A port safety intelligent video analysis method and system

Abstract

The application provides a port safety intelligent video analysis method and system, and relates to the technical field of visual analysis. First, at least one first key coupling feature vector for representing the association between operation data and identification error types is obtained based on historical operation information of a first port. Then, the occurrence probability of various identification error events is adjusted based on the deviation between historical traffic and actual predicted traffic. Finally, a multi-branch port video safety analysis model is set for various identification error types, and the weight of each branch in the port video safety analysis model is adjusted based on the identification error probability. Through the technical scheme, the association between various business data of the port and the occurrence probability of identification error types can be obtained based on big data analysis, so that the port video safety analysis model obtained by training can accurately analyze and process the port video.

Description

Technical Field

[0001] This invention relates to the field of visual analysis technology, and in particular to a method and system for intelligent video analysis of port safety. Background Technology

[0002] Port video security monitoring systems are crucial in port operations, but existing systems have significant shortcomings. Current port video recognition systems do not handle different error types independently, and there is a lack of correlation optimization between algorithms. Furthermore, the system cannot dynamically adjust recognition algorithm parameters based on actual business conditions, resulting in a large discrepancy between prediction results and the actual business environment. Therefore, there is an urgent need for an innovative solution that can dynamically correlate error types and optimize algorithms based on business predictions. Summary of the Invention

[0003] The purpose of this invention is to provide a port safety intelligent video analysis method and system, which can achieve accurate visual recognition of port safety videos.

[0004] This application proposes a port safety intelligent video analysis method, which includes: S1: Determine at least one first key coupling feature vector based on the first business identification data sequence of the first port within the first preset time interval; S2: Based on at least one of the first key coupling feature vectors and historical business data, predict the probability of at least one first target error type corresponding to each work area within a future second preset time interval; S3: Perform an association adjustment process on at least one of the first target error type probabilities to obtain at least one second target error type probability; S4: Train a port video security analysis model based on the probability of at least one of the second target error types; S4 includes: S41: Set up a shared feature extraction network to extract at least one first basic image feature from real-time video frames; S42: Set up at least one first visual branch network for each of at least one type of recognition error; S43: Determine the basic model for outbound port video security analysis based on the shared feature extraction network and at least one of the first visual branch networks; S44: Perform a weight adjustment operation on at least one of the first visual branch networks based on at least one second target error type probability to obtain a port video security analysis model.

[0005] Preferably, S1 includes: S11: Collect the first business identification data sequence of the first port within the first preset time interval; S12: Perform data saliency analysis on the first business identification data sequence to obtain at least one target identification error dataset; S13: Determine at least one first key coupling feature vector based on at least one of the target recognition error datasets.

[0006] Preferably, S12 includes: S121: Extract at least one first identification error dataset from the first service identification data sequence; S122: For each of the first identification error datasets, obtain at least one first salient data feature; S123: Obtain at least one target identification error dataset based on at least one of the first identification error datasets and at least one of the first salient data features.

[0007] Preferably, S2 includes: S21: Based on at least one of the first key coupling feature vectors and historical business data, predict the first business volume and at least one first error type probability of each operating area of ​​the first port in the future second preset time interval. S22: For each work area, based on the first traffic volume and the actual forecast traffic volume within the second preset time interval in the future, adjust at least one of the first error type probabilities to obtain at least one first target error type probability.

[0008] Preferably, S3 includes: S31: For each work area, determine the primary target error type and at least one secondary target error type based on the probability of at least one of the first target error types; S32: Determine the second traffic volume and the first forecast traffic volume based on the main target error type; S33: Based on the first deviation between the second traffic volume and the first forecast traffic volume, determine the first error type probability adjustment value corresponding to the main target error type; S34: Determine at least one second error type probability adjustment value based on the first job information, at least one second job information, and the first error type probability adjustment value; S35: Determine at least one second target error type probability based on the first error type probability adjustment value and at least one second error type probability adjustment value.

