 5, wherein the information of the progress of the repairs comprises photos of the motor vehicle that is being repaired.

7. The method according to claim 6, wherein the said photos are used to train the second Al model.

8. The method according to claims 4 to 7, further comprising the step of training the first Al model based on the recognized objects and / or the time at which the objects are recognized.

9. The method according to any of the preceding claims, further comprising the step of providing an upload function integrated into a graphical user interface (GUI), wherein the upload function is configured to upload data relating to the one or more instances of damage, and wherein, during the step of receiving data from a damaged motor vehicle, said data is received via said upload function, and wherein the graphical user interface (GUI) is suitable for implementation on a mobile device, preferably a mobile phone and / or tablet.

10. The method according to any of the preceding claims, further comprising the step of receiving a description of the one or more instances of damage, and wherein the step of characterizing the one or more instances of damage occurs based on the said description using a Natural Language Processing (NLP) model.

11. The method according to any of the preceding claims wherein the first Al model and the second Al model are integrated into one Al model.

12. The method according to any of the preceding claims, further comprising comparing the generated quotation, or a part thereof, with a quotation, or a part thereof, drawn up by the motor vehicle repair company, wherein the compared quotations, or parts thereof, relate to similar repair processes or similar subprocesses within repair processes, and adjusting the first Al model and / or approving or adjusting the generated quotation, or part thereof, on the basis of the comparison.

13. The method according to any of the preceding claims, further comprising the step of adjusting a determined repair process on the basis of information provided by one or more employees of the motor vehicle repair company, wherein the information relates to a change in the determined repair process and / or a change in the characterized one or more instances of damage.

14. The method according to any of the preceding claims, further comprising the steps of repairing the damaged motor vehicle according to a determined repair process, and training the first Al model on the basis of the information provided by one or more employees of the motor vehicle repair companies, wherein the provided information relates to the repairs.

Images