Data processing method and system based on vehicle remote camera monitoring system

Abstract

The invention provides a data processing method and system based on a vehicle remote camera monitoring system, and the method comprises the steps: integrating a first vehicle monitoring video and an associated target vehicle monitoring video, employing a candidate video analysis model comprising a basic visual angle vector and a supplementary visual angle vector, and generating a basic video analysis model through restoration setting, according to the method, a basic video analysis model is established, the supplementary view vector is particularly subjected to random reduction, other network parameters are reduced according to a pre-learned general video analysis model, targeted learning is carried out on the supplementary view vector of the basic video analysis model, and a target video analysis model with higher analysis capability is generated; according to the technical scheme, the key information in the vehicle monitoring video can be captured more comprehensively, so that accurate video semantic feature data is generated, accurate risk level assessment is performed on the to-be-assessed vehicle through the vehicle insurance risk assessment system, a scientific basis is provided for decision-making of a vehicle insurance service, and the efficiency and accuracy of the insurance service are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a data processing method and system based on a vehicle remote camera monitoring system. Background Art

[0002] With the widespread application of vehicle remote camera monitoring systems, the effective processing and analysis of vehicle monitoring videos has become an important topic in the fields of vehicle management, safety monitoring, and vehicle insurance risk assessment. Traditional vehicle monitoring video processing methods often rely on manual observation or simple image processing algorithms, which have many limitations when processing large-scale and diverse vehicle monitoring video data. For example, manual observation is time-consuming and laborious, and is easily affected by subjective factors, while simple image processing algorithms have difficulty accurately capturing key information in the video, resulting in low accuracy and reliability of the analysis results.

[0003] In order to overcome these limitations, in recent years, deep learning-based video analysis models have been gradually applied to the processing of vehicle surveillance videos. These models can automatically learn the feature representation of video data, thereby accurately identifying and analyzing key information such as vehicle behavior and road conditions in the video. However, existing video analysis models still have some problems when processing vehicle surveillance videos. Due to the diverse and complex perspectives of vehicle surveillance videos, a single video analysis model often cannot cover all possible perspectives, resulting in poor analysis results at certain specific perspectives. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a data processing method based on a vehicle remote camera monitoring system, the method comprising: Acquire first sample vehicle surveillance video data consisting of a first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video; Acquire a candidate video analysis model, wherein the spatial feature mapping space of the candidate video analysis model includes a basic viewing angle vector and a supplementary viewing angle vector, the viewing angle range of the basic viewing angle vector is a basic viewing angle range, the supplementary viewing angle vector is a viewing angle vector used to supplement the basic viewing angle vector, and the basic viewing angle vector is supplemented based on the supplementary viewing angle vector so that the viewing angle range of the spatial feature mapping space is supplemented from the basic viewing angle range to a target viewing angle range, and the target viewing angle range is larger than the basic viewing angle range; The network parameter information of the candidate video analysis model is restored and set to generate a basic video analysis model, wherein the restored parameter data of the supplementary viewing angle vector in the basic video analysis model is obtained by random restoration setting, and the restored parameter data of the remaining network parameter information in the basic video analysis model is obtained by restoration setting based on the network parameter information of a pre-learned universal video analysis model, wherein the remaining network parameter information is the network parameter information of all network parameter information of the basic video analysis model except the supplementary viewing angle vector, and the remaining network parameter information includes the basic viewing angle vector; Based on the first sample vehicle monitoring video data, learning the supplementary view vector of the basic video analysis model to generate a target video analysis model; The remote monitoring video data of the vehicle to be evaluated is obtained, and the obtained remote monitoring video data of the vehicle to be evaluated is input into the target video analysis model to generate corresponding video semantic feature data, and risk feature extraction is performed on the video semantic feature data. After the risk feature data related to the vehicle insurance risk is generated, a corresponding vehicle insurance risk assessment system is constructed to perform risk level assessment on the vehicle to be evaluated.

[0005] In a possible implementation manner of the first aspect, the learning of the supplementary view vector of the basic video analysis model based on the first sample vehicle monitoring video data to generate a target video analysis model includes: Based on the first sample vehicle monitoring video data, learning the supplementary view vector of the basic video analysis model to generate a temporary video analysis model; Acquire second sample vehicle surveillance video data consisting of a second vehicle surveillance video and a target vehicle surveillance video associated with the second vehicle surveillance video; All network parameter information of the temporary video analysis model is updated based on the second sample vehicle monitoring video data to generate the target video analysis model.

[0006] In a possible implementation manner of the first aspect, the first sample vehicle surveillance video data includes a plurality of data, each of the first sample vehicle surveillance video data includes a first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video, and the learning of the supplementary view vector of the basic video analysis model based on the first sample vehicle surveillance video data to generate a temporary video analysis model includes: Traversing the first vehicle surveillance video in each of the first sample vehicle surveillance video data, obtaining positive sample data and negative sample data of the first vehicle surveillance video, wherein the positive sample data of the first vehicle surveillance video is a target vehicle surveillance video corresponding to the same first sample vehicle surveillance video data as the first vehicle surveillance video, and the negative sample data of the first vehicle surveillance video is a target vehicle surveillance video corresponding to a different first sample vehicle surveillance video data as the first vehicle surveillance video; Traversing each first vehicle surveillance video, generating a first video semantic feature of the first vehicle surveillance video, a second video semantic feature of the positive sample data of the first vehicle surveillance video, and a third video semantic feature of the negative sample data of the first vehicle surveillance video based on the basic video analysis model; Calculating a first feature distance between the first vehicle surveillance video and positive sample data of the first vehicle surveillance video based on the first video semantic feature and the second video semantic feature, and calculating a second feature distance between the first vehicle surveillance video and negative sample data of the first vehicle surveillance video based on the first video semantic feature and the third video semantic feature; Determining a first training error parameter based on the first feature distance and the second feature distance; The supplementary viewing angle vector of the basic video analysis model is learned according to the first training error parameter to generate the temporary video analysis model.

[0007] In a possible implementation manner of the first aspect, the second sample vehicle surveillance video data includes a plurality of data, each of the second sample vehicle surveillance video data includes a second vehicle surveillance video and a target vehicle surveillance video associated with the second vehicle surveillance video, and the updating all network parameter information of the temporary video analysis model based on the second sample vehicle surveillance video data to generate the target video analysis model includes: Traversing the second vehicle surveillance video in each of the second sample vehicle surveillance video data, obtaining positive sample data and negative sample data of the second vehicle surveillance video, wherein the positive sample data of the second vehicle surveillance video is a target vehicle surveillance video corresponding to the same second sample vehicle surveillance video data as the second vehicle surveillance video, and the negative sample data of the second vehicle surveillance video is a target vehicle surveillance video corresponding to a different second sample vehicle surveillance video data as the second vehicle surveillance video; Traversing each second vehicle surveillance video, generating a fourth video semantic feature of the second vehicle surveillance video, a fifth video semantic feature of the positive sample data of the second vehicle surveillance video, and a sixth video semantic feature of the negative sample data of the second vehicle surveillance video based on the temporary video analysis model; Calculating a third feature distance between the second vehicle surveillance video and the positive sample data of the second vehicle surveillance video based on the fourth video semantic feature and the fifth video semantic feature, and calculating a fourth feature distance between the second vehicle surveillance video and the negative sample data of the second vehicle surveillance video based on the fourth video semantic feature and the sixth video semantic feature; Determine a second training error parameter based on the third characteristic distance and the fourth characteristic distance; Determining a target training error parameter according to the second training error parameter; All network parameter information of the temporary video analysis model is updated according to the target training error parameter to generate the target video analysis model.

[0008] In a possible implementation manner of the first aspect, determining a target training error parameter according to the second training error parameter includes: Obtaining the remaining network parameter information of the temporary video analysis model; Determining a third training error parameter according to the restored parameter data and the updated parameter data of the remaining network parameter information; The target training error parameter is determined according to the second training error parameter and the third training error parameter.

