WO2021073213A1

WO2021073213A1 - Method and apparatus for generating loss assessment file for damaged vehicle

Info

Publication number: WO2021073213A1
Application number: PCT/CN2020/107934
Authority: WO
Inventors: 蒋晨; 程远; 王清
Original assignee: 支付宝(杭州)信息技术有限公司
Priority date: 2019-10-14
Filing date: 2020-08-07
Publication date: 2021-04-22
Also published as: CN110738576A; CN110738576B

Abstract

A method and apparatus for generating a loss assessment file for a damaged vehicle. The method comprises: firstly, acquiring a damaged part marking result and a damaged degree marking result determined on the basis of a field video of a damaged vehicle, wherein the damaged part marking result indicates a damaged part and a damaged area, and the damaged degree marking result is the result of identifying the damaged degree of each damaged part on the basis of the damaged part marking result; then, at least inputting the damaged part marking result and the damaged degree marking result into a pre-trained loss assessment model, wherein the loss assessment model is trained on the basis of historical loss assessment data of a plurality of loss assessment personnel and/or historical maintenance data of a plurality of maintenance personnel and is used for generating a damage maintenance scheme; and then generating a loss assessment file for the damaged vehicle on the basis of an output result of the loss assessment model. In this way, the effectiveness of vehicle loss assessment can be improved.

Description

Method and device for generating damage assessment file for damaged vehicle

Technical field

One or more embodiments of this specification relate to the field of computer technology, and in particular to a method and device for generating damage assessment files for damaged vehicles through a computer.

Background technique

In the traditional car insurance inspection scene, the car is often inspected by professional investigators of the insurance company. For example, when applying for insurance, it is necessary to check whether the vehicle is damaged. In the scene of auto insurance claims, the insurance company needs to send professional damage survey personnel to the scene of the accident to conduct on-site damage survey. Due to the need for manual damage assessment, insurance companies need to invest a lot of labor costs and professional knowledge training costs. From the experience of ordinary users, due to waiting for on-site inspection by manual surveyors, the user's waiting time is longer and the experience is poor. From the perspective of an insurance company, the surveying personnel require high professionalism. A surveying personnel needs to have the professional skills of component recognition, damage degree recognition, and determination of maintenance plans based on components and damage degrees. Moreover, they often have a wealth of definitions. Only the investigators with loss experience and professional knowledge can quickly produce high-quality loss assessment documents, and the training cost of such investigators is relatively high.

In view of the above-mentioned problems, artificial intelligence and machine learning can be applied to the scene of vehicle damage detection. It is hoped that computer vision image recognition technology in the field of artificial intelligence can be used to automatically identify the images reflected in the pictures based on the on-site images taken by ordinary users. Car damage condition. In this way, labor costs can be greatly reduced, and user experience can be improved.

Summary of the invention

One or more embodiments of this specification describe a method and device for generating damage assessment files for damaged vehicles, which can combine computer vision image recognition technology with manual annotation to provide high-quality damage recognition results and give more effective results. The loss assessment file.

According to a first aspect, there is provided a method for generating a damage assessment file for a damaged vehicle, the method comprising: obtaining a damaged component marking result and a damage degree marking result determined based on a live video of the damaged vehicle, wherein, The marking result of the damaged part indicates the damaged part and the damaged area, and the marking result of the damage degree is based on the marking result of the damaged part, the result of identifying the degree of damage of each damaged part; at least the damaged part is The labeling result and the damage degree labeling result are input into a pre-trained damage assessment model, wherein the damage assessment model is trained based on the historical damage assessment data of multiple damage assessment personnel and/or the historical maintenance data of multiple maintenance personnel, Used to generate a damage repair plan; based on the output result of the damage damage model, generate a damage file for the damaged vehicle.

In one embodiment, the marking result of the damaged part is determined by: acquiring a first image frame marked with the first damaged part and the first damaged area in the live video; The image frame is in adjacent image frames within a predetermined number of frames, and it is detected whether there is a second image frame, and there is a second damaged area in the second image frame, and the second damaged area is the same as the first damaged area. Area correlation; in the case where the second image frame exists, the labeling result of the second image frame is determined, and the labeling result of the second image frame corresponds to the first damaged part and the second damaged area.

In one embodiment, the damage degree annotation result is determined by: inputting the live video and the damaged component annotation result into a pre-trained damage degree prediction model; according to the damage degree prediction model The output result determines the marking result of the damage degree.