[0009] Preferably, the first job information is job information corresponding to the primary target error type, and the second job information is job information corresponding to the secondary target error type.

[0010] Preferably, the first error type probability adjustment value is an adjustment value based on the first target error type probability.

[0011] Preferably, the port video security analysis basic model is the Inception network.

[0012] This application also proposes a port safety intelligent video analysis system to implement the aforementioned port safety intelligent video analysis method.

[0013] This application proposes a port safety intelligent video analysis method and system, relating to the field of visual analysis technology. First, based on historical operational information of a first port, at least one first key coupling feature vector is analyzed to characterize the association between operational data and error types. Then, based on the deviation between historical traffic volume and actual predicted traffic volume, the probability of occurrence of various error events is adjusted. Finally, a multi-branch port video safety analysis model is set up for each error type, and the weights of each branch in the port video safety analysis model are adjusted based on the probability of each error. Through the technical solution of this application, the correlation between various port operational data and the probability of occurrence of error types can be derived based on big data analysis, thereby enabling the trained port video safety analysis model to perform accurate safety analysis and processing operations on port videos. Attached Figure Description

[0014] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0015] Figure 1 This is an execution flowchart of a port safety intelligent video analysis method according to the present invention.

[0016] Figure 2 This is the method for obtaining the target recognition error dataset in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0019] The present invention provides a detailed description of a port safety intelligent video analysis method and system.

[0020] This embodiment proposes a port safety intelligent video analysis method, the specific process of which is as follows: Figure 1 As shown.

[0021] S1: Determine at least one first key coupling feature vector based on the first business identification data sequence of the first port within the first preset time interval.

[0022] In this step, it is necessary to collect and analyze the historical business data and identification error data of the first port, so as to extract at least one first key coupling feature vector between the business data and the identification error data to characterize the association rules between the two, thereby preparing for subsequent error identification work.

[0023] Each of the first key coupling feature vectors corresponds to a type of recognition error.

[0024] S1 includes the following sub-steps: S11: Collect the first business identification data sequence of the first port within the first preset time interval.

[0025] In this step, it is necessary to collect the daily business data of the first port within the first preset time interval, and to collect the error types corresponding to the daily business data accordingly.

[0026] To ensure high data coverage, daily business data from the past three years is collected, and the daily business data is correlated with the corresponding error types. Therefore, the first business identification data sequence includes at least one time node on a daily basis. Furthermore, each data node in the first business identification data sequence includes business data and its corresponding error type data.

[0027] The daily business data includes at least: Container throughput (TEU), vehicle traffic (vehicles / day), and personnel flow (person-times / day). The distribution of work areas, such as: Yard 1, Yard 2, loading and unloading area, warehouse area, etc.; Job type distribution, such as: loading and unloading, handling, maintenance, inspection, etc.; Weather data, such as temperature, humidity, rainfall, visibility, etc.

[0028] Date characteristics, such as weekdays / weekends, holidays, seasons, etc.

[0029] The error type refers to the identification error data recorded by the video surveillance system during the same period, including: Error types specifically include: operator-container coupling, operator-spreader coupling, and operator-safety area coupling. The above error types refer to visual recognition errors caused by the overlap of the container, spreader, and operator, as well as recognition errors of the operator's operating area when the operator is working at the edge of the safety area. Location of the error, including the specific work area; The time when the error occurred, including the specific time period; The scope of the error's impact includes the types of operations and equipment affected.

[0030] S12: Perform data saliency analysis on the first business identification data sequence to obtain at least one target identification error dataset.

[0031] In this step, it is necessary to determine at least one target identification error dataset to characterize the coupling relationship between various business data and various error types, based on the coupling relationship between various business data and various error types.

[0032] The specific operating procedure is as follows: Figure 2 As shown: S121: Extract at least one first identification error dataset from the first business identification data sequence.

