Damage assessment method, device, equipment and storage medium based on video stream

Abstract

The present invention relates to the field of artificial intelligence, and discloses a method, device, equipment and storage medium for assessing damages based on video streams, which are used to improve the accuracy of claim assessment. The method for assessing damages based on video streams includes: identifying road conditions and weather on multiple video frames to obtain target road condition information and target weather information; identifying accident vehicles on multiple video frames to obtain target accident vehicles; identifying injured persons on multiple video frames to obtain target injured persons; extracting audio from the on-site survey video stream to obtain target audio, and performing voice analysis on the target audio to obtain conversation fluency and conversation rationality; inputting target road condition information, target weather information, target accident vehicle, target injured person, conversation fluency and conversation rationality into a claim amount prediction model to predict the claim amount and obtain a target claim amount. In addition, the present invention also relates to blockchain technology, and the target claim amount can be stored in a blockchain node.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a damage assessment method, device, equipment and storage medium based on video stream. Background Art

[0002] With the development of economy, the world is becoming more closely connected, and any change may cause economic fluctuations. Economic development is driven by independent events, and each event may be an opportunity. Due to the huge number of cars in the market, various traffic incidents occur on the road every day. The occurrence of the incident has a great impact on the preparation of the claim amount. If a major incident occurs and sufficient claim amount is not prepared in time, it will cause a huge impact. Therefore, it is very important to quickly assess the damage and claim for vehicles involved in traffic incidents.

[0003] The existing solution is to generate the claim amount by using the vehicle factor as the input variable of the claims ratio model and the vehicle factor having the largest weight. However, a large weight of the vehicle factor will lead to a large deviation in the claim amount generated by the claims ratio model, resulting in low accuracy in claim assessment. Summary of the invention

[0004] The present invention provides a damage assessment method, device, equipment and storage medium based on video stream, which are used to improve the accuracy of claim damage assessment.

[0005] The first aspect of the present invention provides a damage assessment method based on video stream, which includes: obtaining a site survey video stream to be processed, and dividing the site survey video stream into frames to obtain multiple video frames; calling a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information; calling a preset target detection model to perform accident vehicle recognition on the multiple video frames to obtain a target accident vehicle; performing face recognition on the multiple video frames to obtain multiple face data, and performing injured person recognition on the multiple face data to obtain a target injured person; performing audio extraction on the site survey video stream to obtain target audio, and performing speech analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person; inputting the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount.

[0006] Optionally, in a first implementation method of the first aspect of the present invention, calling a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information, including: performing feature extraction on the multiple video frames through an encoder in a preset semantic segmentation model to obtain a high-dimensional array of features corresponding to each video frame; performing pixel point type prediction on the high-dimensional array of features corresponding to each video frame through a decoder in the semantic segmentation model to obtain target road condition information; obtaining weather information to be processed, and inputting the weather information into the semantic segmentation model to perform weather recognition to obtain target weather information.

[0007] Optionally, in a second implementation method of the first aspect of the present invention, calling a preset target detection model to perform accident vehicle identification on the multiple video frames to obtain a target accident vehicle includes: inputting the multiple video frames into a preset target detection model to perform accident vehicle identification to obtain multiple candidate accident vehicles; performing stationary state judgment on the multiple candidate accident vehicles to obtain the relative sizes of the multiple candidate accident vehicles; performing geometric calculation on the relative sizes of the multiple candidate accident vehicles using a preset pixel ratio to obtain the target accident vehicle.

[0008] Optionally, in a third implementation method of the first aspect of the present invention, performing face recognition on the multiple video frames to obtain multiple face data, and performing injured person identification on the multiple face data to obtain a target injured person, includes: performing face recognition on the multiple video frames to obtain multiple face data; inputting the multiple face data into a preset character expression recognition model to perform character labeling to obtain multiple character tags; assigning values ​​to the multiple character tags to obtain a score ranking, and determining the character tag with the highest score in the score ranking as the target injured person.

[0009] Optionally, in a fourth implementation manner of the first aspect of the present invention, audio extraction is performed on the on-site survey video stream to obtain target audio, and speech analysis is performed on the target audio to obtain conversation fluency and conversation rationality corresponding to the target injured person, including: audio extraction is performed on the on-site survey video stream to obtain target audio, and the target audio is converted into text to obtain target text; word segmentation is performed on the target text through a preset natural language processing model to obtain a word segmentation result, and fluency analysis is performed on the word segmentation result and the target audio to obtain conversation fluency corresponding to the target injured person; and rationality analysis is performed on the target audio based on the conversation fluency and preset standard audio to obtain conversation rationality corresponding to the target injured person.

[0010] Optionally, in a fifth implementation method of the first aspect of the present invention, the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality are input into a preset claim amount prediction model to predict the claim amount, and obtaining the target claim amount includes: performing vector conversion on the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality to obtain a target vector; inputting the target vector into a preset claim amount prediction model, performing a logistic regression operation on the target vector through the claim amount prediction model to obtain the target claim amount.

[0011] Optionally, in a sixth implementation manner of the first aspect of the present invention, after the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality are input into a preset claim amount prediction model to predict the claim amount and obtain the target claim amount, the damage assessment method based on video stream also includes: acquiring expression data and conversation data of the target injured person; generating a target satisfaction degree based on the expression data and the conversation data, the target satisfaction degree being used to indicate the satisfaction degree of the target injured person with the target claim amount; constructing a linear relationship between the target satisfaction degree and the target claim amount, and adjusting the target claim amount based on the linear relationship and a preset amount adjustment range.