[0009] In a possible implementation of the first aspect, a video viewing angle range of a target vehicle monitoring video associated with the first vehicle monitoring video is greater than the basic viewing angle range, and acquiring first sample vehicle monitoring video data consisting of the first vehicle monitoring video and the target vehicle monitoring video associated with the first vehicle monitoring video includes: Determine the target vehicle monitoring video from vehicle monitoring videos whose video viewing angle range is larger than the basic viewing angle range, and determine a first vehicle monitoring video associated with the target vehicle monitoring video; The first sample vehicle surveillance video data is determined according to the first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video.

[0010] In a possible implementation of the first aspect, determining the target vehicle monitoring video from a vehicle monitoring video having a video viewing angle range greater than the basic viewing angle range includes: Extracting a vehicle monitoring video having a video viewing angle range greater than the basic viewing angle range from the shared vehicle monitoring video sequence as a reference vehicle monitoring video; Determining the target vehicle monitoring video based on the reference vehicle monitoring video; The determining of the first vehicle monitoring video associated with the target vehicle monitoring video includes: The vehicle monitoring video associated with the target vehicle monitoring video in the shared vehicle monitoring video sequence is output as a first vehicle monitoring video associated with the target vehicle monitoring video.

[0011] In a possible implementation manner of the first aspect, the reference vehicle monitoring videos include a plurality of reference vehicle monitoring videos, and determining the target vehicle monitoring video according to the reference vehicle monitoring videos includes: Traversing each of the reference vehicle surveillance videos, determining a viewing angle of a surveillance semantic element of the reference vehicle surveillance video in the reference vehicle surveillance video, the surveillance semantic element being data matched by a first vehicle surveillance video corresponding to the reference vehicle surveillance video; If the viewing angle is a target viewing angle, the reference vehicle monitoring video is removed from the plurality of reference vehicle monitoring videos according to a set ratio to generate the target vehicle monitoring video.

[0012] In a possible implementation manner of the first aspect, determining a viewing angle of the surveillance semantic element of the reference vehicle surveillance video in the reference vehicle surveillance video includes: Decomposing the reference vehicle monitoring video according to the basic viewing angle range to generate multiple viewing angle video segments; respectively calculating fifth characteristic distances between the plurality of perspective video segments and a first vehicle surveillance video corresponding to the reference vehicle surveillance video; The viewing angle video segment with the smallest fifth characteristic distance is output as the viewing angle of the monitoring semantic element of the reference vehicle monitoring video in the reference vehicle monitoring video.

[0013] On the other hand, an embodiment of the present invention also provides a data processing system based on a vehicle remote camera monitoring system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0014] Based on the above aspects, the embodiment of the present application significantly improves the analysis capability of vehicle monitoring videos and the accuracy of vehicle insurance risk assessment. Specifically, the method first integrates the first vehicle monitoring video and its associated target vehicle monitoring video, uses a candidate video analysis model containing a basic view vector and a supplementary view vector, generates a basic video analysis model through restoration settings, and especially randomly restores the supplementary view vector, while other network parameters are restored according to a pre-learned general video analysis model. This differentiated restoration strategy effectively improves the flexibility and adaptability of the model. By conducting targeted learning on the supplementary view vector of the basic video analysis model, the present invention generates a target video analysis model with stronger analysis capabilities. The target video analysis model can capture key information in the vehicle monitoring video more comprehensively, thereby generating accurate video semantic feature data. Furthermore, by extracting risk features from the video semantic feature data, a complete vehicle insurance risk assessment system is constructed, which can accurately assess the risk level of the vehicle to be assessed, provide a scientific basis for the decision-making of the vehicle insurance business, effectively reduce the insurance risk, and improve the efficiency and accuracy of the insurance business. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the execution flow of a data processing method based on a vehicle remote camera monitoring system provided by an embodiment of the present invention.

[0016] Figure 2 It is a schematic diagram of the hardware architecture of a data processing system based on a vehicle remote camera monitoring system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The figure is a flow chart of a data processing method based on a vehicle remote camera monitoring system provided by an embodiment of the present invention. The data processing method based on a vehicle remote camera monitoring system is introduced in detail below.

[0018] Step S110: Acquire first sample vehicle surveillance video data consisting of a first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video.

[0019] In this embodiment, for a vehicle monitoring database, which includes vehicle monitoring videos from different intersections and sections in multiple cities, the basic viewing angle range is set to a specific angle range in front of the vehicle, for example, a viewing angle range of 30 degrees to the left and right from the center line in front of the vehicle.

[0020] First, we filter out vehicle surveillance videos whose video viewing angles are larger than the basic viewing angles from this database. For example, some surveillance cameras are installed at higher positions and can cover a larger area. The videos captured by these cameras may contain perspectives from the front to the side and even part of the rear of the vehicle. These videos are the source of surveillance videos for finding the target vehicle.

[0021] Take a large intersection in the city center as an example. There is a panoramic surveillance camera installed on the top of a high-rise building, which can capture the driving conditions of vehicles in all directions at the intersection. From the video captured by this camera, the target vehicle surveillance video that meets the requirements can be extracted. For example, a 10-minute video is selected from this panoramic video. This video contains the driving, turning, and parking behaviors of multiple cars at the intersection, and the viewing angle range is 60 degrees from the front of the vehicle to the left and right sides. This video is determined as the target vehicle surveillance video.

[0022] Then, it is necessary to determine the first vehicle monitoring video associated with the target vehicle monitoring video. At this intersection, there are some ordinary monitoring cameras for specific lanes. The viewing angles of these cameras may be relatively small and belong to the basic viewing angle range. The video shot by one of the ordinary cameras just contains the driving situation of a certain vehicle in the target vehicle monitoring video on a specific lane. The video shot by this ordinary camera is the first vehicle monitoring video associated with the target vehicle monitoring video.

[0023] Combining such a first vehicle monitoring video and a target vehicle monitoring video together forms the first example vehicle monitoring video data. Assuming that 5 sets of such first vehicle monitoring video and target vehicle monitoring video combinations are found at this intersection, then these 5 sets of data constitute the first example vehicle monitoring video data for subsequent analysis.

[0024] Step S120, obtaining a candidate video analysis model, wherein the spatial feature mapping space of the candidate video analysis model includes a basic viewing angle vector and a supplementary viewing angle vector, wherein the viewing angle range of the basic viewing angle vector is a basic viewing angle range, and the supplementary viewing angle vector is a viewing angle vector used to supplement the basic viewing angle vector, and the basic viewing angle vector is supplemented based on the supplementary viewing angle vector so that the viewing angle range of the spatial feature mapping space is supplemented from the basic viewing angle range to the target viewing angle range, and the target viewing angle range is larger than the basic viewing angle range.

[0025] Consider a research and development scenario of vehicle surveillance video analysis. There is a pre-built video analysis model library that contains candidate video analysis models with various functions and structures.

[0026] The selected candidate video analysis model is specially designed for vehicle surveillance video analysis. In the spatial feature mapping space of this model, the basic view vector is defined as a vector representation corresponding to the basic view range mentioned above. For example, if the basic view range is 30 degrees to the left and right of the front of the vehicle, then the basic view vector mathematically represents the feature mapping relationship within this specific view range.

[0027] The purpose of the supplementary view vector is to expand the range of view that the model can handle. Assuming that the basic view vector can handle the features in front of the vehicle, the supplementary view vector is constructed to handle the view features of the side and part of the rear of the vehicle. For example, the supplementary view vector is designed to expand the view range from the basic 30 degrees in front and left and right to the target view range, such as 60 degrees in front and left and right.

[0028] The construction of this candidate video analysis model is based on the statistical analysis of a large amount of vehicle surveillance video data and the optimization of deep learning algorithms. Its spatial feature mapping space is constructed by mathematically modeling the shape, color, motion trajectory and other features of the vehicle under different viewing angles. The basic view vector and the supplementary view vector work together in this space. The basic view vector is responsible for processing the core forward view information, and the supplementary view vector supplements the information of the basic view vector when necessary, so that the entire model can adapt to vehicle surveillance video analysis with a wider range of viewing angles.