[0033] Since the first business identification data sequence includes multiple data sets in days within the first preset time interval, in order to analyze the coupling relationship between business data and error type data, this step requires extracting at least one of the first identification error datasets from the first business identification data sequence in which an identification error event has occurred.

[0034] Each of the first error identification datasets includes business data and error type data, respectively.

[0035] S122: For each of the first identification error datasets, obtain at least one first salient data feature.

[0036] Each of the first error identification datasets contains at least one set of business data, such as daily container loading and unloading volume and spreader operation frequency.

[0037] In this step, each piece of business data needs to be compared with the corresponding comparison threshold in turn, so that the data exceeding the threshold is identified as the first significant data feature.

[0038] For example, if the comparison threshold for daily container loading and unloading volume is 1000 TEU, then when the daily container loading and unloading volume in the first identification error dataset is greater than 1000 TEU, it is identified as the first significant data feature.

[0039] S123: Obtain at least one target identification error dataset based on at least one of the first identification error datasets and at least one of the first salient data features.

[0040] In S122, each of the first identification error datasets determined corresponds to one type of identification error among operator-container coupling, operator-spreader coupling, and operator-safety area coupling.

[0041] In order to extract significant data corresponding to each type of identification error in subsequent steps, at least one target identification error dataset needs to be obtained in this step, on a per-identification error-type basis.

[0042] For example, the target identification error dataset corresponding to the operator-container coupling contains at least one of the first salient data features corresponding to all such identification errors.

[0043] S13: Determine at least one first key coupling feature vector based on at least one of the target recognition error datasets.

[0044] In this step, for at least one target recognition error dataset obtained in S12, PCA analysis is used to derive at least one salient data feature corresponding to each recognition error type, thereby forming three first key coupling feature vectors. The salient data feature refers to the top N most frequent business data points in each target recognition error dataset.

[0045] Preferably, each error type corresponds to three features, thereby forming the first key coupling feature vector corresponding to each error type.

[0046] For example, if container handling volume, personnel flow, and container yard density are among the top three most frequent occurrences in at least one of the first significant data features, then based on the above business data combination forming the first key coupling feature vector, the first key coupling feature vector can be: Operator-container coupling feature vector: C1 = [Container loading / unloading volume, Personnel flow, Container yard density] Operator-scraper coupling feature vector: C2 = [scraper operation frequency, distance between operator and scrapper, scrapper movement speed] Operator-safety zone coupling feature vector: C3 = [safety zone workload, operator movement trajectory, safety zone boundary clarity].

[0047] S2: Based on at least one of the first key coupling feature vectors and historical business data, predict the probability of at least one first target error type corresponding to each work area within a future second preset time interval.

[0048] In S1, the first key coupling feature vectors corresponding to various types of recognition errors have been determined. In this step, it is necessary to determine the occurrence probability of each type of recognition error within a future second preset time interval based on historical operation data and at least one of the first key coupling feature vectors, so as to prepare for subsequent adjustments to the visual recognition algorithm.

[0049] S2 includes the following sub-steps: S21: Based on at least one of the first key coupling feature vectors and historical business data, predict the first business volume and at least one first error type probability of each operating area of ​​the first port in the future second preset time interval.

[0050] In this step, based on historical business data from the past 7 days and 3 of the first key coupling feature vectors, the first business volume and the probability of the first error type in each operating area of ​​the first port within the second preset time interval are predicted.

[0051] The first error type probability includes the probability of the above three error types occurring in each work area within a future second preset time interval.

[0052] The error type probability in this step is obtained by predicting the error type probability prediction model, which uses an LSTM neural network. The input is historical business data from the past 7 days and at least one of the first key coupling feature vectors obtained in S1.