[0012] The second aspect of the present invention provides a damage assessment device based on video stream, which includes: an acquisition module for acquiring a site survey video stream to be processed, and dividing the site survey video stream into frames to obtain multiple video frames; an information recognition module for calling a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information; a vehicle recognition module for calling a preset target detection model to perform accident vehicle recognition on the multiple video frames to obtain a target accident vehicle; and a wounded person recognition module for performing accident vehicle recognition on the multiple video frames. The frame is used for performing face recognition to obtain multiple face data, and the injured person is identified on the multiple face data to obtain the target injured person; the speech analysis module is used to extract audio from the on-site survey video stream to obtain target audio, and perform speech analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person; the prediction module is used to input the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain the target claim amount.

[0013] Optionally, in a first implementation method of the second aspect of the present invention, the information identification module is specifically used to: perform feature extraction on the multiple video frames through an encoder in a preset semantic segmentation model to obtain a high-dimensional array of features corresponding to each video frame; perform pixel point type prediction on the high-dimensional array of features corresponding to each video frame through a decoder in the semantic segmentation model to obtain target road condition information; obtain weather information to be processed, and input the weather information into the semantic segmentation model for weather identification to obtain target weather information.

[0014] Optionally, in a second implementation method of the second aspect of the present invention, the vehicle identification module is specifically used to: input the multiple video frames into a preset target detection model to perform accident vehicle identification to obtain multiple candidate accident vehicles; perform stationary state judgment on the multiple candidate accident vehicles to obtain the relative sizes of the multiple candidate accident vehicles; perform geometric calculations on the relative sizes of the multiple candidate accident vehicles through a preset pixel ratio to obtain a target accident vehicle.

[0015] Optionally, in a third implementation method of the second aspect of the present invention, the injured person identification module is specifically used to: perform face recognition on the multiple video frames to obtain multiple face data; input the multiple face data into a preset character expression recognition model to perform character labeling to obtain multiple character tags; assign values ​​to the multiple character tags to obtain a score ranking, and determine the character tag with the highest score in the score ranking as the target injured person.

[0016] Optionally, in a fourth implementation manner of the second aspect of the present invention, the speech analysis module is specifically used to: perform audio extraction on the on-site survey video stream to obtain target audio, and convert the target audio into text to obtain target text; perform word segmentation on the target text through a preset natural language processing model to obtain a word segmentation result, and perform a fluency analysis on the word segmentation result and the target audio to obtain the conversation fluency corresponding to the target injured person; perform a rationality analysis on the target audio based on the conversation fluency and a preset standard audio to obtain the conversation rationality corresponding to the target injured person.

[0017] Optionally, in a fifth implementation of the second aspect of the present invention, the prediction module is specifically used to: perform vector conversion on the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality to obtain a target vector; input the target vector into a preset claim amount prediction model, and perform a logistic regression operation on the target vector through the claim amount prediction model to obtain a target claim amount.

[0018] Optionally, in a sixth implementation method of the second aspect of the present invention, the video stream-based damage assessment device also includes: an adjustment module, used to obtain expression data and conversation data of the target injured person; generate a target satisfaction rate based on the expression data and the conversation data, and the target satisfaction rate is used to indicate the target injured person's satisfaction with the target claim amount; construct a linear relationship between the target satisfaction rate and the target claim amount, and adjust the target claim amount based on the linear relationship and a preset amount adjustment range.

[0019] The third aspect of the present invention provides a video stream-based damage assessment device, comprising: a memory and at least one processor, wherein the memory stores a computer program; the at least one processor calls the computer program in the memory so that the video stream-based damage assessment device executes the above-mentioned video stream-based damage assessment method.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer-readable storage medium is run on a computer, the computer is enabled to execute the above-mentioned video stream-based damage assessment method.

[0021] In the technical solution provided by the present invention, a site survey video stream to be processed is obtained, and the site survey video stream is divided into frames to obtain multiple video frames; a preset semantic segmentation model is called to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information; a preset target detection model is called to perform accident vehicle recognition on the multiple video frames to obtain a target accident vehicle; face recognition is performed on the multiple video frames to obtain multiple face data, and the injured are identified on the multiple face data to obtain a target injured person; audio extraction is performed on the site survey video stream to obtain target audio, and speech analysis is performed on the target audio to obtain the dialogue fluency and dialogue rationality corresponding to the target injured person; the target road condition information, target weather information, target accident vehicle, target injured person, dialogue fluency and dialogue rationality are input into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount. In an embodiment of the present invention, a field survey video stream is segmented into video frames to obtain a plurality of video frames; road conditions, weather, accident vehicles and injured persons are respectively identified for the plurality of video frames to obtain target road condition information, target weather information, target accident vehicle and target injured persons; voice analysis is performed on the field survey video stream to obtain the target injured persons' conversation fluency and conversation rationality; the target road condition information, target weather information, target accident vehicle, target injured persons, conversation fluency and conversation rationality are input into a claim amount prediction model to predict the claim amount and obtain a target claim amount. The present invention uses multiple types of information as the input of the claim amount prediction model, thereby improving the amount prediction accuracy of the claim amount prediction model, thereby improving the accuracy of claim loss assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of an embodiment of a method for assessing damage based on video streams in an embodiment of the present invention;

[0023] Figure 2 Schematic diagram of another embodiment of the damage assessment method based on video stream in an embodiment of the present invention;

[0024] Figure 3 A schematic diagram of an embodiment of a damage assessment device based on video stream in an embodiment of the present invention;

[0025] Figure 4 It is a schematic diagram of another embodiment of a damage assessment device based on video stream in an embodiment of the present invention;

[0026] Figure 5 The figure is a schematic diagram of an embodiment of a damage assessment device based on video stream in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The embodiments of the present invention provide a damage assessment method, apparatus, device and storage medium based on video stream, which are used to use various types of information as input to a claim amount prediction model to improve the accuracy of claim damage assessment.