In one embodiment, the inputting at least the damaged component labeling result and the damaged degree labeling result into a pre-trained damage assessment model includes: marking the damaged component result and the damaged degree labeling result , And inputting the fixed loss model based on the vehicle type information determined based on the live video.

In an embodiment, the output result of the damage assessment model includes a damage repair plan determined for the damaged vehicle; and the output result of the damage assessment model is based on the output result of the damage assessment model to generate a damage assessment file for the damaged vehicle The method includes: obtaining the output result of the damage assessment model and the mapping relationship between a predetermined damage maintenance plan and a predetermined resource; determining the resource share corresponding to the damage maintenance plan determined for the damaged vehicle based on the mapping relationship; The damage repair plan determined for the damaged vehicle and its corresponding resource share are combined to generate a damage assessment file for the damaged vehicle.

In an embodiment, the method further includes: obtaining a detection result of the damage assessment personnel on whether the damage assessment file generated for the damaged vehicle is correct; in the case that the detection result is wrong, obtaining the corrected assessment result A loss file; a new sample is generated for the loss setting model according to the revised loss setting file; the new sample is used to update the loss setting model.

In one embodiment, the damage repair plan includes at least one of painting, sheet metal, replacement, polishing, machine repair, external repair, disassembly and assembly, and electrician.

According to a second aspect, there is provided a device for generating a damage assessment file for a damaged vehicle, the device comprising: an obtaining unit configured to obtain a damaged component marking result and a damage degree marking determined based on a live video of the damaged vehicle As a result, wherein the marking result of the damaged part indicates the damaged part and the damaged area, and the marking result of the damage degree is based on the marking result of the damaged part, the result of identifying the degree of damage of each damaged component; , Configured to input at least the damaged component labeling result and the damaged degree labeling result into a pre-trained damage assessment model, wherein the damage assessment model is based on historical damage assessment data and/or multiple damage assessment personnel. The historical maintenance data of each maintenance personnel is trained to generate a damage maintenance plan; the generating unit is configured to generate a damage assessment file for the damaged vehicle based on the output result of the damage assessment model.

According to a third aspect, there is provided a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of the first aspect.

According to a fourth aspect, there is provided a computing device, including a memory and a processor, characterized in that executable code is stored in the memory, and when the processor executes the executable code, the method of the first aspect is implemented .

Through the method and device for generating a damage assessment file for a damaged vehicle provided by the embodiments of this specification, in the process of generating a damage assessment file for a damaged vehicle, the damaged component marking result and the damage degree marking result are used as the characteristics of the damage assessment model. Modules that are independent of the damage assessment model can be used for labeling and training, so as to disperse the work of labelers and reduce the overall professionalism of labelers. In the damage model training stage, the historical damage assessment data of multiple damage assessment personnel and/or the historical maintenance data of multiple maintenance personnel are used to diversify the damage assessment model samples, eliminate the influence of personal habits or experience, and improve the damage assessment model The effectiveness of the results, thereby generating high-quality loss-making files. In short, the embodiments of this specification can improve the effectiveness of vehicle damage assessment.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present invention more clearly, the following will briefly introduce the drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative work, other drawings can be obtained from these drawings.

Figure 1 shows a schematic diagram of an implementation scenario of an embodiment disclosed in this specification;

Fig. 2 shows a flowchart of a method for generating a damage assessment file for a damaged vehicle according to an embodiment;

Figure 3 shows a schematic diagram of a specific example of damaged parts marking;

FIG. 4 shows a schematic diagram of marking the degree of damage of a specific example;

Fig. 5 shows a specific example of a flow chart of generating a fixed-loss file based on a fixed-loss model and optimization of the fixed-loss model;

Fig. 6 shows a schematic block diagram of an apparatus for generating a damage file for a damaged vehicle according to an embodiment.

Detailed ways

The following describes the solutions provided in this specification with reference to the accompanying drawings.

For ease of description, the description will be made with reference to a specific application scenario of the embodiment of this specification shown in FIG. Figure 1 shows a vehicle inspection scene, the purpose is: to generate a damage file for the vehicle. The damage assessment file can be used to provide users with maintenance suggestions, or to provide insurance compensation reference for insurance companies, or for maintenance reference at vehicle repair points. The damage assessment document may include a repair plan and/or repair costs, and may also include at least one of damaged parts, damage types, damaged materials, and so on.