[0053] The specific structure of the error type probability prediction model includes: Input layer: 7 days of historical business data (e.g., 30 features) + 3 coupled feature vectors (9 features) → 39-dimensional input For example, the 30 features may include: 1. Daily container throughput (TEU) 2. Container handling volume at the yard 3. Container handling volume at yard 2 4. Daily vehicle entry and exit volume (vehicles) 5. Daily worker turnover (person-times) 6. Total daily personnel flow (person-times) 7. Percentage of loading and unloading operations (%) 8. Percentage of handling operations (%) 9. Maintenance work percentage (%) 10. Percentage of testing operations (%) 11. Average temperature (°C) 12. Average humidity (%) 13. Rainfall (mm) 14. Visibility (km) 15. Weekdays / Weekends (0 / 1) 16. Holidays (0 / 1) 17. Seasons (1-4) Date Database: 1=Spring, 2=Summer, 3=Autumn, 4=Winter 18. Work Hours (0-23) 19. Operational activity level at Yard 1 20. Operational activity level in Yard 2 21. Activity level of loading and unloading area operations 22. Warehouse area operational activity level 23. Container yard density (units / m²) 24. Utilization rate of operating equipment (%) 25. Worker density (persons / m²) 26. Work Area Congestion Level (0-1) In the business system, 0 = idle, 1 = congested. 27. Job Type Diversity Index: 0 = Single type, 1 = High diversity in the business system. 28. Job Continuity Index: 0 = Frequent interruptions, 1 = Continuous operation. 29. Operational Complexity Index: 0 = Simple, 1 = Complex for Business Systems 30. Operational Safety Risk Index Hidden layer: 2 LSTM units, each with 128 neurons Output layer: Predicts the probability distribution of business volume and error type for each work area within the second preset time interval in the future.

[0054] Preferably, in the specific training process of the error type probability prediction model, the Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 50 training rounds.

[0055] S22: For each work area, based on the first traffic volume and the actual forecast traffic volume within the second preset time interval in the future, adjust at least one of the first error type probabilities to obtain at least one first target error type probability.

[0056] In this step, the probability of at least one of the first error types needs to be adjusted based on the actual forecast business data within the second preset time interval in the future.

[0057] In the actual adjustment process, the deviation calculation can be performed based on the business data within the second preset time interval in the future and the corresponding data in the historical business data. The larger the deviation, the larger the adjustment range required for the first error type probability, thereby obtaining at least one first target error type probability corresponding to each work area.

[0058] Preferably, appropriate adjustment rules can be determined based on expert experience to complete the adjustment operation of the probability of the first error type.

[0059] Preferably, a first error type probability adjustment model can be obtained in advance, so as to obtain at least one first target error type probability in each work area based on at least one first error type probability and the actual forecast business data within the second preset time interval in the future.

[0060] When training the first error type probability adjustment model, the predicted error type probability value of each work area in the historical data, the historical work volume, and the actual forecast business data in the second preset time interval in the future can be used as input data, and the adjusted error type probability value can be used as output data to complete the above training process.

[0061] S3: Perform an association adjustment process on at least one of the first target error type probabilities to obtain at least one second target error type probability.

[0062] For each operational area of ​​the first port, there are three corresponding error type probabilities. However, these three error type probabilities are not usually independent; they may be correlated with each other. For example, when an operator-container coupling results in an identification error, the error type probability of operator-container coupling may also affect the error type probability of operator-spreader coupling because the container and the spreader are in close proximity.

[0063] Therefore, in this step, the error probability of the error category needs to be adjusted based on the business volume corresponding to the business type with the highest error probability among the three error probabilities and the actual forecast business volume of the corresponding business type within the second preset time interval. Based on this, the occurrence probabilities of the other two error categories are then adjusted according to their correlation.

[0064] The logic behind the above method is that the error type corresponding to the highest error type probability is the dominant error type in the second future time interval, and its influence on error identification is the primary factor. Therefore, the probabilities of the other two error types are adaptively adjusted based on the dominant error type.

[0065] The aforementioned correlation adjustment process refers to adjusting the probability of occurrence of error types based on the relevance of various error type job information.

[0066] S3 includes the following sub-steps: S31: For each work area, determine the primary target error type and at least one secondary target error type based on the probability of at least one of the first target error types.