[0028] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 The first embodiment of the damage assessment method based on video stream in the embodiment of the present invention includes:

[0030] 101. Obtain a site survey video stream to be processed, and divide the site survey video stream into frames to obtain multiple video frames;

[0031] It is understandable that the execution subject of the present invention may be a damage assessment device based on video stream, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0032] Specifically, the server reads the video stream recorded by the preset survey instrument through the preset decision-making system to obtain the field survey video stream to be processed. The server decomposes the field survey video stream into multiple pictures through the preset video decoding algorithm to obtain multiple video frames, wherein the video decoding algorithm can be a high-efficiency video encoding algorithm (for example, HEVC / H.265), which is not specifically limited here.

[0033] 102. Calling a preset semantic segmentation model to perform road condition and weather recognition on multiple video frames to obtain target road condition information and target weather information;

[0034] Specifically, the server uses a preset semantic segmentation model to identify the road conditions of multiple video frames to obtain the target road condition information. Furthermore, the speech segmentation model can be a residual learning model, which is not specifically limited here. It should be noted that the principle of the semantic segmentation model specifically includes: the semantic segmentation model includes an encoder and a decoder, the server extracts features from multiple video frames through the encoder in the semantic segmentation model to decompose the multiple video frames into a high-dimensional array of features corresponding to each video frame, and the server processes the high-dimensional array of features corresponding to each video frame through the decoder to predict the classification type of each pixel in the video frame to obtain the target road condition information. It should be noted that the model training process of the semantic segmentation model includes: the server collects training images, wherein the training images are images with road condition annotation information, the server inputs the training images into the preset training model, extracts features from the training images through the neural network in the training model, obtains the prediction information of the training images, and calculates the loss value of the prediction information of the training images and the annotation information in the training images. When the loss value is less than the preset threshold, the model training is completed, and the trained training model is used as the semantic segmentation model to obtain the speech segmentation model. Furthermore, when the server semantically segments multiple video frames, because there is a sequence and information correlation between the multiple video frames, the server superimposes the video frames before and after the video frame (such as the nth frame) (such as the n-1th frame and the n+1th frame) as supplementary information to increase the recognition accuracy of the multiple video frames. Furthermore, the server splices the multiple video frames into a more complete large-size image (that is, a large-size video frame) through additional location information, and the server's semantic segmentation recognition of the large-size image can obtain more accurate recognition results. For example: if the server recognizes a part of the road, it cannot obtain accurate road condition information. If the server expands the video frame to the front and back of the road, it can more accurately judge the road condition information. The server obtains the weather information to be processed from the preset weather forecast website through the preset crawler tool, and inputs the weather information to be processed into the preset semantic segmentation model for processing to obtain the target weather information.

[0035] 103. Calling a preset target detection model to identify the accident vehicle on multiple video frames to obtain a target accident vehicle;

[0036] Specifically, the server uses a preset target detection model to identify the accident vehicle on multiple video frames to obtain the target accident vehicle, wherein the target detection model can be a continuous convolutional network model, which is not specifically limited here. Further, the continuous convolutional network model is composed of a continuous convolutional network, which extracts different features of the input image through convolution operations. A single convolutional layer can only extract some low-level features such as edges, lines, and corners. The continuous convolutional layer can iteratively extract more complex features from low-level features. The server trains the continuous convolutional network through multiple sample images to obtain a continuous convolutional network model (that is, a target detection model), wherein the sample image is an image of an accident vehicle and carries vehicle annotation information. The server processes the sample image through the continuous convolutional network in the target detection model, and the server identifies the target object in the sample image. The server displays the target object. Further, when the server displays the target object, the continuous convolutional network model can add a detection box to the target object, wherein the detection box is used to display the category of the target object. Furthermore, the server uses a target detection model to temporarily number the identified vehicle entity (i.e., the candidate accident vehicle) with a random number. The target accident vehicle for the auto insurance inspection will not move, so the server temporarily registers the target accident vehicle through target recognition of large-size images.

[0037] 104. Perform face recognition on multiple video frames to obtain multiple face data, perform injured person recognition on the multiple face data to obtain target injured persons;

[0038] Specifically, the server performs face recognition on multiple video frames, and the server temporarily registers the people present to obtain multiple face data, wherein the people present are all the people included in the multiple video frames. The server completes the label of each registered person through the person's posture, multiple face data and the collected distance of the target accident vehicle through the person expression recognition model, and obtains multiple person labels. The server assigns values ​​to the multiple person labels, and the server determines the person corresponding to the person label with the highest score after assignment as the target injured person. For example: the postures of person A, person B, and person C are standing, sitting, and lying respectively, then person A, person B, and person C are ranked 1, 2, and 3 respectively, where the ranking is the distance between the person and the target accident vehicle sorted from large to small, and then assigned in natural number order, person A is assigned 1, person B is assigned 2, and person C is assigned 3. Furthermore, the server divides the degree of pain from light to severe into 10 levels through the person expression recognition model, and the more severe the expression of pain, the higher the score; the person's consciousness is divided into three categories according to whether the eyes are open and the mouth is open through the person expression recognition model: awake (that is, eyes open, mouth open), unclear (that is, other states), and coma (that is, eyes closed, mouth closed), and awake (that is, eyes open, mouth open) is assigned 0, unclear (that is, other states) is assigned 5, and coma (that is, eyes closed, mouth closed) is assigned 10. The server adds the assigned values ​​and determines the highest score as the target injured person.