In this implementation scenario, the user can collect on-site video of the vehicle through a terminal that can collect on-site information, such as a smart phone, a camera, and a sensor. The user can send the live video to the loss assessment platform through the terminal. The damage assessment platform may include three modules: a damaged component labeling module, a damaged degree labeling module, and a damage assessment module. The three modules in the loss assessment platform can be integrated or independent of each other.

Among them, the damaged parts marking module uses live video to mark damaged vehicles, such as marking damaged parts and damaged areas. Damaged parts and damaged areas are described by marking the results of damaged parts. The module may be an independent module executed by a computer, or it may include an interactive module executed by humans interacting with a computer.

The damage degree labeling module can further determine the damage degree of the damaged component based on the damaged component labeling result of the live video. The degree of damage of the damaged component is described by the result of the degree of damage.

The damage assessment module can store a pre-trained damage assessment model, and process the above-mentioned damaged component marking results and damage degree marking results through the damage assessment model to determine a repair plan for the damaged vehicle, and generate a loss assessment file according to the repair plan.

In the implementation scenario shown in FIG. 1, the loss-fixing platform can be set on the server side that provides support for the terminal's loss-fixing application, but it does not rule out the possibility of setting the loss-fixing platform on other devices. The three modules in the loss-assessment platform can also be separately installed in different devices, which is not limited in this manual. In the embodiments of this specification, the loss determination process is described with the loss determination module as the main body.

It is understandable that when generating the damage assessment file, it is necessary to combine the damage repair plan. Different professionals may also have different maintenance preferences, and may also give different maintenance plans for the same damage. Therefore, under the technical concept of this manual, in order to improve the quality of the damage assessment document, it is necessary to avoid the bias of the maintenance plan caused by personal color as much as possible, and give pertinent suggestions, such as not increasing the insurance company's maintenance compensation cost due to unnecessary replacements. , Do not cause loss of car owners due to incomplete maintenance, etc., for the training of the damage model, use training samples of multi-party data.

In addition, the separate setting of the three modules in the loss assessment platform can greatly reduce the difficulty of manual labeling during the actual loss assessment of samples or vehicles. For example, if the damaged parts are marked with modules, the marking staff only needs to identify the corresponding parts, and the corresponding parts can be identified at the obvious damaged parts. The damaged parts are marked with the module, and the marking staff only needs to grasp the degree of damage of certain damaged parts. It greatly reduces the professionalism and experience requirements of the loss assessment personnel, saves the cost of manpower training for the loss assessment business side, and improves efficiency.

In short, under the implementation framework of this manual, the damaged component marking result and the damaged degree marking result are used together as the characteristics of the damage assessment model. The damaged component and its degree of damage jointly determine its repair plan, the damaged component marking result and the damage The degree marking result can be carried out in a module independent of the damage assessment model, which can decentralize the work of the marking staff and reduce the comprehensive professionalism of the marking staff. On the other hand, in the training phase of the damage model, the historical damage assessment data of multiple damage assessors and/or the historical repair data of multiple maintenance personnel are used, combined with the diversification of samples, to eliminate the influence of personal preference and improve the damage assessment model. The effectiveness of the results, thereby generating high-quality loss-making files.

Fig. 2 shows a process of generating a damage assessment file for a damaged vehicle according to an embodiment. The execution subject of the method can be any system, equipment, device, platform or server with computing and processing capabilities. For example, the computing platform shown in Figure 1. As shown in FIG. 2, the method for generating a damage assessment file for a damaged vehicle may include the following steps: Step 201: Obtain the damaged component marking result and the damage degree marking result determined based on the live video of the damaged vehicle, wherein the damaged vehicle The component labeling result includes the damaged component name and the damaged frame. The damage frame is used to enclose the damaged area. The damage degree labeling result is based on the damaged component labeling result and the damage degree identification result of each damaged component; step 202, at least The labeling results of damaged parts and the degree of damage are input into the pre-trained damage assessment model, where the damage assessment model is trained based on the historical damage assessment data of multiple damage assessors and/or the historical maintenance data of multiple maintenance personnel. To generate a damage repair plan; step 203, based on the output result of the damage model, generate a damage file for the damaged vehicle.