[0029] Step S130, restore and set the network parameter information of the candidate video analysis model to generate a basic video analysis model, the restored parameter data of the supplementary view vector in the basic video analysis model is obtained by random restoration setting, the restored parameter data of the remaining network parameter information in the basic video analysis model is obtained by restoration setting based on the network parameter information of a pre-learned universal video analysis model, the remaining network parameter information is the network parameter information of all the network parameter information of the basic video analysis model except the supplementary view vector, and the remaining network parameter information includes the basic view vector.

[0030] In a laboratory environment for video analysis algorithm development, the candidate video analysis model mentioned above has been obtained. In order to generate a basic video analysis model, its network parameter information needs to be restored and set.

[0031] For the restored parameter data of the supplementary view vector, a random restoration setting method is used. For example, assuming that the supplementary view vector has 10 related parameters in the model, a random number generator is used to assign initial values ​​to these 10 parameters within a certain numerical range. This numerical range is determined based on the structure of the model and previous experience, for example, these parameters may be random numbers between -1 and 1.

[0032] As for the remaining network parameter information, restore the settings based on the network parameter information of the pre-learned general video analysis model. The general video analysis model is a model learned from a large number of different types of video data, and it has a certain degree of universality. Extract relevant network parameter information from this general video analysis model. For example, the parameters related to the basic view vector and other parameters related to the basic structure of the model, copy the values ​​of these parameters to the basic video analysis model according to the settings in the general video analysis model.

[0033] To illustrate with a specific example, suppose that the value of a weight parameter corresponding to the basic view vector in the general video analysis model is 0.5. In the basic video analysis model, this weight parameter is also set to 0.5. In this way, a basic video analysis model is generated through this mixed restoration setting method. This basic video analysis model has certain basic capabilities based on the general video analysis model in its initial state. At the same time, the random restoration setting of the supplementary view vector also provides a certain exploration space for subsequent learning.

[0034] Step S140: Based on the first sample vehicle monitoring video data, the supplementary viewing angle vector of the basic video analysis model is learned to generate a target video analysis model.

[0035] Let’s continue with the example of the vehicle surveillance video at the intersection. We already have a basic video analysis model and the first example of vehicle surveillance video data.

[0036] First, the first vehicle monitoring video in the first example vehicle monitoring video data contains the driving information of the vehicle within the basic viewing angle range, while the target vehicle monitoring video contains the vehicle information in a wider viewing angle range. The first vehicle monitoring video in each first example vehicle monitoring video data is traversed.

[0037] Take one of the first vehicle monitoring videos as an example. This video records the acceleration, deceleration and turning behavior of a car in a specific lane. The first vehicle monitoring video is used as the input of the basic video analysis model, and the model will generate the first video semantic features of the first vehicle monitoring video. The first video semantic features may include information such as the vehicle's motion pattern and its relative position to surrounding vehicles.

[0038] At the same time, the target vehicle monitoring video corresponding to the first vehicle monitoring video (that is, the positive sample data) is also input into the model, and the model generates its second video semantic features. In addition, the target vehicle monitoring video (negative sample data) corresponding to the first sample vehicle monitoring video data different from the first vehicle monitoring video is also input into the model to generate the third video semantic features.

[0039] Assume that the motion trajectory of the vehicle in the first vehicle monitoring video when turning is smooth, which is represented by a specific curve pattern in the first video semantic feature. The corresponding target vehicle monitoring video (positive sample data) shows the turning of the vehicle from a wider range of viewing angles, and its second video semantic feature contains more information about the posture of the side of the vehicle when turning. The vehicle in the negative sample data may turn in different scenarios, and its third video semantic feature is quite different from the first video semantic feature.

[0040] Then, based on the first video semantic feature and the second video semantic feature, a first feature distance between the two is calculated. For example, the first feature distance is obtained by calculating the Euclidean distance between the two on semantic features such as vehicle motion trajectory and posture. Similarly, the second feature distance is calculated based on the first video semantic feature and the third video semantic feature.

[0041] Assume that the first feature distance is 0.3 (indicating that the positive sample data and the first vehicle monitoring video are relatively close in semantic features), and the second feature distance is 0.8 (indicating that the negative sample data and the first vehicle monitoring video are quite different in semantic features). Based on these two feature distances, the first training error parameter can be determined. For example, a simple weighted calculation method can be used, where the first training error parameter = 0.3 * 0.5 + 0.8 * 0.5 = 0.55.

[0042] The supplementary view vector of the basic video analysis model is learned based on the first training error parameter. This learning process may involve adjusting the values ​​of various parameters in the supplementary view vector to reduce the first training error parameter. For example, if a parameter in the supplementary view vector is related to the shape feature extraction from the side view of the vehicle, when the first training error parameter is large, this parameter will be adjusted, which may increase its weight or change its calculation method.

[0043] After multiple rounds of such processing on all first vehicle monitoring videos in the first example vehicle monitoring video data, a temporary video analysis model is generated.

[0044] Then, a second sample vehicle monitoring video data consisting of a second vehicle monitoring video and a target vehicle monitoring video associated with the second vehicle monitoring video is obtained. Assume that the second sample vehicle monitoring video data comes from a monitoring video of another intersection with a different layout.

[0045] The second vehicle surveillance video in each second sample vehicle surveillance video data is traversed again, positive sample data and negative sample data are obtained in a similar manner to the first sample vehicle surveillance video data, and the fourth video semantic feature, the fifth video semantic feature and the sixth video semantic feature are calculated.

[0046] For example, for a vehicle parking behavior in a second vehicle monitoring video, its fourth video semantic feature contains information such as the location and speed change of the vehicle when it is parked from a basic perspective. The fifth video semantic feature of the corresponding positive sample data (the target vehicle monitoring video in the same second sample vehicle monitoring video data) contains information about the surrounding environment when the vehicle is parked from a wider perspective, etc. The sixth video semantic feature of the negative sample data contains relevant information about vehicle parking in different scenarios.

[0047] The third characteristic distance and the fourth characteristic distance are calculated, for example, the third characteristic distance is 0.2 and the fourth characteristic distance is 0.7. The second training error parameter is determined based on the two characteristic distances, and it is assumed that the second training error parameter is 0.45 obtained by using the same weighted calculation method.

[0048] Next, the remaining network parameter information of the temporary video analysis model is obtained. Assume that the value of one parameter in the remaining network parameter information in the temporary video analysis model is 0.6. After learning the second sample vehicle monitoring video data, the updated value of this parameter is 0.65. The third training error parameter is determined based on the restored parameter data and the updated parameter data of this parameter. For example, the absolute value of the difference between the two can be calculated as the third training error parameter, that is, 0.05.

[0049] A target training error parameter is determined according to the second training error parameter and the third training error parameter, for example, the target training error parameter = 0.45 + 0.05 = 0.5.

[0050] Finally, all network parameter information of the temporary video analysis model is updated according to the target training error parameter to generate the target video analysis model. This update process adjusts various parameters in the temporary video analysis model, including the supplementary view vector and other remaining network parameter information, so that the model can process vehicle surveillance videos with different view ranges more accurately.

[0051] Step S150, obtaining remote monitoring video data of the vehicle to be evaluated, inputting the obtained remote monitoring video data of the vehicle to be evaluated into the target video analysis model, generating corresponding video semantic feature data, and extracting risk features from the video semantic feature data. After generating risk feature data related to vehicle insurance risk, a corresponding vehicle insurance risk assessment system is constructed to conduct risk level assessment on the vehicle to be evaluated.

[0052] For example, for a vehicle insurance platform, remote monitoring video data of the vehicle to be evaluated is obtained from remote monitoring devices installed on roads in different regions. These remote monitoring devices may be gantry cameras on highways or surveillance cameras on urban expressways.

[0053] For example, a 30-minute remote monitoring video of a car driving on a highway is obtained and the remote monitoring video data is input into the target video analysis model.