[0039] 105. Perform audio extraction on the on-site survey video stream to obtain target audio, and perform voice analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person;

[0040] Specifically, the server imports the on-site survey video stream into a preset video editor to perform audio extraction to obtain target audio. The server performs voice analysis on the target audio to obtain the conversation fluency corresponding to the target injured person, wherein the server performs voice analysis on the target audio to obtain conversation fluency and conversation rationality by classifying the target injured person's emotions. Further, the server converts the target audio into text to obtain target text. The server performs word segmentation on the target text through a natural speech processing model to obtain a word segmentation result. The server compares the difference between the speech segmentation of the target injured person and the word segmentation result according to the word segmentation result. The server identifies whether the target injured person's speech speed has been abnormally slowed down to obtain a recognition result. The recognition result includes whether the target injured person's speech speed has been abnormally slowed down and whether the target injured person's speech speed has not been abnormally slowed down. The server determines whether the target injured person has a breathing disorder based on the recognition result. When the target injured person's speech speed has been abnormally slowed down, it is determined that the target injured person has a breathing disorder, and the conversation fluency is not smooth; when the target injured person's speech speed has not been abnormally slowed down, it is determined that the target injured person does not have a breathing disorder, and the conversation fluency is smooth. The server uses voice recognition to determine the rationality of the target injured person's conversation by comparing it with the conversation fluency and the preset standard audio to obtain the conversation rationality, where the conversation rationality is used to indicate the target injured person's consciousness. The server compares the target injured person's conversation with a smooth conversation fluency with the preset standard audio to obtain the conversation rationality, where the conversation rationality includes: reasonable and unreasonable. For example, when the target injured person's conversation is "I feel very bad now" and the preset standard audio is "I feel bad", the target injured person's conversation is compared with the preset standard audio to obtain the target injured person's conversation rationality as reasonable.

[0041] 106. Input the target road condition information, target weather information, target accident vehicle, target injured person, conversation fluency and conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount.

[0042] Specifically, the server encodes the target road condition information, target weather information, target accident vehicle, target injured person, conversation fluency and conversation rationality to obtain a target vector. For example, when the target vector is [0.2, 40, 0.7, 4, 2, 101, 5], [0.2, 40] is conversation fluency, [0.7] is conversation rationality, [4] is target road condition information, [2] is target weather information,

[101] is target accident vehicle, and [5] is target injured person. The server uses the target vector as the input of the claim amount prediction model, and the server predicts the claim amount for the target vector to obtain the target claim amount. It should be noted that the claim amount prediction model is a logistic regression analysis model. The principle of the claim amount prediction model includes: the independent variable can be either continuous or categorical, the independent variable is processed through logistic regression analysis, the weight of the independent variable is obtained, and the claim amount is generated according to the weight of the independent variable. The training process of the claim amount prediction model includes: obtaining a sample vector, inputting the sample vector into a logistic regression analysis model to predict the claim amount, obtaining a training amount, calculating a loss value of the training amount, and adjusting the parameters of the logistic regression analysis model according to the loss value until the logistic regression analysis model converges to obtain a claim amount prediction model.

[0043] Furthermore, the server stores the target claim amount in the blockchain database, which is not specifically limited here.

[0044] In an embodiment of the present invention, a field survey video stream is segmented into video frames to obtain a plurality of video frames; road conditions, weather, accident vehicles and injured persons are respectively identified for the plurality of video frames to obtain target road condition information, target weather information, target accident vehicle and target injured persons; voice analysis is performed on the field survey video stream to obtain the target injured persons' conversation fluency and conversation rationality; the target road condition information, target weather information, target accident vehicle, target injured persons, conversation fluency and conversation rationality are input into a claim amount prediction model to predict the claim amount and obtain a target claim amount. The present invention uses multiple types of information as the input of the claim amount prediction model, thereby improving the amount prediction accuracy of the claim amount prediction model, thereby improving the accuracy of claim loss assessment.

[0045] See also Figure 2 The second embodiment of the damage assessment method based on video stream in the embodiment of the present invention includes:

[0046] 201. Obtain a site survey video stream to be processed, and divide the site survey video stream into frames to obtain multiple video frames;

[0047] Specifically, the server receives a service damage assessment request, extracts a to-be-processed on-site survey video stream from the service damage assessment request, and divides the on-site survey video stream into frames using a preset video decoding algorithm to obtain multiple video frames, wherein the video decoding algorithm may be a high-efficiency video encoding algorithm (for example, H.261), which is not specifically limited here.

[0048] 202. Calling a preset semantic segmentation model to perform road condition and weather recognition on multiple video frames to obtain target road condition information and target weather information;

[0049] Specifically, the server adds additional position information to the current video frame (belonging to any one of the multiple video frames), that is, the server splices the multiple video frames into a more complete large-size picture. The server can obtain more accurate recognition results by performing semantic segmentation recognition on the large-size picture.

[0050] Optionally, the server extracts features from multiple video frames through an encoder in a preset semantic segmentation model to obtain a high-dimensional array of features corresponding to each video frame. Furthermore, when the server semantically segments multiple video frames, because multiple video frames have order and information correlation before and after, the server superimposes the previous and next video frames of the video frame as supplementary information to increase the recognition accuracy of multiple video frames; the server predicts the pixel type of the high-dimensional array of features corresponding to each video frame through a decoder in the semantic segmentation model to obtain target road condition information; the server obtains the weather information to be processed from the preset weather forecast website through a preset crawler tool, and the server encodes the obtained weather information into a two-dimensional array, and the server inputs the two-dimensional array as a channel of multiple video frames into the semantic segmentation model for weather recognition to obtain the target weather information. For example: when the weather information obtained by the server is sunny, the server encodes the sunny weather into a two-dimensional array, and the server inputs the two-dimensional array into the semantic segmentation model for weather recognition, and obtains the target weather information as sunny, wherein the encoding of the weather information is implemented by the server through the neural network layer of the semantic segmentation model.