[0067] In this step, for each work area, the probability of at least one first target error type is compared, and the error type with the highest probability is determined as the primary target error type. The other error types are then determined as at least one secondary target error type.

[0068] S32: Determine the second traffic volume and the first forecast traffic volume based on the main target error type.

[0069] In this step, the second traffic volume within a historical preset time interval corresponding to the main target error type and the first forecast traffic volume within a future second preset time interval are determined based on the main target error type.

[0070] For example, when the primary target error type is the operator-container coupling type, the business volume related to container operation within the historical preset time interval is determined as the second business volume, and the business volume related to container operation within the future second preset time interval is taken as the first forecast business volume.

[0071] S33: Based on the first deviation between the second traffic volume and the first forecast traffic volume, determine the first error type probability adjustment value corresponding to the main target error type.

[0072] In this step, the probability of the first target error type corresponding to the main target error type is adjusted based on the first deviation between the second traffic volume and the first forecast traffic volume corresponding to the main target error type, thereby obtaining the first error type probability adjustment value.

[0073] The adjustment probability of the first error type is positively correlated with the first deviation.

[0074] Preferably, appropriate adjustment rules can be determined based on expert experience to complete the adjustment operation of the first error type adjustment probability.

[0075] Preferably, a second error type probability adjustment model can be obtained in advance, thereby obtaining the corresponding first error type probability adjustment value based on the first deviation and the first target error type probability. Historical error type adjustment probability sample data can be used for training.

[0076] The first error type probability adjustment value is an adjustment value based on the first target error type probability.

[0077] S34: Determine at least one second error type probability adjustment value based on the first job information, at least one second job information, and the first error type probability adjustment value.

[0078] Wherein, the first job information is job information corresponding to the primary target error type, and the second job information is job information corresponding to the secondary target error type.

[0079] In step S33, the first error type probability adjustment value corresponding to the primary target error type has been determined. In this step, the error type probability of at least one secondary target error type needs to be adjusted by combining the correlation between the first job information and at least one second job information, as well as the first error type probability adjustment value, to obtain at least one second error type probability adjustment value.

[0080] Specifically, the second error type probability adjustment value corresponding to each of the secondary target error types is determined by a pre-trained third error type probability adjustment model.

[0081] To explore and utilize the correlation between job information of different error types, job information contained in at least one of the first key coupling feature vectors in historical data can be used as input data, and the adjustment value corresponding to the main target error type in historical data can be used as input data, and the adjustment value corresponding to each secondary target error type can be used as output data, so as to train and obtain the third error type probability adjustment model.

[0082] S35: Determine at least one second target error type probability based on the first error type probability adjustment value and at least one second error type probability adjustment value.

[0083] In this step, at least one first target error type probability can be used as a basis, and at least one second target error type probability corresponding to each error type can be obtained by adjusting the first error type probability adjustment value and at least one second error type probability adjustment value.

[0084] S4: Train a port video security analysis model based on at least one of the second target error type probabilities.

[0085] In step S3, the probabilities of the second target error type corresponding to each error type have been determined. Since the visual recognition networks adapted to each error category have specific characteristics, a multi-branch model is employed to perform error category identification on the first port in order to better identify the error category. Furthermore, at least one probability of the second target error type is used as the weight of each branch to further improve the accuracy of visual recognition.

[0086] S4 includes the following sub-steps: S41: Set up a shared feature extraction network to extract at least one first basic image feature from real-time video frames.

[0087] In order to train a port video security analysis model, this step requires first extracting image features that are applicable to all three types of models.

[0088] Once the video surveillance system captures real-time video frames of the port's operational area, the system first processes the images through a shared feature extraction network to extract common basic features from the entire video frame that are relevant to all error types, such as: the outline and movement trajectory of operators, the shape and position of containers, the movement characteristics of spreaders, the boundary features of safe areas, and the degree of congestion in the operational area.