[0054] The target video analysis model first analyzes the video and generates corresponding video semantic feature data. This video semantic feature data may include information such as the vehicle's speed change, the following distance from other vehicles, and whether the lane is changed frequently. For example, the video semantic feature data shows that the vehicle has several sudden acceleration and decelerations during driving, and the following distance is too close on some sections of the road.

[0055] Risk features are then extracted from these video semantic feature data. In the case of sudden acceleration and deceleration, this may be related to the driver's driving habits or the mechanical condition of the vehicle, which is a potential risk factor. Following the vehicle too closely is also a dangerous driving behavior that increases the risk of collision. These risk feature data related to vehicle insurance risks are extracted.

[0056] Construct a vehicle insurance risk assessment system. For example, if the vehicle has frequent sudden acceleration and deceleration (more than 3 times within a certain period of time), the risk level will increase by 1; if the following distance is too close (the average following distance is less than the safety distance standard), the risk level will increase by 1. Based on the extracted risk feature data, the risk level of the vehicle to be assessed can be determined. If the risk level of the vehicle reaches 3 (assuming the risk level range is 1-5), it will be regarded as a high-risk vehicle when pricing or formulating insurance policies, and the insurance premium may be increased or the insurance terms may be adjusted.

[0057] In this way, the target video analysis model can be used to analyze a large amount of remote monitoring video data of vehicles to be evaluated, build an accurate vehicle insurance risk assessment system, and thus better manage insurance risks.

[0058] Furthermore, in one embodiment, step S150 includes the following steps: First, obtain the remote monitoring video data of the vehicle to be evaluated. This step obtains the remote monitoring video data by establishing a connection with the remote monitoring equipment installed in the road network. These remote monitoring devices are distributed in different geographical locations, including but not limited to highways, urban main roads, intersections and other key traffic areas. These devices have the function of shooting vehicle driving pictures, and send the captured video data to the designated data collection center through network transmission technology. The data collection center has the functions of data reception, storage and preliminary screening. When receiving the video data from each remote monitoring device, it is screened according to the vehicle's identity information (such as license plate number) or the pre-set list of vehicles to be evaluated, and the video data belonging to the vehicle to be evaluated is extracted and stored, thereby obtaining the remote monitoring video data of the vehicle to be evaluated, and this remote monitoring video data is used as the input for the next step.

[0059] Next, the acquired remote monitoring video data of the vehicle to be evaluated is input into the target video analysis model. The target video analysis model is a pre-trained and optimized model that has the ability to perform in-depth analysis of vehicle monitoring videos. The target video analysis model contains multiple processing modules that work together to process the input remote monitoring video data. When the remote monitoring video data is input into the target video analysis model, the target video analysis model first decodes the remote monitoring video data and converts the remote monitoring video data into an image frame sequence that can be analyzed. Then, the image frame sequence is extracted by the feature extraction module. These features include but are not limited to the appearance features of the vehicle (such as vehicle model, color, etc.), the driving state features of the vehicle (such as speed, acceleration, steering, etc.), and the relationship features between the vehicle and the surrounding environment (such as the distance from other vehicles, the position in the lane, etc.). Based on the extraction of these features, the model generates the corresponding video semantic feature data, which is used as the input of the next step.

[0060] After that, the risk features are extracted from the video semantic feature data. The risk feature extraction module analyzes the video semantic feature data according to the pre-set risk feature indicators. These risk feature indicators are determined based on historical data related to vehicle insurance risks, industry standards, and traffic rules. The risk feature extraction module first classifies the features in the video semantic feature data and identifies the feature categories related to risks, such as dangerous driving behaviors (such as speeding, sudden braking, frequent lane changes, etc.), abnormal vehicle status (such as potential mechanical problems that may be implied by damage to the vehicle appearance, etc.), and environmental risks (such as driving in bad weather, driving in road construction areas, etc.). Then, for each risk feature category, the corresponding judgment criteria and quantification methods are set. For example, for speeding behavior, according to the speed limit regulations of different road sections, it is judged whether the vehicle speed exceeds the speed limit, and quantified according to the proportion of exceeding the speed limit; for sudden braking behavior, it is quantified according to the change amplitude and duration of the vehicle acceleration. In this way, the risk feature data related to vehicle insurance risk is extracted from the video semantic feature data, and this risk feature data is used as the input of the next step.

[0061] Then, the corresponding vehicle insurance risk assessment system is constructed. The construction process is based on risk feature data and the risk assessment principles of the insurance industry. First, the framework structure of the assessment system is determined, including the weight allocation of different risk features and the risk level classification criteria. The weight allocation depends on the degree of impact of the risk feature on the insurance risk. For example, dangerous driving behavior has a greater impact on the insurance risk and may be assigned a higher weight; the abnormal vehicle state has a relatively small impact, but it will also be assigned an appropriate weight according to its severity. The risk level classification criteria are determined according to the quantitative range of the risk feature data. For example, the total value of the quantified risk feature data is divided into different intervals, and each interval corresponds to a risk level, such as low risk, medium risk, high risk, etc. At the same time, when constructing the assessment system, the particularities of different regions, different types of vehicles and different insurance policies are also taken into account to ensure the universality and accuracy of the assessment system. The vehicle insurance risk assessment system constructed based on the above steps is used to assess the risk level of the vehicle to be assessed.

[0062] Finally, the risk level of the vehicle to be evaluated is evaluated. The risk characteristic data of the vehicle to be evaluated is input into the constructed vehicle insurance risk assessment system. The assessment system calculates and analyzes the risk characteristic data according to the pre-set weight distribution and risk level classification standards. For example, first, the weighted risk characteristic value is obtained by multiplying each risk characteristic data by its corresponding weight, and then these weighted risk characteristic values ​​are added to obtain the total risk value. According to the risk level classification interval where the total risk value is located, the risk level of the vehicle to be evaluated is determined. This risk level assessment result can provide a decision-making basis for vehicle insurance companies, such as determining insurance premiums, adjusting insurance terms, etc.

[0063] Based on the above steps, the embodiment of the present application significantly improves the analysis capability of vehicle monitoring videos and the accuracy of vehicle insurance risk assessment. Specifically, the method first integrates the first vehicle monitoring video and its associated target vehicle monitoring video, uses a candidate video analysis model containing a basic view vector and a supplementary view vector, generates a basic video analysis model through restoration settings, and especially randomly restores the supplementary view vector, while other network parameters are restored according to a pre-learned general video analysis model. This differentiated restoration strategy effectively improves the flexibility and adaptability of the model. By conducting targeted learning on the supplementary view vector of the basic video analysis model, the present invention generates a target video analysis model with stronger analysis capabilities. The target video analysis model can capture key information in the vehicle monitoring video more comprehensively, thereby generating accurate video semantic feature data. Furthermore, by extracting risk features from the video semantic feature data, a complete vehicle insurance risk assessment system is constructed, which can accurately assess the risk level of the vehicle to be assessed, provide a scientific basis for the decision-making of the vehicle insurance business, effectively reduce the insurance risk, and improve the efficiency and accuracy of the insurance business.

[0064] In a possible implementation, step S140 includes: Step S141: Based on the first sample vehicle monitoring video data, the supplementary viewing angle vector of the basic video analysis model is learned to generate a temporary video analysis model.

[0065] Step S142: Acquire second sample vehicle surveillance video data consisting of a second vehicle surveillance video and a target vehicle surveillance video associated with the second vehicle surveillance video.

[0066] Step S143: Update all network parameter information of the temporary video analysis model based on the second sample vehicle monitoring video data to generate the target video analysis model.