[0051] 203. Calling a preset target detection model to identify accident vehicles on multiple video frames to obtain a target accident vehicle;

[0052] Specifically, the server uses a target detection model to temporarily number the identified vehicle entity with a random number. The target accident vehicle for the auto insurance inspection will not move, so the server registers the vehicle through target recognition of large-size images.

[0053] Optionally, the server inputs multiple video frames into a preset target detection model to identify the accident vehicle, and obtains multiple candidate accident vehicles. Further, the server creates a large-size picture through a preset acceleration sensor and a preset gyroscope. Two minutes after the preset surveyor is started, the server uses the preset compass information to establish a polar coordinate system of a large-size picture with east-west as the horizontal coordinate with the location as the center to mark the vehicle position identified in the picture, and obtains multiple candidate accident vehicles, wherein the angle information of the polar coordinates is the angle between the center of the observed target projection on the ground and the line connecting the surveyor, and the distance information is expressed as a relative distance; the server judges the stationary state of multiple candidate accident vehicles, and obtains the relative sizes of multiple candidate accident vehicles, wherein the stationary state The judgment is obtained by the degree of change of the coordinates of multiple candidate accident vehicles and the degree of change of the pixel value of the pixel value of multiple superimposed average synthetic images, wherein the coordinate change is calculated by the moving distance, the pixel value change degree is measured by the change of the pixel variance of the identification area over time, and the relative size is obtained by geometric calculation through the pixel ratio of the wheel height in the image; the server performs geometric calculation on the relative sizes of multiple candidate accident vehicles through a preset pixel ratio to obtain the target accident vehicle, and the server judges the stationary vehicle through the degree of change of the vehicle coordinates and the degree of change of the pixel value of the pixel value of multiple superimposed average synthetic images to obtain the target accident vehicle, further, the coordinate change is calculated by the moving distance, and the pixel value change degree is used to indicate the change of the pixel variance of the identification area over time.

[0054] 204. Perform face recognition on multiple video frames to obtain multiple face data, perform injured person recognition on the multiple face data to obtain a target injured person;

[0055] Specifically, the server generates a label for each registered person through a character expression recognition model to obtain multiple character labels. The server assigns values ​​to the multiple character labels, and the server determines the person corresponding to the character label with the highest score after assignment as the target injured person.

[0056] Optionally, the server performs face recognition on multiple video frames to obtain multiple face data; the server inputs the multiple face data into a preset human expression recognition model for character annotation to obtain multiple character tags; the server assigns values ​​to the multiple character tags to obtain a score ranking, and determines the character tag with the highest score in the score ranking as the target injured person. The server sorts the postures of different characters as standing, sitting, squatting, and lying down, and assigns values ​​in natural number order. For example: the postures of character A, character B, and character C are standing, sitting, and lying down, respectively, then character A, character B, and character C are sorted as 1, 2, and 3, respectively; the distances between the characters and the target accident vehicle are sorted from large to small and assigned values ​​in natural number order. Furthermore, the server divides the degree of pain from light to severe into 10 levels through the character expression recognition model. The more severe the pain expression, the higher the score. The character's consciousness is divided into three categories according to whether the eyes are open and the mouth is open through the character expression recognition model: awake (that is, eyes open, mouth open), unclear (that is, other states), and coma (that is, eyes closed, mouth closed), and assigned values ​​of 0, 5, and 10. The server adds up the assigned values ​​and determines the target injured person with the highest score.

[0057] 205. Perform audio extraction on the on-site survey video stream to obtain target audio, and perform voice analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person;

[0058] Specifically, the server determines the rationality of the target injured person's conversation through voice recognition, conversation fluency and preset standard audio, and obtains the conversation rationality, wherein the conversation rationality is used to indicate the consciousness of the target injured person.

[0059] Optionally, the server extracts audio from the on-site survey video stream to obtain target audio, and converts the target audio into text to obtain target text; the server segments the target text through a preset natural language processing model to obtain a segmentation result. For example, if the target text is "there are often differences of opinion", the server calculates the word probabilities in the target text through the Viterbi algorithm in the natural language processing model, and obtains the word probabilities in the target text: "often" -2.3, "by" -3, "there are" -2.3, "have opinions" -2.3, "opinions" -1.6, "differences" -1.6, "see" -3, "intention" -3, "see differences" -3, "divide" -2.3, and use words with high word probabilities as segmentation results, and the obtained segmentation results are: "often", "there are", "intention", "see", and "differences", wherein the natural language processing model is a natural language processing (NLP) model. The server performs a fluency analysis on the word segmentation results and the target audio to obtain the conversation fluency corresponding to the target injured person; based on the conversation fluency and the preset standard audio, the server performs a rationality analysis on the target audio to obtain the conversation rationality corresponding to the target injured person. Furthermore, the server recognizes the text in the target voice and the start and end time of each sentence through a preset speech recognition algorithm. The server calculates the speaking speed of each sentence, and the speaking speed is in "words / minute". The server marks the speaking speed less than 200 words / minute. The server calculates the variation rate of the speaking speed of the target injured person, that is, the server divides the standard deviation of the speaking speed of each sentence by the mean. The server returns the proportion of sentences with a speaking speed of less than 200 words / minute and the speaking speed variation rate of the target injured person as a set of data. In the process of judging the conversation fluency, the server will prompt some commonly used simple questions to obtain relatively standardized answers and assist in the judgment of rationality.