[0089] These fundamental features are the common raw data for the identification of the three error types. Once extracted, they are simultaneously fed into three different feature processing branches. This design ensures that the identification of all error types is based on the same visual foundation, avoiding duplication and inconsistency in feature extraction.

[0090] S42: Set up at least one first visual branch network for at least one type of recognition error.

[0091] Since each type of recognition error targets a different scene and visual feature, this step requires setting up a corresponding first visual branch network for each type of recognition error.

[0092] The first visual branch network specifically includes: 1. Operator-Container Coupling Main scene: Container yard area, with containers densely arranged.

[0093] Visual characteristics: Operators are usually located at the rear or side of the container, causing the container to be partially or completely obscured; the container is large (average size 12m × 2.4m × 2.6m), and the obscured area is obvious; the relative position of the operator and the container changes slowly.

[0094] Challenges in identification: Multiple containers overlap in a container yard, making it difficult to distinguish the target container; when operators are behind containers, their faces and bodies are mostly obscured; the degree of obscuration is related to the operator's position and the arrangement of containers.

[0095] Specific identification methods: 1) In the HSV color space, separate the operator (orange / yellow safety clothing) from the container (gray / silver); 2) Threshold adjustment: Adjust the color space segmentation threshold; 3) Overlapping area detection: Calculate the overlap rate of the operator's and container's outlines; Overlap rate threshold: >30% is considered coupling.

[0096] 2. Operator-spreading device coupling Main scenario: Loading and unloading area, where lifting equipment moves frequently.

[0097] Visual characteristics: The lifting device moves at a high speed with a fixed trajectory; the operator is positioned along the lifting device's trajectory and is obscured by the device; the relative position of the operator and the lifting device changes rapidly. Challenges in identification: The spreader moves at high speed and the detection window is short; the operator may briefly enter the trajectory during the spreader's movement; the obstruction time is short.

[0098] Specific identification methods: Motion trajectory tracking: The SORT target tracking algorithm is used to track the operator and the lifting device separately; Motion model parameter adjustment: For example, from 0.50 to 0.47.

[0099] 3. Operator-safety area coupling Main scenarios: safety zone boundaries, such as hazardous areas and equipment operation areas.

[0100] Visual characteristics: The boundaries of the safe zone are usually marked in yellow, but may be blurred due to lighting and weather conditions; operators approach or enter the boundary area; operators stay in the boundary area for an extended period of time.

[0101] Challenges in identification: The boundaries of the safe zone are unclear (affected by weather and lighting); operators may linger near the boundaries instead of entering directly; the ambiguity of the boundaries is related to environmental factors.

[0102] Specific identification methods: Boundary detection and ambiguity assessment: The Canny edge detection algorithm is used to detect the boundary of the safe area; Boundary ambiguity calculation: The difference between yellow and the surrounding colors in HSV space; Boundary expansion coefficient adjustment: For example, from 50.0 to 49.26.

[0103] S43: Determine the basic model for port video security analysis based on the shared feature extraction network and at least one of the first visual branch networks.

[0104] The port video security analysis basic model can adopt a conventional multi-branch network structure in the existing technology, preferably the Inception network.

[0105] The shared feature extraction network is used to extract shared basic features, and specific recognition error types are identified based on the structure of each of the first visual branch networks.

[0106] S44: Perform a weight adjustment operation on at least one of the first visual branch networks based on at least one second target error type probability to obtain a port video security analysis model.

[0107] Specifically, in the forward function of the Inception module, each branch output can be multiplied by a preset weight before splicing.

[0108] Preferably, after training and obtaining the port video security analysis model, the method further includes performing security intelligent analysis operations on the video of the first port based on the port video security analysis model to identify various error types.

[0109] This application also proposes a port safety intelligent video analysis system for performing the aforementioned port safety intelligent video analysis method.