[0067] In a possible implementation manner, the first sample vehicle monitoring video data includes a plurality of data, each of which includes a first vehicle monitoring video and a target vehicle monitoring video associated with the first vehicle monitoring video, and step S141 includes: Step S1411, traverse the first vehicle surveillance video in each of the first sample vehicle surveillance video data, and obtain positive sample data and negative sample data of the first vehicle surveillance video, wherein the positive sample data of the first vehicle surveillance video is a target vehicle surveillance video corresponding to the same first sample vehicle surveillance video data as the first vehicle surveillance video, and the negative sample data of the first vehicle surveillance video is a target vehicle surveillance video corresponding to a different first sample vehicle surveillance video data as the first vehicle surveillance video.

[0068] Step S1412, traverse each first vehicle surveillance video, and generate a first video semantic feature of the first vehicle surveillance video, a second video semantic feature of the positive sample data of the first vehicle surveillance video, and a third video semantic feature of the negative sample data of the first vehicle surveillance video based on the basic video analysis model.

[0069] Step S1413, calculating the first feature distance between the first vehicle surveillance video and the positive sample data of the first vehicle surveillance video based on the first video semantic feature and the second video semantic feature, and calculating the second feature distance between the first vehicle surveillance video and the negative sample data of the first vehicle surveillance video based on the first video semantic feature and the third video semantic feature.

[0070] Step S1414: determine a first training error parameter based on the first feature distance and the second feature distance.

[0071] Step S1415, learning the supplementary viewing angle vector of the basic video analysis model according to the first training error parameter to generate the temporary video analysis model.

[0072] In a possible implementation manner, the second sample vehicle monitoring video data includes a plurality of data, each of which includes a second vehicle monitoring video and a target vehicle monitoring video associated with the second vehicle monitoring video, and step S143 includes: Step S1431, traverse each second vehicle surveillance video in the second sample vehicle surveillance video data, and obtain positive sample data and negative sample data of the second vehicle surveillance video, the positive sample data of the second vehicle surveillance video is a target vehicle surveillance video corresponding to the same second sample vehicle surveillance video data as the second vehicle surveillance video, and the negative sample data of the second vehicle surveillance video is a target vehicle surveillance video corresponding to a different second sample vehicle surveillance video data as the second vehicle surveillance video.

[0073] Step S1432, traverse each second vehicle surveillance video, and generate the fourth video semantic feature of the second vehicle surveillance video, the fifth video semantic feature of the positive sample data of the second vehicle surveillance video, and the sixth video semantic feature of the negative sample data of the second vehicle surveillance video based on the temporary video analysis model.

[0074] Step S1433, based on the fourth video semantic feature and the fifth video semantic feature, calculate the third feature distance between the second vehicle surveillance video and the positive sample data of the second vehicle surveillance video, and based on the fourth video semantic feature and the sixth video semantic feature, calculate the fourth feature distance between the second vehicle surveillance video and the negative sample data of the second vehicle surveillance video.

[0075] Step S1434: determine a second training error parameter based on the third characteristic distance and the fourth characteristic distance.

[0076] Step S1435: determine a target training error parameter according to the second training error parameter.

[0077] Step S1436, updating all network parameter information of the temporary video analysis model according to the target training error parameter to generate the target video analysis model.

[0078] In a possible implementation, step S1435 includes: obtaining the remaining network parameter information of the temporary video analysis model, determining a third training error parameter based on the restored parameter data and the updated parameter data of the remaining network parameter information, and determining the target training error parameter based on the second training error parameter and the third training error parameter.

[0079] In the process of learning the supplementary view vector of the basic video analysis model based on the first example vehicle monitoring video data to generate the target video analysis model, the supplementary view vector of the basic video analysis model is firstly learned based on the first example vehicle monitoring video data to generate a temporary video analysis model. The first example vehicle monitoring video data includes multiple data groups, each data group has a first vehicle monitoring video and a target vehicle monitoring video associated therewith.

[0080] The traversal starts from each first vehicle monitoring video in the first sample vehicle monitoring video data. Taking the monitoring video in the urban traffic monitoring system as an example, for each first vehicle monitoring video, there are positive sample data and negative sample data. The positive sample data is the target vehicle monitoring video that belongs to the same first sample vehicle monitoring video data as the first vehicle monitoring video. The negative sample data is the target vehicle monitoring video corresponding to a different first sample vehicle monitoring video data from the first vehicle monitoring video. For example, in a monitoring scene at an intersection, a first vehicle monitoring video records the driving situation of a car on a specific lane, and the corresponding target vehicle monitoring video (positive sample data) in the same group can show the full picture of the vehicle driving at the intersection from a wider perspective, including the side and part of the rear of the vehicle. The target vehicle monitoring video (negative sample data) from other groups may be the driving situation of different vehicles at similar intersections, which is somewhat different from the vehicle driving scene in the current first vehicle monitoring video.

[0081] When traversing each first vehicle monitoring video, the corresponding video semantic features are generated based on the basic video analysis model. For the first vehicle monitoring video, a first video semantic feature is generated. This feature may include information such as the vehicle's driving speed, driving direction, and vehicle type. For the positive sample data of the first vehicle monitoring video, a second video semantic feature is generated. In addition to containing some information in the first video semantic feature, this feature may also contain additional information such as the vehicle's posture in a wider perspective and its relative position to surrounding vehicles. For the negative sample data, a third video semantic feature is generated. Since it corresponds to different vehicle driving scenes, this feature differs from the first video semantic feature and the second video semantic feature in terms of vehicle driving speed, driving trajectory, surrounding environment, etc.

[0082] The first feature distance is calculated based on the first video semantic feature and the second video semantic feature. For example, the Euclidean distance calculation method can be used. In terms of vehicle speed features, if the speed value in the first video semantic feature is 60 kilometers per hour and the speed value in the second video semantic feature is 58 kilometers per hour, the distance in the speed dimension is 2 kilometers per hour; in terms of vehicle position features, if the vehicle is located at one-third of the left side of the lane in the first video semantic feature and the vehicle is located at one-half of the left side of the lane in the second video semantic feature, the distance between the two is calculated by a predefined position quantification method. Combining the distance calculation results of these different dimensions, the first feature distance between the first vehicle monitoring video and the positive sample data is obtained. Similarly, the second feature distance is calculated based on the first video semantic feature and the third video semantic feature. Since the difference between the negative sample data and the first vehicle monitoring video is large, the difference values ​​in the dimensions of speed and position will be relatively large, thereby calculating the second feature distance.

[0083] A first training error parameter is determined based on the first feature distance and the second feature distance. Assume that a weighted calculation method is adopted, and the weight of the first feature distance is set to 0.6 and the weight of the second feature distance is set to 0.4. If the first feature distance is 0.3 and the second feature distance is 0.8, then the first training error parameter is 0.3 * 0.6 + 0.8 * 0.4 = 0.5. The supplementary perspective vector of the basic video analysis model is learned based on this first training error parameter. During the learning process, the supplementary perspective vector in the basic video analysis model will be adjusted according to the first training error parameter. For example, if a parameter in the supplementary perspective vector is related to feature extraction from the side perspective of the vehicle, when the first training error parameter is large, the value of this parameter will be adjusted according to a pre-set learning algorithm to improve the model's analysis accuracy of the first sample vehicle monitoring video data, thereby generating a temporary video analysis model.

[0084] Then, a second sample vehicle monitoring video data consisting of a second vehicle monitoring video and a target vehicle monitoring video associated with the second vehicle monitoring video is obtained. These data also come from the vehicle monitoring system and may be monitoring videos of intersections or road sections from different areas. The second sample vehicle monitoring video data also includes multiple data groups.

[0085] Traverse the second vehicle surveillance videos in each second sample vehicle surveillance video data. Similar to the processing method of the first sample vehicle surveillance video data, obtain the positive sample data and negative sample data of each second vehicle surveillance video. For example, on the main road of another city, a second vehicle surveillance video records the driving process of a car on the road. The corresponding target vehicle surveillance video (positive sample data) in the same group can show the driving conditions of the vehicle from a more comprehensive perspective, including the side and surrounding environment information of the vehicle. The target vehicle surveillance videos (negative sample data) from other groups are the driving conditions of different vehicles on similar roads.