[0060] 206. Perform vector conversion on target road condition information, target weather information, target accident vehicle, target injured person, dialogue fluency and dialogue rationality to obtain a target vector;

[0061] Specifically, the server performs vector conversion on target road condition information, target weather information, target accident vehicle, target injured person, conversation fluency and conversation rationality to obtain a target vector, and the server inputs the target vector as a predictor into logistic regression (i.e., linear regression, ridge regression, etc.). For example, when the target vector is [0.1, 30, 0.8, 5, 1, 100, 3], [0.1, 30] is conversation fluency, [0.8] is conversation rationality, [5] is target road condition information, [1] is target weather information,

[100] is target accident vehicle, and [3] is target injured person.

[0062] 207. Input the target vector into a preset claim amount prediction model, perform a logistic regression operation on the target vector through the claim amount prediction model, and obtain a target claim amount;

[0063] Specifically, the server inputs the target vector into the claim amount prediction model, and the server performs a logistic regression operation on the target vector through the claim amount prediction model to obtain the target claim amount. Furthermore, the logistic regression operation is that the server creates a linear regression task for each column vector in the target vector, wherein each column vector in the target vector is a value for different objects of the same feature, and the operation of the linear regression task created by the server for each column vector specifically includes: the server corrects each column vector to obtain a regression value vector for each column vector; the server determines the collinearity parameter of each column vector based on the regression value vector of each column vector and a preset prediction value vector and returns the prediction value to obtain the target claim amount.

[0064] 208. Obtain the target injured person's expression data and conversation data;

[0065] Specifically, the server collects the facial expression data of the target injured person through a preset camera, and the server collects the conversation data of the target injured person through a preset voice collector.

[0066] 209. Generate target satisfaction based on expression data and conversation data, where the target satisfaction is used to indicate the target injured person's satisfaction with the target claim amount;

[0067] Specifically, the server obtains the facial expression data and dialogue data of the target injured person. Further, the server takes 80% of the target claim amount as the starting quotation amount, and the server quotes in a step-by-step manner according to 5% of the target claim amount. The server collects the facial emotions of the target injured person during the quotation process to obtain the facial expression data of the target injured person, wherein the probability of being happy in the emotion of the facial expression data is subtracted from the probability value of being angry to obtain the probability difference corresponding to the facial expression data. When the probability difference is a positive value, the facial expression data of the target injured person is considered to be satisfied, and when the probability difference is a negative value, the facial expression data of the target injured person is considered to be dissatisfied. The server generates a target satisfaction level based on the probability difference corresponding to the facial expression data and the dialogue data. Specifically, the server identifies the dialogue data, and when it is identified that the dialogue data includes the words "satisfied" and "not bad", the target satisfaction level is generated according to the probability difference corresponding to the facial expression data. Further, when the dialogue data includes the words "satisfied" and "not bad", and the probability difference is a positive value, the server uses the probability difference as the target satisfaction level corresponding to the target injured person, and the target satisfaction level is used to indicate the target injured person's satisfaction with the target claim amount.

[0068] 210. Construct a linear relationship between target satisfaction and target claim amount, and adjust the target claim amount based on the linear relationship and the preset amount adjustment range.

[0069] Specifically, the server fits a regression model of price and probability difference through linear regression to obtain the target satisfaction; the server constructs a linear relationship between the target satisfaction and the target claim amount, and the server adjusts the target claim amount through the linear relationship and a preset amount adjustment range.

[0070] Furthermore, the server stores the target claim amount in the blockchain database, which is not specifically limited here.

[0071] In an embodiment of the present invention, a field survey video stream is segmented into video frames to obtain a plurality of video frames; road conditions, weather, accident vehicles and injured persons are respectively identified for the plurality of video frames to obtain target road condition information, target weather information, target accident vehicle and target injured persons; voice analysis is performed on the field survey video stream to obtain the target injured persons' conversation fluency and conversation rationality; the target road condition information, target weather information, target accident vehicle, target injured persons, conversation fluency and conversation rationality are input into a claim amount prediction model to predict the claim amount and obtain a target claim amount. The present invention uses multiple types of information as the input of the claim amount prediction model, thereby improving the amount prediction accuracy of the claim amount prediction model, thereby improving the accuracy of claim loss assessment.

[0072] The above describes the damage assessment method based on video stream in the embodiment of the present invention. The following describes the damage assessment device based on video stream in the embodiment of the present invention. Figure 3 In the first embodiment of the damage assessment device based on video stream in the embodiment of the present invention, the device comprises:

[0073] The acquisition module 301 is used to acquire a site survey video stream to be processed, and divide the site survey video stream into frames to obtain multiple video frames;

[0074] The information recognition module 302 is used to call a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information;

[0075] The vehicle identification module 303 is used to call a preset target detection model to identify the accident vehicle on the multiple video frames to obtain a target accident vehicle;

[0076] The injured person identification module 304 is used to perform face recognition on the multiple video frames to obtain multiple face data, and perform injured person identification on the multiple face data to obtain a target injured person;

[0077] The speech analysis module 305 is used to extract audio from the on-site survey video stream to obtain target audio, and perform speech analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person;

[0078] The prediction module 306 is used to input the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount.

[0079] Furthermore, the target claim amount is stored in the blockchain database, which is not specifically limited here.

[0080] In an embodiment of the present invention, a field survey video stream is segmented into video frames to obtain a plurality of video frames; road conditions, weather, accident vehicles and injured persons are respectively identified for the plurality of video frames to obtain target road condition information, target weather information, target accident vehicle and target injured persons; voice analysis is performed on the field survey video stream to obtain the target injured persons' conversation fluency and conversation rationality; the target road condition information, target weather information, target accident vehicle, target injured persons, conversation fluency and conversation rationality are input into a claim amount prediction model to predict the claim amount and obtain a target claim amount. The present invention uses multiple types of information as the input of the claim amount prediction model, thereby improving the amount prediction accuracy of the claim amount prediction model, thereby improving the accuracy of claim loss assessment.