[0110] This application proposes a port safety intelligent video analysis method and system, relating to the field of visual analysis technology. First, based on historical operational information of a first port, at least one first key coupling feature vector is analyzed to characterize the association between operational data and error types. Then, based on the deviation between historical traffic volume and actual predicted traffic volume, the probability of occurrence of various error events is adjusted. Finally, a multi-branch port video safety analysis model is set up for each error type, and the weights of each branch in the port video safety analysis model are adjusted based on the probability of each error. Through the technical solution of this application, the correlation between various port operational data and the probability of occurrence of error types can be derived based on big data analysis, thereby enabling the trained port video safety analysis model to perform accurate safety analysis and processing operations on port videos.

[0111] The above description is only a preferred embodiment of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the claims of this patent application are included in the scope of this patent application.

Claims

1. A port safety intelligent video analysis method, characterized in that, The method includes: S1: Determine at least one first key coupling feature vector based on the first business identification data sequence of the first port within the first preset time interval; S2: Based on at least one of the first key coupling feature vectors and historical business data, predict the probability of at least one first target error type corresponding to each work area within a future second preset time interval; S3: Perform an association adjustment process on at least one of the first target error type probabilities to obtain at least one second target error type probability; S4: Train a port video security analysis model based on the probability of at least one of the second target error types; S4 includes: S41: Set up a shared feature extraction network to extract at least one first basic image feature from real-time video frames; S42: Set up at least one first visual branch network for each of at least one type of recognition error; S43: Determine the basic model for outbound port video security analysis based on the shared feature extraction network and at least one of the first visual branch networks; S44: Perform a weight adjustment operation on at least one of the first visual branch networks based on at least one second target error type probability to obtain a port video security analysis model.

2. The intelligent video analysis method for port safety according to claim 1, characterized in that, S1 includes: S11: Collect the first business identification data sequence of the first port within the first preset time interval; S12: Perform data saliency analysis on the first business identification data sequence to obtain at least one target identification error dataset; S13: Determine at least one first key coupling feature vector based on at least one of the target recognition error datasets.

3. The intelligent video analysis method for port safety according to claim 2, characterized in that, S12 includes: S121: Extract at least one first identification error dataset from the first service identification data sequence; S122: For each of the first identification error datasets, obtain at least one first salient data feature; S123: Obtain at least one target identification error dataset based on at least one of the first identification error datasets and at least one of the first salient data features.

4. The intelligent video analysis method for port safety according to claim 1, characterized in that, S2 includes: S21: Based on at least one of the first key coupling feature vectors and historical business data, predict the first business volume and at least one first error type probability of each operating area of ​​the first port in the future second preset time interval. S22: For each work area, based on the first traffic volume and the actual forecast traffic volume within the second preset time interval in the future, adjust at least one of the first error type probabilities to obtain at least one first target error type probability.

5. The intelligent video analysis method for port safety according to claim 4, characterized in that, S3 includes: S31: For each work area, determine the primary target error type and at least one secondary target error type based on the probability of at least one of the first target error types; S32: Determine the second traffic volume and the first forecast traffic volume based on the main target error type; S33: Based on the first deviation between the second traffic volume and the first forecast traffic volume, determine the first error type probability adjustment value corresponding to the main target error type; S34: Determine at least one second error type probability adjustment value based on the first job information, at least one second job information, and the first error type probability adjustment value; S35: Determine at least one second target error type probability based on the first error type probability adjustment value and at least one second error type probability adjustment value.

6. The intelligent video analysis method for port safety according to claim 5, characterized in that, The first job information is job information corresponding to the primary target error type, and the second job information is job information corresponding to the secondary target error type.

7. The intelligent video analysis method for port safety according to claim 6, characterized in that, The first error type probability adjustment value is an adjustment value based on the first target error type probability.

8. The intelligent video analysis method for port safety according to claim 1, characterized in that, The basic model for port video security analysis is the Inception network.

9. A port safety intelligent video analysis system, used to implement the port safety intelligent video analysis method according to any one of claims 1-8.

Images