First, in step 201, the marking result of the damaged component and the marking result of the damage degree determined based on the live video of the damaged vehicle are obtained. The marking result of the damaged component and the marking result of the damage degree may be pre-stored, or may be marked based on the on-site video of the damaged vehicle in the execution subject of the process. On-site video of the damaged vehicle can be collected by the user using on-site terminals (such as smart phones, cameras, etc.). When the execution body of the process is a server that provides support for loss-making applications, the live video can be remotely transmitted to the above-mentioned execution body by the user through a terminal device running the loss-making application.

Here, the result of marking damaged parts can indicate damaged parts and damaged areas. For example, the damaged label result contains the name of the damaged part, indicating the damaged part, and the damaged frame, which encloses the damaged area. For example, the marking results of damaged parts include: the front bumper, slightly scratched; the left front light, slightly cracked. The determination of damaged components can be based on a pre-trained damage detection model based on video streams, can also be determined by manual labeling, or can be determined by a combination of manual and computer.

The following takes the combination of manual and computer as an example to illustrate the specific process of marking damaged parts in detail.

Please refer to Figure 3. First, the person who marked the damaged parts can manually watch the live video. When an image frame that is obviously damaged (hereinafter referred to as the obvious lossy frame) is found, mark the damaged part of the obviously damaged frame. Parts and damaged areas. Due to the continuity of video playback, there are lossy areas in multiple consecutive frames, and the marking staff can select one of the frames for marking. After that, the computer can automatically fill in the damage recognition results of multiple frames before and after the obvious lossy frame through a predetermined model. Among them, the predetermined model can include algorithms such as image similarity (pixel similarity, area connectivity, etc.). When an image frame contains an area that is related (similar or connected) to the damaged area marked by the marker, Mark damaged parts and damaged areas on the image frame. The computer can also apply the artificially labeled damaged parts to the frames before and after the obvious lossy frame in an adaptive manner.

For example, the first image frame is a manually marked obviously damaged frame, and is marked with the first damaged part "front bumper" and the damage frame surrounding the first damaged area. The computer can pre-order before and after the first image frame. Among adjacent frames within the number of frames (such as k), it is detected whether there is a second image frame that includes a second damaged area associated with the first damaged area. If it exists, determine the annotation result of the second image frame. It can be understood that the annotation result of the second image frame corresponds to the first damaged component and the second damaged area. For example, when the predetermined number of frames is 2, supplementary annotations are performed on the two image frames before and after the first image frame. The second damaged area associated with the first damaged area can be similar to the first damaged area (for example, the similarity is higher than a predetermined threshold), or it can be an area that overlaps with the first damaged area, or It is an area that is consistent with the edge of the first area and can be connected into one area.

As shown in Figure 3, in some embodiments, it is also possible to manually check whether the damaged component and the damaged area are correct in the annotation result automatically supplemented by the computer. If there is an error, the error frame can be corrected, and the correction method can include Correction and deletion. Specifically, if an image frame without damage is labeled, the label of the frame can be deleted, and if there is an error in the damaged labeling area, the position of the labeling frame can be manually corrected to make it correctly enclose the damaged area.

Then, manually continue to detect whether there are image frames that obviously contain obvious damage (obviously damaged frames). If there is no new obvious lossy frame, confirm that the labeling is complete, and output the labeling results, such as the names of all damaged parts, and the corresponding image frames that mark the damaged area. If there are new obviously lossy frames, repeat the above process.

It is understandable that the combination of manual and computer labeling is suitable for damage labeling when the labeling model is not accurate enough. In this process, only the damaged parts are marked. For the marking personnel, only the damaged parts need to be marked. At least one damaged part only needs to mark the obvious damage area on the image frame, and there is no need to accurately mark all the parts at one time. The damaged parts and the degree of damage, therefore, the professional and experience requirements of the labeling personnel are lower, and the training cost is also lower. In addition, for a damaged part, only one of the image frames may be marked, and the other marking processes are automatically supplemented by the computer, which greatly improves the marking efficiency. Further, when the accuracy of the computer marking result meets the predetermined condition, the above marking process can be completely realized by the computer without manual participation, which further improves the marking efficiency.

Further, as shown in FIG. 4, it can be determined based on the marking result of the damaged component, and the identification result of the damage degree of each damaged component is obtained, and the corresponding damage degree marking result is obtained. For example, the damaged part is marked with "left front bumper", and the degree of damage can be further marked with "slight scratch". The degree of damage can be related to the damaged component and/or the damaged area, so it can be marked in conjunction with the damaged component and the damaged area.