[0081] See also Figure 4 The second embodiment of the damage assessment device based on video stream in the embodiment of the present invention includes:

[0082] The acquisition module 301 is used to acquire a site survey video stream to be processed, and divide the site survey video stream into frames to obtain multiple video frames;

[0083] The information recognition module 302 is used to call a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information;

[0084] The vehicle identification module 303 is used to call a preset target detection model to identify the accident vehicle on the multiple video frames to obtain a target accident vehicle;

[0085] The injured person identification module 304 is used to perform face recognition on the multiple video frames to obtain multiple face data, and perform injured person identification on the multiple face data to obtain a target injured person;

[0086] The speech analysis module 305 is used to extract audio from the on-site survey video stream to obtain target audio, and perform speech analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person;

[0087] The prediction module 306 is used to input the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount.

[0088] Optionally, the information identification module 302 is specifically used to: perform feature extraction on the multiple video frames through an encoder in a preset semantic segmentation model to obtain a high-dimensional array of features corresponding to each video frame; perform pixel point type prediction on the high-dimensional array of features corresponding to each video frame through a decoder in the semantic segmentation model to obtain target road condition information; obtain weather information to be processed, and input the weather information into the semantic segmentation model for weather identification to obtain target weather information.

[0089] Optionally, the vehicle identification module 303 is specifically used to: input the multiple video frames into a preset target detection model to perform accident vehicle identification to obtain multiple candidate accident vehicles; perform stationary state judgment on the multiple candidate accident vehicles to obtain the relative sizes of the multiple candidate accident vehicles; perform geometric calculations on the relative sizes of the multiple candidate accident vehicles through a preset pixel ratio to obtain a target accident vehicle.

[0090] Optionally, the injured person identification module 304 is specifically used to: perform face recognition on the multiple video frames to obtain multiple face data; input the multiple face data into a preset character expression recognition model for character labeling to obtain multiple character tags; assign values ​​to the multiple character tags to obtain a score ranking, and determine the character tag with the highest score in the score ranking as the target injured person.

[0091] Optionally, the speech analysis module 305 is specifically used to: extract audio from the on-site survey video stream to obtain target audio, and convert the target audio into text to obtain target text; segment the target text through a preset natural language processing model to obtain a segmentation result, and perform a fluency analysis on the segmentation result and the target audio to obtain the conversation fluency corresponding to the target injured person; perform a rationality analysis on the target audio based on the conversation fluency and a preset standard audio to obtain the conversation rationality corresponding to the target injured person.

[0092] Optionally, the prediction module 306 is specifically used to: perform vector conversion on the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality to obtain a target vector; input the target vector into a preset claim amount prediction model, and perform a logistic regression operation on the target vector through the claim amount prediction model to obtain a target claim amount.

[0093] Optionally, the damage assessment device based on video stream also includes: an adjustment module 307, which is used to obtain the expression data and conversation data of the target injured person; generate a target satisfaction degree based on the expression data and the conversation data, and the target satisfaction degree is used to indicate the satisfaction of the target injured person with the target claim amount; construct a linear relationship between the target satisfaction degree and the target claim amount, and adjust the target claim amount based on the linear relationship and a preset amount adjustment range.

[0094] Furthermore, the target claim amount is stored in the blockchain database, which is not specifically limited here.

[0095] In an embodiment of the present invention, a field survey video stream is segmented into video frames to obtain a plurality of video frames; road conditions, weather, accident vehicles and injured persons are respectively identified for the plurality of video frames to obtain target road condition information, target weather information, target accident vehicle and target injured persons; voice analysis is performed on the field survey video stream to obtain the target injured persons' conversation fluency and conversation rationality; the target road condition information, target weather information, target accident vehicle, target injured persons, conversation fluency and conversation rationality are input into a claim amount prediction model to predict the claim amount and obtain a target claim amount. The present invention uses multiple types of information as the input of the claim amount prediction model, thereby improving the amount prediction accuracy of the claim amount prediction model, thereby improving the accuracy of claim loss assessment.

[0096] above Figure 3 and Figure 4 The damage assessment device based on video stream in the embodiment of the present invention is described in detail from the perspective of modular functional entities, and the damage assessment device based on video stream in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0097] Figure 5 : is a schematic diagram of the structure of a video stream-based damage assessment device provided by an embodiment of the present invention. The video stream-based damage assessment device 500 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 510 (for example, one or more processors) and a memory 520, and one or more storage media 530 (for example, one or more massive storage devices) storing application programs 533 or data 532. Among them, the memory 520 and the storage medium 530 can be short-term storage or permanent storage. The program stored in the storage medium 530 may include one or more modules (not shown in the figure), and each module may include a series of computer program operations in the video stream-based damage assessment device 500. Furthermore, the processor 510 may be configured to communicate with the storage medium 530 to execute a series of computer program operations in the storage medium 530 on the video stream-based damage assessment device 500.

[0098] The damage assessment device 500 based on video stream may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input and output interfaces 560, and / or one or more operating systems 531, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 5 The structure of the video stream-based damage assessment device shown does not constitute a limitation of the video stream-based damage assessment device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0099] The present invention also provides a damage assessment device based on video stream, which includes a memory and a processor. The memory stores a computer-readable computer program. When the computer-readable computer program is executed by the processor, the processor executes the steps of the damage assessment method based on video stream in the above-mentioned embodiments.

[0100] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. A computer program is stored in the computer-readable storage medium. When the computer program runs on a computer, the computer executes the steps of the video stream-based damage assessment method.

[0101] Furthermore, the computer-readable storage medium may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function, etc.; the data storage area may store data created according to the use of the blockchain node, etc.