[0086] When traversing each second vehicle monitoring video, the corresponding video semantic features are generated based on the temporary video analysis model. The fourth video semantic feature is generated for the second vehicle monitoring video, which contains information such as the vehicle's driving speed, acceleration, and whether it complies with traffic regulations. The fifth video semantic feature is generated for the positive sample data of the second vehicle monitoring video. In addition to the information in the fourth video semantic feature, it may also contain information such as the interaction between the vehicle and other vehicles from a wider perspective. The sixth video semantic feature is generated for the negative sample data. Since it corresponds to different vehicle driving scenes, it differs from the fourth video semantic feature and the fifth video semantic feature in terms of vehicle driving behavior, surrounding environment, etc.

[0087] The third feature distance is calculated based on the fourth video semantic feature and the fifth video semantic feature. For example, in terms of vehicle speed, if the speed value in the fourth video semantic feature is 55 kilometers per hour and the speed value in the fifth video semantic feature is 53 kilometers per hour, the distance in the speed dimension is 2 kilometers per hour; in terms of the vehicle's compliance with traffic rules, if the vehicle in the fourth video semantic feature does not fully comply with the traffic light instructions once, and the vehicle in the fifth video semantic feature fully complies with the traffic light instructions, the distance between the two is calculated using a predefined traffic rule compliance quantification method. The third feature distance between the second vehicle monitoring video and the positive sample data is obtained by combining the distance calculation results of different dimensions. Similarly, the fourth feature distance is calculated based on the fourth video semantic feature and the sixth video semantic feature. Since the difference between the negative sample data and the second vehicle monitoring video is large, the difference values ​​in the dimensions of speed and traffic rule compliance will be relatively large, thereby calculating the fourth feature distance.

[0088] The second training error parameter is determined based on the third feature distance and the fourth feature distance. Assume that a weighted calculation method is adopted, and the weight of the third feature distance is set to 0.7 and the weight of the fourth feature distance is set to 0.3. If the third feature distance is 0.2 and the fourth feature distance is 0.7, then the second training error parameter is 0.2 * 0.7 + 0.7 * 0.3 = 0.35.

[0089] Determine the target training error parameter based on the second training error parameter. First, obtain the remaining network parameter information of the temporary video analysis model. For example, the initial value of a network parameter in the remaining network parameter information in the temporary video analysis model is 0.5. After preliminary processing of the second sample vehicle monitoring video data, the updated value of the parameter is 0.53. Determine the third training error parameter based on the initial value and the updated value of this parameter. For example, the absolute value of the difference between the two is calculated to obtain the third training error parameter of 0.03. Then determine the target training error parameter based on the second training error parameter and the third training error parameter. For example, the target training error parameter is 0.35 + 0.03 = 0.38.

[0090] Finally, all network parameter information of the temporary video analysis model is updated according to the target training error parameter to generate the target video analysis model. During the update process, all network parameter information in the temporary video analysis model will be adjusted according to the target training error parameter. For example, if a certain network parameter is related to the feature extraction of the vehicle speed, the value of the parameter is adjusted according to the target training error parameter and the pre-set update algorithm to improve the model's analysis accuracy for the second sample vehicle monitoring video data and a wider range of vehicle monitoring video data, thereby generating the final target video analysis model.

[0091] In a possible implementation, the video viewing angle range of the target vehicle monitoring video associated with the first vehicle monitoring video is greater than the basic viewing angle range, and step S110 includes: Step S111, determining the target vehicle monitoring video from the vehicle monitoring videos whose video viewing angle range is larger than the basic viewing angle range, and determining a first vehicle monitoring video associated with the target vehicle monitoring video.

[0092] Step S112: determining the first sample vehicle surveillance video data according to the first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video.

[0093] For example, in a widely covered urban traffic monitoring system, there are many vehicle monitoring videos shot by monitoring cameras with different viewing angles. The basic viewing angle is pre-set, such as a viewing angle of 30 degrees to the left and right of the center line in front of the vehicle. Videos shot by cameras that can capture a wider range are vehicle monitoring videos with a video viewing angle greater than the basic viewing angle. For example, at large intersections or traffic hubs in the city, there are some panoramic cameras installed on the top of high-rise buildings or in special locations. These cameras can capture the situation in front of, on the side, and even part of the rear of the vehicle, and their viewing angles may reach 60 degrees to the left and right of the front of the vehicle or even larger. From the many vehicle monitoring videos shot by these panoramic cameras, the target vehicle monitoring video is determined according to specific screening criteria. This screening criterion may be based on multiple factors such as video clarity, the proportion of vehicle types included, and traffic flow conditions. For example, if you want to determine the target vehicle monitoring video related to a specific type of vehicle (such as a large passenger car), you will filter out videos shot by panoramic cameras that contain a high proportion of large passenger cars, have moderate traffic flow, and meet the video clarity requirements as the target vehicle monitoring video.

[0094] After determining the target vehicle monitoring video, the first vehicle monitoring video associated with the target vehicle monitoring video is then determined. In the same traffic monitoring system, in addition to the panoramic camera, there are many ordinary cameras for specific lanes or local areas. The video viewing angle range captured by these ordinary cameras may be close to or belong to the basic viewing angle range. For each determined target vehicle monitoring video, it is necessary to find the first vehicle monitoring video associated with it in the videos captured by many ordinary cameras. This association may be based on information such as the vehicle's identity (such as the license plate number) and the vehicle's driving trajectory. For example, if the target vehicle monitoring video shows that a car enters from an intersection within a certain time period and drives in a specific area, then in the video captured by the ordinary camera, a video of the same vehicle driving on a specific lane within the same time period can be found. This video is the first vehicle monitoring video associated with the target vehicle monitoring video.

[0095] Finally, the first example vehicle monitoring video data is determined based on the first vehicle monitoring video and the target vehicle monitoring video associated with the first vehicle monitoring video. The first vehicle monitoring video found and the corresponding target vehicle monitoring video are combined to form the first example vehicle monitoring video data. This set of data contains the driving information of the vehicle under different viewing angles. The first vehicle monitoring video provides the driving details of the vehicle within the basic viewing angle, such as the precise position of the vehicle in a specific lane, speed changes and other information; while the target vehicle monitoring video provides more comprehensive viewing angle information, such as the relative position relationship between the vehicle and the surrounding vehicles, the side posture of the vehicle when turning at the intersection and other information. The combination of these information provides a rich data foundation for the subsequent learning of the supplementary viewing angle vector of the basic video analysis model based on the first example vehicle monitoring video data, which helps to improve the accuracy and comprehensiveness of the vehicle monitoring video analysis model, and thus plays an important role in related applications such as vehicle insurance risk assessment.

[0096] In a possible implementation, step S111 includes: Step S1111: extracting a vehicle monitoring video whose video viewing angle range is larger than the basic viewing angle range from the shared vehicle monitoring video sequence as a reference vehicle monitoring video.

[0097] Step S1112, determining the target vehicle monitoring video based on the reference vehicle monitoring video.

[0098] Step S1113: outputting the vehicle surveillance video associated with the target vehicle surveillance video in the shared vehicle surveillance video sequence as a first vehicle surveillance video associated with the target vehicle surveillance video.

[0099] In a possible implementation, the reference vehicle monitoring video includes multiple reference vehicle monitoring videos, and step S1112 includes: Step S1112-1, traverse each of the reference vehicle surveillance videos, determine the viewing angle of the surveillance semantic element of the reference vehicle surveillance video in the reference vehicle surveillance video, and the surveillance semantic element is data matched by the first vehicle surveillance video corresponding to the reference vehicle surveillance video.

[0100] Step S1112-2: if the viewing angle is a target viewing angle, remove the reference vehicle monitoring video from the plurality of reference vehicle monitoring videos according to a set ratio to generate the target vehicle monitoring video.

[0101] In a possible implementation, step S1112-1 includes: Step S1112-11, decomposing the reference vehicle monitoring video according to the basic viewing angle range to generate multiple viewing angle video segments.