This part can be completed by computer or manually. In the case of being completed by a computer, the computer can pre-store a trained damage prediction model that recognizes the damage degree according to the damaged component and the damaged area. A training sample of the damage prediction model can correspond to one or more pictures marking a certain damaged part and damaged area, and a damage degree label. The recognition model can be a convolutional neural network CNN, etc., which is not limited in this specification. The damage degree label may include, for example, slight scratches, severe scratches, slight cracks, chipping, slight deformation, severe deformation, and so on. Input the image frame marked by the damaged part and the name of the damaged part into the trained damage prediction model, and the model outputs the classification result of the damage degree. The damage degree label corresponding to the classification result can be used as the damage of the corresponding damaged component. Mark the result by the degree. In the case that the damage degree marking is done by computer, manual participation is not required, which reduces the labor cost of related business parties (such as insurance companies).

Those skilled in the art can understand that the damage degree prediction model may be a model used to identify the degree of damage of various components, or it may include multiple models, and each model only recognizes the degree of damage of one or a few components. When the damage degree prediction model includes multiple models, it is also possible to determine which model is used to predict the damage degree of the damaged component according to the damaged component name in the damaged component labeling result.

In the case that the degree of damage is marked manually, since the labeling personnel only pays attention to the degree of damage, and does not need to mark the parts and damaged areas, and does not need to pay attention to the follow-up procedures such as maintenance plans, the professional degree requirements are also greatly reduced, and accordingly , The training cost of related business parties (such as insurance companies) is much lower than that of traditional loss determination methods.

In step 201, the execution subject may obtain pre-stored damaged component marking results and damage degree marking results locally, or may obtain damaged component marking results and damage degree marking results remotely in real time, which is not limited herein.

Next, in step 202, at least the above-mentioned damaged component labeling result and damage degree labeling result are input into a pre-trained damage model. Wherein, the damage assessment model is trained based on the historical damage assessment data of multiple damage assessment personnel and/or the historical maintenance data of multiple maintenance personnel, and is used to generate a damage repair plan. The maintenance plan here includes, but is not limited to, one or more types of painting, sheet metal, replacement, polishing, machine repair, external repair, disassembly, electrical engineering, and so on. Among them, each category can also correspond to a more detailed maintenance plan according to the degree of damage of the damaged parts, for example, according to the degree of scratching from light to heavy, the painting treatment is divided into: touch-up, half-spray and full-spray.

According to one embodiment, the training samples of the loss assessment model can be obtained from historical loss assessment data of multiple loss assessment personnel. A piece of historical damage assessment data of a damage assessment person can include a marked damaged component and its degree of damage (corresponding features), as well as repair suggestions (corresponding labels). For a piece of historical damage data, the determined training sample can be, for example, the damaged part "left front bumper", the degree of damage "slight scratch", and the corresponding maintenance plan label "paint repair".

According to another embodiment, the training samples of the damage estimation model can be obtained from historical maintenance data of multiple maintenance personnel. A piece of historical maintenance data of a maintenance person can include the name of the damaged component and its degree of damage (corresponding characteristics), and a maintenance plan (corresponding label). For a piece of historical maintenance data, the determined training sample may be, for example, the damaged component "left front bumper", the degree of damage "severely deformed", and the corresponding maintenance plan label "replace left front bumper".

It can be understood that, in some implementations, the maintenance schemes given for the same damaged component and the degree of damage, different models, different damage assessment or maintenance personnel, are different.

On the one hand, for the same damaged component and the degree of damage, different vehicle models may have different repair plans. For example, the bumper of some car models is made of metal and can only be repaired in the case of deformation, while the bumper of some car models may be made of plastic material. Once deformed, it needs to be replaced. To this end, the characteristics of the fixed-loss model may also include vehicle model information. At this time, the car model information can be directly input as a feature into the loss-fixing model, or pictures or videos of the damaged vehicle can be input as a feature into the loss-fixing model, and the model's characteristics can be automatically extracted from the loss-fixing model through a neural network. In this way, the repair plan can be determined based on the vehicle type, so that the damage repair plan is more accurate.