[0102] The blockchain referred to in this invention is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0103] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0104] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several computer programs to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0105] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A damage assessment method based on video stream, characterized in that: The damage assessment method based on video stream includes: Acquire a site survey video stream to be processed, and divide the site survey video stream into frames to obtain multiple video frames; Calling a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information; Calling a preset target detection model to identify the accident vehicle on the multiple video frames to obtain a target accident vehicle; Performing face recognition on the multiple video frames to obtain multiple face data, and performing injured person recognition on the multiple face data to obtain a target injured person; Performing audio extraction on the on-site survey video stream to obtain target audio, and performing voice analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person; Input the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount; The calling of a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information includes: Extracting features from the multiple video frames through an encoder in a preset semantic segmentation model to obtain a high-dimensional array of features corresponding to each video frame; The decoder in the semantic segmentation model predicts the pixel type of the feature high-dimensional array corresponding to each video frame to obtain target road condition information; The weather information to be processed is obtained, and the weather information is input into the semantic segmentation model for weather recognition to obtain target weather information.

2. The damage assessment method based on video stream according to claim 1 is characterized in that: The calling of a preset target detection model to identify the accident vehicle on the multiple video frames to obtain the target accident vehicle includes: Inputting the plurality of video frames into a preset target detection model to perform accident vehicle identification, and obtaining a plurality of candidate accident vehicles; Determining the stationary state of the multiple candidate accident vehicles to obtain the relative sizes of the multiple candidate accident vehicles; The relative sizes of the plurality of candidate accident vehicles are proportionally calculated using a preset pixel ratio to obtain a target accident vehicle.

3. The damage assessment method based on video stream according to claim 1, characterized in that: The performing face recognition on the multiple video frames to obtain multiple face data, and performing injured person recognition on the multiple face data to obtain a target injured person, includes: Performing face recognition on the multiple video frames to obtain multiple face data; Inputting the plurality of face data into a preset human expression recognition model for human character labeling to obtain a plurality of human character labels; The multiple character tags are assigned values ​​to obtain a score ranking, and the character tag with the highest score in the score ranking is determined as the target injured person.

4. The damage assessment method based on video stream according to claim 1 is characterized in that: The extracting audio from the on-site survey video stream to obtain target audio, and performing voice analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person include: Extracting audio from the on-site survey video stream to obtain target audio, and converting the target audio into text to obtain target text; The target text is segmented by a preset natural language processing model to obtain a segmentation result, and the segmentation result and the target audio are analyzed for fluency to obtain the conversation fluency corresponding to the target injured person; Based on the conversation fluency and the preset standard audio, the target audio is analyzed for rationality to obtain the conversation rationality corresponding to the target injured person.

5. The damage assessment method based on video stream according to claim 1 is characterized in that: The step of inputting the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount includes: Performing vector conversion on the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality to obtain a target vector; The target vector is input into a preset claim amount prediction model, and a logistic regression operation is performed on the target vector through the claim amount prediction model to obtain a target claim amount.

6. The video stream-based damage assessment method according to any one of claims 1 to 5, characterized in that: After inputting the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain the target claim amount, the damage assessment method based on video stream further includes: Acquiring expression data and conversation data of the target injured person; generating a target satisfaction level based on the expression data and the conversation data, wherein the target satisfaction level is used to indicate the target injured person's satisfaction level with the target claim amount; A linear relationship between the target satisfaction and the target claim amount is constructed, and the target claim amount is adjusted based on the linear relationship and a preset amount adjustment range.

7. A damage assessment device based on video stream, characterized in that: The damage assessment device based on video stream includes: An acquisition module is used to acquire a site survey video stream to be processed, and divide the site survey video stream into frames to obtain multiple video frames; An information recognition module, used to call a preset semantic segmentation model to perform road condition and weather recognition on the multiple video frames to obtain target road condition information and target weather information; A vehicle identification module, used for calling a preset target detection model to identify the accident vehicle on the multiple video frames to obtain a target accident vehicle; A wounded person identification module, configured to perform face recognition on the plurality of video frames to obtain a plurality of face data, and perform wounded person identification on the plurality of face data to obtain a target wounded person; A speech analysis module is used to extract audio from the on-site survey video stream to obtain target audio, and perform speech analysis on the target audio to obtain the conversation fluency and conversation rationality corresponding to the target injured person; A prediction module, used for inputting the target road condition information, the target weather information, the target accident vehicle, the target injured person, the conversation fluency and the conversation rationality into a preset claim amount prediction model to predict the claim amount and obtain a target claim amount; The information identification module is specifically used for: Extracting features from the multiple video frames through an encoder in a preset semantic segmentation model to obtain a high-dimensional array of features corresponding to each video frame; The decoder in the semantic segmentation model predicts the pixel type of the feature high-dimensional array corresponding to each video frame to obtain target road condition information; The weather information to be processed is obtained, and the weather information is input into the semantic segmentation model for weather recognition to obtain target weather information.

8. The damage assessment device based on video stream according to claim 7, characterized in that: The injured person identification module is specifically used for: Performing face recognition on the multiple video frames to obtain multiple face data; Inputting the plurality of face data into a preset human expression recognition model for human character labeling to obtain a plurality of human character labels; The multiple character tags are assigned values ​​to obtain a score ranking, and the character tag with the highest score in the score ranking is determined as the target injured person.

9. A damage assessment device based on video stream, characterized in that: The damage assessment device based on video stream includes: a memory and at least one processor, wherein a computer program is stored in the memory; The at least one processor calls the computer program in the memory to enable the video stream-based damage assessment device to execute the video stream-based damage assessment method according to any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the video stream-based damage assessment method as described in any one of claims 1-6 is implemented.

Images