[0102] Step S1112-12, respectively calculating fifth characteristic distances between the multiple perspective video segments and the first vehicle surveillance video corresponding to the reference vehicle surveillance video.

[0103] Step S1112-13: output the viewing angle video segment with the smallest fifth characteristic distance as the viewing angle of the monitoring semantic element of the reference vehicle monitoring video in the reference vehicle monitoring video.

[0104] In this embodiment, in determining the target vehicle monitoring video from the vehicle monitoring video whose video viewing angle range is larger than the basic viewing angle range, firstly, the vehicle monitoring video whose video viewing angle range is larger than the basic viewing angle range is extracted from the shared vehicle monitoring video sequence as the reference vehicle monitoring video. In the urban traffic monitoring system, the shared vehicle monitoring video sequence contains video data from multiple monitoring cameras, which are distributed in various corners of the city, such as intersections, main roads, parking lot entrances, etc. Some of these cameras can capture the driving conditions of vehicles with a larger viewing angle range due to their installation position and functional characteristics, and their video viewing angle range is larger than the basic viewing angle range. For example, the basic viewing angle range is set to 30 degrees to the left and right of the center line directly in front of the vehicle, while some panoramic cameras installed on the top of a high-rise building or in the center of a large transportation hub can capture a video viewing angle range of 60 degrees or even larger to the left and right of the vehicle directly in front of it. From this huge shared vehicle monitoring video sequence, these videos with a larger viewing angle range are screened out through specific image recognition technology and viewing angle range detection algorithm, and used as reference vehicle monitoring videos.

[0105] Then, the target vehicle monitoring video is determined based on the reference vehicle monitoring video. When there are multiple reference vehicle monitoring videos, each reference vehicle monitoring video needs to be traversed. In this process, the viewing angle of the monitoring semantic element of the reference vehicle monitoring video in the reference vehicle monitoring video needs to be determined. Here, the monitoring semantic element is the data matched by the first vehicle monitoring video corresponding to the reference vehicle monitoring video. Specifically, the reference vehicle monitoring video is decomposed according to the basic viewing angle range to generate multiple viewing angle video clips. Taking a reference vehicle monitoring video as an example, assuming that the video duration is 10 minutes, and the basic viewing angle range is 30 degrees to the left and right of the front of the vehicle. Through a specific video processing algorithm, the 10-minute video is divided into multiple viewing angle video clips according to this basic viewing angle range, and several video clips under different viewing angles can be obtained, and each video clip contains the driving conditions of the vehicle within the basic viewing angle range at a certain moment.

[0106] Then, the fifth characteristic distance between the multiple perspective video clips and the first vehicle monitoring video corresponding to the reference vehicle monitoring video is calculated respectively. For each perspective video clip and the corresponding first vehicle monitoring video, the calculation of the characteristic distance involves multiple aspects. For example, in terms of the position feature of the vehicle, if the first vehicle monitoring video shows that the vehicle is located at the left third of the lane at a certain moment, and the position of the vehicle in the perspective video clip is at the left half of the lane, the distance between the two is calculated by a predefined position quantification method; in terms of the speed feature of the vehicle, if the speed of the vehicle in the first vehicle monitoring video is 50 kilometers per hour, and the speed of the vehicle in the perspective video clip is 48 kilometers per hour, the speed difference is calculated; in terms of the driving direction feature of the vehicle, if the vehicle in the first vehicle monitoring video is driving in a straight line, and the vehicle in the perspective video clip has a slight turning tendency, the difference between the two is calculated according to the quantized value of the turning angle. The fifth characteristic distance between each perspective video clip and the first vehicle monitoring video is obtained by combining the calculation results of the differences in these different features.

[0107] Then, the perspective video segment with the smallest fifth characteristic distance is output as the perspective of the monitoring semantic element of the reference vehicle monitoring video in the reference vehicle monitoring video. For example, in the comparison between the multiple perspective video segments and the first vehicle monitoring video, if there is a perspective video segment with the smallest difference in features such as vehicle position, speed, and driving direction with the first vehicle monitoring video and the smallest fifth characteristic distance, then this perspective video segment is determined as the perspective of the monitoring semantic element in the reference vehicle monitoring video.

[0108] If this perspective is the target perspective, remove the reference vehicle monitoring video from multiple reference vehicle monitoring videos according to the set ratio to generate the target vehicle monitoring video. For example, the ratio is set to 50%. After the reference vehicle monitoring video with the target perspective is determined, remove the videos that do not meet the requirements from all reference vehicle monitoring videos according to this ratio. Assuming that there are 10 reference vehicle monitoring videos in total, after the perspective is judged, the perspectives in 5 reference vehicle monitoring videos are the target perspectives, then remove the other 5 videos that do not meet the requirements from these 10 reference vehicle monitoring videos, and the remaining 5 reference vehicle monitoring videos constitute the target vehicle monitoring video.

[0109] In determining the first vehicle monitoring video associated with the target vehicle monitoring video, the vehicle monitoring video associated with the target vehicle monitoring video in the shared vehicle monitoring video sequence is output as the first vehicle monitoring video associated with the target vehicle monitoring video. In the shared vehicle monitoring video sequence, there is a certain association relationship between the various vehicle monitoring videos. This association relationship is based on multiple factors such as the vehicle's identity (such as license plate number), the vehicle's driving trajectory, and the order in which the vehicle appears in different monitoring areas. For example, the target vehicle monitoring video shows that a car with a specific license plate number enters from an intersection within a certain time period and then drives within a specific area. In the shared vehicle monitoring video sequence, through license plate number recognition and driving trajectory matching, a video of the same vehicle driving on a specific lane within the same time period can be found. This video is the first vehicle monitoring video associated with the target vehicle monitoring video. The first vehicle monitoring video and the target vehicle monitoring video determined in this way provide a basis for the subsequent construction of the first example vehicle monitoring video data, and further provide an accurate data source for performing related operations on the basic video analysis model based on the first example vehicle monitoring video data (such as learning the supplementary perspective vector, etc.), which helps to improve the accuracy and effectiveness of the entire vehicle monitoring video analysis model, and is of great significance in related applications such as vehicle insurance risk assessment.

[0110] Figure 2 FIG. 1 shows the hardware structure of a data processing system 100 based on a vehicle remote camera monitoring system for implementing the above-mentioned data processing method based on a vehicle remote camera monitoring system provided by an embodiment of the present invention, as shown in FIG. Figure 2 As shown, the data processing system 100 based on the vehicle remote camera monitoring system may include a processor 110 , a machine-readable storage medium 120 , a bus 130 and a communication unit 140 .

[0111] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the data processing system 100 based on the vehicle remote camera monitoring system uses to execute or use to complete the exemplary method described in the present invention.

[0112] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the data processing method based on the vehicle remote camera monitoring system in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0113] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the data processing system 100 based on the vehicle remote camera monitoring system mentioned above. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0114] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above data processing method based on the vehicle remote camera monitoring system is implemented.

[0115] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A data processing method based on a vehicle remote camera monitoring system, characterized in that: The method comprises: Acquire first sample vehicle surveillance video data consisting of a first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video; Acquire a candidate video analysis model, wherein the spatial feature mapping space of the candidate video analysis model includes a basic viewing angle vector and a supplementary viewing angle vector, the viewing angle range of the basic viewing angle vector is a basic viewing angle range, the supplementary viewing angle vector is a viewing angle vector used to supplement the basic viewing angle vector, and the basic viewing angle vector is supplemented based on the supplementary viewing angle vector so that the viewing angle range of the spatial feature mapping space is supplemented from the basic viewing angle range to a target viewing angle range, and the target viewing angle range is larger than the basic viewing angle range; The network parameter information of the candidate video analysis model is restored and set to generate a basic video analysis model, wherein the restored parameter data of the supplementary viewing angle vector in the basic video analysis model is obtained by random restoration setting, and the restored parameter data of the remaining network parameter information in the basic video analysis model is obtained by restoration setting based on the network parameter information of a pre-learned universal video analysis model, wherein the remaining network parameter information is the network parameter information of all network parameter information of the basic video analysis model except the supplementary viewing angle vector, and the remaining network parameter information includes the basic viewing angle vector; Based on the first sample vehicle monitoring video data, learning the supplementary view vector of the basic video analysis model to generate a target video analysis model; The remote monitoring video data of the vehicle to be evaluated is obtained, and the obtained remote monitoring video data of the vehicle to be evaluated is input into the target video analysis model to generate corresponding video semantic feature data, and risk feature extraction is performed on the video semantic feature data. After the risk feature data related to the vehicle insurance risk is generated, a corresponding vehicle insurance risk assessment system is constructed to perform risk level assessment on the vehicle to be evaluated.