On the other hand, for the same damaged component and degree of damage of the same vehicle model, different damage fixing or maintenance personnel will give different repair plans. For example, the bumper is also slightly scratched. Some damage or repair personnel may only recommend paint touch-up treatment, and some damage or repair personnel may recommend full spray treatment. In the case of particularly slight scratches, some damage or repair personnel may recommend polishing, and some damage or repair personnel may recommend painting. For this reason, in the embodiments of this specification, collect historical data of multiple damage assessment personnel/maintenance personnel as much as possible to diversify the samples and avoid label bias caused by differences in individual experience and habits, thereby improving the effectiveness of the output results of the damage assessment model. .

Then, through step 203, based on the output result of the damage damage model, a damage file is generated for the damaged vehicle. The fixed loss file here can be in the form of a document, a list, etc. Damage assessment documents can be used to describe vehicle damage and its maintenance plan, as well as vehicle damage and the amount of resources lost by the vehicle owner or insurance company based on the vehicle damage, such as the amount of compensation paid by the insurance company.

According to the description of step 202, the output result of the constant-loss model is a maintenance plan, such as repainting the left front bumper, and so on. In the process of generating the damage file based on the output result of the damage damage model, if the damage file is used to describe the damage of the vehicle and its repair plan, the repair plan can be generated into a damage file. If the damage file is also used to describe the vehicle ownership If the amount of resources lost by the insurance company or the insurance company, it is necessary to generate a loss assessment document together with the maintenance plan and the share of the resources involved. In an optional implementation manner, the resource share corresponding to the maintenance plan may be determined through a predetermined mapping relationship. For example, for the maintenance plan of “replacement of the left front bumper”, the resource share (maintenance cost) corresponding to at least one method can be mapped to the 4S shop to replace the original configuration bumper, and the ordinary vehicle repair shop to replace the ordinary bumper. Optionally, the damaged file may also include the marking result of the damaged component and the marking result of the degree of damage.

According to a possible design, as shown in Figure 5, the generated loss assessment file can also be inspected by experienced loss assessment personnel. If the damage assessment document is accurate, the inspection ends. Conversely, if the damage assessment file is inaccurate, for example, the maintenance plan for a slight scratch on the bumper is to replace the bumper, etc., then the generated damage assessment file should be corrected by professional damage assessment personnel. Further, it is also possible to generate a new sample based on the revised loss assessment file, and use the new sample to update the loss assessment model. In this way, the fixed-loss model can be continuously optimized, and the accuracy of the fixed-loss model can be improved.

In FIG. 5, the solid line depicts the process of generating the constant loss file in the process shown in FIG. 2, and the dotted line describes the update process of the constant loss model.

Recalling the above process, in the process of generating damage assessment files for damaged vehicles, the damaged component marking results and the damage degree marking results are used as the characteristics of the damage assessment model. Modules independent of the damage assessment model can be used for labeling and training. The work of annotators reduces the comprehensive professionalism of annotators. In the damage model training stage, the historical damage assessment data of multiple damage assessment personnel and/or the historical maintenance data of multiple maintenance personnel are used to diversify the damage assessment model samples, eliminate the influence of personal habits or experience, and improve the damage assessment model The effectiveness of the results, thereby generating high-quality loss-making files. In short, the embodiments of this specification can improve the effectiveness of vehicle damage assessment.

According to another embodiment, there is also provided a device for generating a damage file for a damaged vehicle. Fig. 6 shows a schematic block diagram of an apparatus for generating a damage file for a damaged vehicle according to an embodiment. As shown in FIG. 6, the apparatus 600 for generating a damage assessment file for a damaged vehicle includes: an obtaining unit 61 configured to obtain the damaged component marking result and the damage degree marking result determined based on the live video of the damaged vehicle, wherein, The marking result of the damaged part indicates the damaged part and the damaged area. The marking result of the damage degree is based on the marking result of the damaged part and the result of identifying the degree of damage of each damaged component; the damage determination unit 62 is configured to at least damage the damaged parts. The component labeling results and the damage degree labeling results are input into the pre-trained damage assessment model, where the damage assessment model is trained based on the historical damage assessment data of multiple damage assessors and/or the historical maintenance data of multiple maintenance personnel for generating Damage repair plan; generating unit 63, configured to generate a damage file for the damaged vehicle based on the output result of the damage model.