2. The data processing method based on the vehicle remote camera monitoring system according to claim 1 is characterized in that: The learning of the supplementary view vector of the basic video analysis model based on the first sample vehicle monitoring video data to generate a target video analysis model includes: Based on the first sample vehicle monitoring video data, learning the supplementary view vector of the basic video analysis model to generate a temporary video analysis model; Acquire second sample vehicle surveillance video data consisting of a second vehicle surveillance video and a target vehicle surveillance video associated with the second vehicle surveillance video; All network parameter information of the temporary video analysis model is updated based on the second sample vehicle monitoring video data to generate the target video analysis model.

3. The data processing method based on the vehicle remote camera monitoring system according to claim 2 is characterized in that: The first sample vehicle monitoring video data includes a plurality of data, each of which includes a first vehicle monitoring video and a target vehicle monitoring video associated with the first vehicle monitoring video. The learning of the supplementary view vector of the basic video analysis model based on the first sample vehicle monitoring video data to generate a temporary video analysis model includes: Traversing the first vehicle surveillance video in each of the first sample vehicle surveillance video data, obtaining positive sample data and negative sample data of the first vehicle surveillance video, wherein the positive sample data of the first vehicle surveillance video is a target vehicle surveillance video corresponding to the same first sample vehicle surveillance video data as the first vehicle surveillance video, and the negative sample data of the first vehicle surveillance video is a target vehicle surveillance video corresponding to a different first sample vehicle surveillance video data as the first vehicle surveillance video; Traversing each first vehicle surveillance video, generating a first video semantic feature of the first vehicle surveillance video, a second video semantic feature of the positive sample data of the first vehicle surveillance video, and a third video semantic feature of the negative sample data of the first vehicle surveillance video based on the basic video analysis model; Calculating a first feature distance between the first vehicle surveillance video and positive sample data of the first vehicle surveillance video based on the first video semantic feature and the second video semantic feature, and calculating a second feature distance between the first vehicle surveillance video and negative sample data of the first vehicle surveillance video based on the first video semantic feature and the third video semantic feature; Determining a first training error parameter based on the first feature distance and the second feature distance; The supplementary viewing angle vector of the basic video analysis model is learned according to the first training error parameter to generate the temporary video analysis model.

4. The data processing method based on the vehicle remote camera monitoring system according to claim 2 is characterized in that: The second sample vehicle monitoring video data includes a plurality of data, each of which includes a second vehicle monitoring video and a target vehicle monitoring video associated with the second vehicle monitoring video, and the updating of all network parameter information of the temporary video analysis model based on the second sample vehicle monitoring video data to generate the target video analysis model includes: Traversing the second vehicle surveillance video in each of the second sample vehicle surveillance video data, obtaining positive sample data and negative sample data of the second vehicle surveillance video, wherein the positive sample data of the second vehicle surveillance video is a target vehicle surveillance video corresponding to the same second sample vehicle surveillance video data as the second vehicle surveillance video, and the negative sample data of the second vehicle surveillance video is a target vehicle surveillance video corresponding to a different second sample vehicle surveillance video data as the second vehicle surveillance video; Traversing each second vehicle surveillance video, generating a fourth video semantic feature of the second vehicle surveillance video, a fifth video semantic feature of the positive sample data of the second vehicle surveillance video, and a sixth video semantic feature of the negative sample data of the second vehicle surveillance video based on the temporary video analysis model; Calculating a third feature distance between the second vehicle surveillance video and the positive sample data of the second vehicle surveillance video based on the fourth video semantic feature and the fifth video semantic feature, and calculating a fourth feature distance between the second vehicle surveillance video and the negative sample data of the second vehicle surveillance video based on the fourth video semantic feature and the sixth video semantic feature; Determine a second training error parameter based on the third characteristic distance and the fourth characteristic distance; Determining a target training error parameter according to the second training error parameter; All network parameter information of the temporary video analysis model is updated according to the target training error parameter to generate the target video analysis model.

5. The data processing method based on the vehicle remote camera monitoring system according to claim 4 is characterized in that: The determining a target training error parameter according to the second training error parameter includes: Obtaining the remaining network parameter information of the temporary video analysis model; Determining a third training error parameter according to the restored parameter data and the updated parameter data of the remaining network parameter information; The target training error parameter is determined according to the second training error parameter and the third training error parameter.

6. The data processing method based on the vehicle remote camera monitoring system according to any one of claims 1 to 5, characterized in that: The video viewing angle range of the target vehicle monitoring video associated with the first vehicle monitoring video is greater than the basic viewing angle range, and the acquiring of first sample vehicle monitoring video data consisting of the first vehicle monitoring video and the target vehicle monitoring video associated with the first vehicle monitoring video includes: Determine the target vehicle monitoring video from vehicle monitoring videos whose video viewing angle range is larger than the basic viewing angle range, and determine a first vehicle monitoring video associated with the target vehicle monitoring video; The first sample vehicle surveillance video data is determined according to the first vehicle surveillance video and a target vehicle surveillance video associated with the first vehicle surveillance video.

7. The data processing method based on the vehicle remote camera monitoring system according to claim 6 is characterized in that: The step of determining the target vehicle monitoring video from a vehicle monitoring video having a video viewing angle range greater than the basic viewing angle range includes: Extracting a vehicle monitoring video having a video viewing angle range greater than the basic viewing angle range from the shared vehicle monitoring video sequence as a reference vehicle monitoring video; Determining the target vehicle monitoring video based on the reference vehicle monitoring video; The determining of the first vehicle monitoring video associated with the target vehicle monitoring video includes: The vehicle monitoring video associated with the target vehicle monitoring video in the shared vehicle monitoring video sequence is output as a first vehicle monitoring video associated with the target vehicle monitoring video.

8. The data processing method based on the vehicle remote camera monitoring system according to claim 7 is characterized in that: The reference vehicle monitoring videos include a plurality of reference vehicle monitoring videos, and determining the target vehicle monitoring video based on the reference vehicle monitoring videos includes: Traversing each of the reference vehicle surveillance videos, determining a viewing angle of a surveillance semantic element of the reference vehicle surveillance video in the reference vehicle surveillance video, the surveillance semantic element being data matched by a first vehicle surveillance video corresponding to the reference vehicle surveillance video; If the viewing angle is a target viewing angle, the reference vehicle monitoring video is removed from the plurality of reference vehicle monitoring videos according to a set ratio to generate the target vehicle monitoring video.

9. The data processing method based on the vehicle remote camera monitoring system according to claim 8 is characterized in that: The determining the viewing angle of the monitoring semantic element of the reference vehicle monitoring video in the reference vehicle monitoring video includes: Decomposing the reference vehicle monitoring video according to the basic viewing angle range to generate multiple viewing angle video segments; respectively calculating fifth characteristic distances between the plurality of perspective video segments and a first vehicle surveillance video corresponding to the reference vehicle surveillance video; The viewing angle video segment with the smallest fifth characteristic distance is output as the viewing angle of the monitoring semantic element of the reference vehicle monitoring video in the reference vehicle monitoring video.

10. A data processing system based on a vehicle remote camera monitoring system, characterized in that: The data processing system based on the vehicle remote camera monitoring system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the data processing method based on the vehicle remote camera monitoring system as described in any one of claims 1 to 9 above.

Images