According to one embodiment, the device 600 may further include a damaged part marking unit (not shown), configured to determine the damaged part marking result by: acquiring the first damaged part and the first damaged part marked in the live video. The first image frame of the damaged area; among the image frames adjacent to the first image frame within a predetermined number of frames, detect whether there is a second image frame, and there is a second damaged area in the second image frame, and the second damaged area The area is associated with the first damaged area; if there is a second image frame, the labeling result of the second image frame is determined, and the labeling result of the second image frame corresponds to the first damaged part and the second damaged area.

According to another embodiment, the device 600 may further include a damage degree labeling unit (not shown) configured to determine the damage degree labeling result in the following manner: input the live video and the damaged component labeling result into the pre-trained damage Degree prediction model; the damage degree marking result is determined according to the output result of the damage degree prediction model.

In one embodiment, the damage assessment unit 62 may be further configured to input the damage assessment model into the damaged component marking result, the damage degree marking result, and the vehicle information determined based on the live video.

In one embodiment, the output result of the fixed damage model includes a damage repair plan determined for the damaged vehicle. Among them, the damage repair plan may include, but is not limited to, at least one of painting, sheet metal, replacement, polishing, machine repair, disassembly and assembly, external repair, electrician, and so on.

At this time, the generating unit 63 may also be configured to: obtain the output result of the damage assessment model and the mapping relationship between the predetermined damage repair plan and the predetermined resource; and determine the corresponding damage repair plan for the damaged vehicle based on the above mapping relationship. Resource share: The damage repair plan determined for the damaged vehicle and its corresponding resource share are combined to generate a damage assessment file for the damaged vehicle.

According to a possible design, the device 600 may further include an update unit (not shown), configured to: obtain a detection result of the damage assessment personnel regarding whether the damage assessment file generated for the damaged vehicle is correct; in the case that the detection result is wrong , Obtain the revised loss determination file; generate a new sample for the loss determination model based on the revised loss determination file; update the loss determination model with the new sample.

It is worth noting that the apparatus 600 shown in FIG. 6 is an apparatus embodiment corresponding to the method embodiment shown in FIG. 2, and the corresponding description in the method embodiment shown in FIG. 2 is also applicable to the apparatus 600. Go into details again.

According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method described in conjunction with FIG. 2.

According to an embodiment of still another aspect, there is also provided a computing device, including a memory and a processor, the memory is stored with executable code, and when the processor executes the executable code, it implements the method described in conjunction with FIG. 2 method.

Those skilled in the art should be aware that, in one or more of the foregoing examples, the functions described in the embodiments of this specification can be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on the computer-readable medium.

The specific implementations described above further describe the purpose, technical solutions, and beneficial effects of the technical concept of this specification in further detail. It should be understood that the above are only specific implementations of the technical concept of this specification, and It is not used to limit the protection scope of the technical concept of this specification. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions of the embodiments of this specification shall be included in the protection scope of the technical concept of this specification within.

Claims

A method for generating a damage assessment file for a damaged vehicle, the method comprising:

Acquire the damaged component labeling result and the damage degree labeling result determined based on the on-site video of the damaged vehicle, wherein the damaged component labeling result indicates the damaged component and the damaged area, and the damage degree labeling result is based on Marking results of damaged parts, identification results of the degree of damage of each damaged part;

At least the damaged component marking result and the damaged degree marking result are input into a pre-trained damage assessment model, wherein the damage assessment model is based on historical damage assessment data of multiple damage assessment personnel and/or multiple maintenance personnel Training based on historical maintenance data to generate damage maintenance plans;

Based on the output result of the damage assessment model, a damage assessment file is generated for the damaged vehicle.
The method according to claim 1, wherein the marking result of the damaged part is determined in the following manner:

Acquiring a first image frame marked with the first damaged component and the first damaged area in the live video;

In the image frames adjacent to the first image frame within a predetermined number of frames, it is detected whether there is a second image frame, and there is a second damaged area in the second image frame, and the second damaged area is The first damaged area is associated;

In the case where the second image frame exists, the labeling result of the second image frame is determined, and the labeling result of the second image frame corresponds to the first damaged component and the second damaged area.
The method according to claim 1, wherein the marking result of the degree of damage is determined in the following manner:

Inputting the live video and the marking result of the damaged component into a pre-trained damage degree prediction model;

The marking result of the damage degree is determined according to the output result of the damage degree prediction model.
The method according to claim 1, wherein said inputting at least the marking result of the damaged component and the marking result of the damage degree into a pre-trained damage model comprises:

The marking result of the damaged component, the marking result of the degree of damage, and the vehicle type information determined based on the live video are input into the damage assessment model.
The method according to claim 1, wherein the output result of the damage assessment model includes a damage repair plan determined for the damaged vehicle;

The generating a damage file for the damaged vehicle based on the output result of the damage measurement model includes:

Acquiring the output result of the damage assessment model and the mapping relationship between the predetermined damage maintenance plan and the predetermined resource;

Determining the resource share corresponding to the damage maintenance plan determined for the damaged vehicle based on the mapping relationship;

The damage repair plan determined for the damaged vehicle and its corresponding resource share are combined to generate a damage assessment file for the damaged vehicle.
The method according to claim 1, wherein the method further comprises:

Obtain the detection result of the damage assessment personnel on whether the damage assessment file generated for the damaged vehicle is correct;

In the case where the detection result is wrong, obtain the corrected loss assessment file;

Generate a new sample for the loss-making model according to the revised loss-making file;

The new sample is used to update the loss assessment model.
The method according to any one of claims 1 to 6, wherein the damage repair plan includes at least one of painting, sheet metal, replacement, polishing, mechanical repair, external repair, disassembly and assembly, and electrician.
A device for generating damage assessment files for damaged vehicles, the device comprising:

The obtaining unit is configured to obtain the damaged component marking result and the damage degree marking result determined based on the on-site video of the damaged vehicle, wherein the damaged component marking result indicates the damaged component and the damaged area, and the damaged component The degree marking result is based on the marking result of the damaged component, and the result of identifying the damage degree of each damaged component;

The damage assessment unit is configured to input at least the damaged component labeling result and the damage degree labeling result into a pre-trained damage assessment model, wherein the damage assessment model is based on the historical damage assessment data of a plurality of damage assessment personnel and / Or training based on historical maintenance data of multiple maintenance personnel to generate damage maintenance plans;

The generating unit is configured to generate a damage file for the damaged vehicle based on the output result of the damage measurement model.
8. The device according to claim 8, wherein the device further comprises a damaged part marking unit configured to determine the damaged part marking result in the following manner:

Acquiring a first image frame marked with the first damaged component and the first damaged area in the live video;

In the image frames adjacent to the first image frame within a predetermined number of frames, it is detected whether there is a second image frame, and there is a second damaged area in the second image frame, and the second damaged area is The first damaged area is associated;

In the case where the second image frame exists, the labeling result of the second image frame is determined, and the labeling result of the second image frame corresponds to the first damaged component and the second damaged area.
The device according to claim 8, wherein the device further comprises a damage degree marking unit configured to determine the damage degree marking result in the following manner:

Inputting the live video and the marking result of the damaged component into a pre-trained damage degree prediction model;

The marking result of the damage degree is determined according to the output result of the damage degree prediction model.
The device according to claim 8, wherein the loss assessment unit is further configured to:

The marking result of the damaged component, the marking result of the degree of damage, and the vehicle type information determined based on the live video are input into the damage assessment model.
8. The device according to claim 8, wherein the output result of the damage assessment model comprises a damage repair plan determined for the damaged vehicle;

The generating unit is further configured to:

Acquiring the output result of the damage assessment model and the mapping relationship between the predetermined damage maintenance plan and the predetermined resource;

Determining the resource share corresponding to the damage maintenance plan determined for the damaged vehicle based on the mapping relationship;

The damage repair plan determined for the damaged vehicle and its corresponding resource share are combined to generate a damage assessment file for the damaged vehicle.
The device according to claim 8, wherein the device further comprises an update unit configured to:

Obtain the detection result of the damage assessment personnel on whether the damage assessment file generated for the damaged vehicle is correct;

In the case where the detection result is wrong, obtain the corrected loss assessment file;

Generate a new sample for the loss-making model according to the revised loss-making file;

The new sample is used to update the loss assessment model.
The device according to any one of claims 8-13, wherein the damage repair plan includes at least one of painting, sheet metal, replacement, polishing, mechanical repair, external repair, disassembly and assembly, and electrician.
A computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of any one of claims 1-7.
A computing device, comprising a memory and a processor, characterized in that executable code is stored in the memory, and when the processor executes the executable code, the device described in any one of claims 1-7 is implemented method.