Real-time analysis and dynamic warning method of dangerous behaviors based on deep learning

Abstract

The present invention belongs to the field of dangerous behavior analysis, and discloses a real-time analysis and dynamic alarm method of dangerous behavior based on deep learning. When collecting ship personnel behavior data, the present invention obtains the reference image group corresponding to each monitoring image based on the predefined behavior reference image group selection principle, and then obtains the corresponding personnel behavior data of each monitoring image, which can accurately match the reference image group corresponding to each monitoring image, so as to efficiently and accurately obtain the personnel behavior data in each monitoring image. The accuracy and efficiency of behavior data recognition are significantly improved. When identifying the severity level of personnel dangerous behavior, the present invention analyzes the evaluation of personnel dangerous behavior in each risk image and the environmental danger level of the target ship at the corresponding collection time, which can more comprehensively evaluate the actual degree of harm of dangerous behavior, improve the accuracy and scientificity of identification, and provide a more reliable basis for ship safety management and emergency response.

Description

Technical Field

[0001] The present invention belongs to the technical field of dangerous behavior analysis, and relates to a real-time analysis and dynamic warning method of dangerous behaviors based on deep learning. Background Art

[0002] The hazards of dangerous behaviors on ships are extremely serious. It not only threatens the personal safety of crew members, but also may pose a major threat to the normal operation and navigation safety of ships. During the navigation of ships, timely and accurate identification of dangerous behaviors is an important prerequisite and basis for providing a good driving environment and safe driving conditions. Therefore, the research on real-time analysis and dynamic alarm of dangerous behaviors based on deep learning is of great significance.

[0003] Existing dangerous behavior recognition methods usually confirm the existence of dangerous behaviors through the behavior recognition method of static images, but do not recognize the existence of dynamic dangerous behaviors through the image group comparison method. This analysis method can only capture the scene at a certain moment and cannot reflect the time series and dynamic changes of the behavior. Therefore, it cannot provide key information about the development, duration and change trend of the behavior. This limits the comprehensive and accurate understanding of dangerous behaviors.

[0004] The existing dangerous behavior identification ignores the consideration of the ship's operating environment for corrective analysis when analyzing the severity of dangerous behaviors. This analysis method is difficult to fully and accurately reflect the actual degree of harm caused by dangerous behaviors, which may lead to misjudgment and underestimation of risks, affecting the safety of ship operation and the scientific rationality of decision-making. Summary of the invention

[0005] In view of this, in order to solve the problems raised in the above background technology, a real-time analysis and dynamic alarm method of dangerous behaviors based on deep learning is proposed.

[0006] The purpose of the present invention can be achieved through the following technical solutions: a real-time analysis and dynamic alarm method of dangerous behaviors based on deep learning, including: S1. Analysis of personnel gathering: using a high-definition camera to collect video of the target ship deck, and then selecting a number of monitoring images according to a predefined monitoring image selection principle, and analyzing the evaluation of the degree of personnel gathering in each monitoring image.

[0007] S2. Personnel behavior data collection: based on the predefined behavior reference image group selection principle, obtain the reference image group corresponding to each monitoring image, and obtain the personnel behavior data corresponding to each monitoring image, wherein the personnel behavior data includes the evaluation index of the change of the personnel torso inclination angle, the evaluation index of the movement speed of the personnel limbs and the evaluation index of the change of the personnel spacing.

[0008] S3. Identification of dangerous behaviors of personnel: Analyze the evaluation of dangerous behaviors of personnel in each monitoring image based on the personnel behavior data of each monitoring image, and judge whether there is dangerous behavior of personnel at the collection time corresponding to each monitoring image based on the evaluation of the degree of personnel gathering. If so, issue a warning and record the monitoring image as a risk image.

[0009] S4. Ship environmental hazard analysis: collect safety data on the target ship deck, including rainfall, heave distance, heave period and center offset distance, and analyze the environmental hazard of the target ship at the time of collection of each risk image.

[0010] S5. Identification of severity level of dangerous behaviors of personnel: Based on the evaluation of dangerous behaviors of personnel in each risk image and the environmental hazard level of the target ship at the corresponding collection time, the corrected evaluation of dangerous behaviors of personnel at the corresponding collection time of each risk image is analyzed, and the severity level of dangerous behaviors of personnel is identified.

[0011] Compared with the prior art, the present invention has the following beneficial effects: (1) When collecting ship personnel behavior data, the present invention obtains the reference image group corresponding to each monitoring image based on the predefined behavior reference image group selection principle, and then obtains the corresponding personnel behavior data of each monitoring image. This analysis method can accurately match the reference image group corresponding to each monitoring image, thereby efficiently and accurately obtaining the personnel behavior data in each monitoring image. This analysis method significantly improves the accuracy and efficiency of behavior data recognition, and provides strong support for ship safety monitoring and personnel behavior analysis.

[0012] (2) When identifying the severity level of dangerous behaviors of personnel, the present invention performs analysis based on the evaluation of dangerous behaviors of personnel in each risk image and the environmental danger level of the target ship at the corresponding acquisition time. This analysis method can more comprehensively evaluate the actual degree of harm of dangerous behaviors, improve the accuracy and scientificity of identification, and provide a more accurate and reliable basis for ship safety management and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0014] Figure 1 It is a schematic diagram for implementing the steps of the method of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] See also Figure 1 As shown, the present invention provides a real-time analysis and dynamic alarm method for dangerous behaviors based on deep learning, the method comprising: S1, personnel gathering analysis: using a high-definition camera to collect the deck video of the target ship, and then selecting a number of monitoring images according to a predefined monitoring image selection principle, and analyzing the personnel gathering degree evaluation of each monitoring image.

[0017] In a preferred embodiment of the present invention, the predefined monitoring image selection principle is specifically as follows: extracting the target ship deck video collected by a high-definition camera, and then performing image extraction according to a preset reference interval frame number to obtain a number of target ship deck images, and then performing personnel presence judgment on each target ship deck image to obtain the number of personnel in each target ship deck image; if the number of personnel on a target ship deck is greater than the preset number of monitoring personnel, the target ship deck image is recorded as a ship deck image to be analyzed.

[0018] It should be noted that when selecting monitoring images, the reason for extracting images according to the preset reference interval frame number is: 1. In a continuous video stream, the image content between adjacent frames is often highly similar, especially in static or slowly changing scenes. By setting a reasonable interval frame number for image extraction, the amount of redundant data can be significantly reduced, thereby reducing storage and processing costs. 2. Extracting images with interval frame numbers can reduce the total amount of images that need to be processed, thereby increasing the speed of image processing and analysis.

[0019] Exemplarily, the reference interval frame number is .

[0020] It should be noted that the reason for conducting personnel quantity analysis is that the purpose of the present invention is to conduct personnel dangerous behavior analysis, which usually involves the interaction of two or more persons, so the number of personnel needs to be screened.

[0021] For example, the number of monitoring personnel is .

[0022] Arrange the ship deck images to be analyzed in chronological order, and divide them into several ship deck image groups to be analyzed according to continuity. It is determined whether the number of deck images to be analyzed corresponding to each deck image group to be analyzed is greater than the number of preset reference images. If the number of deck images to be analyzed corresponding to a deck image group to be analyzed is greater than or equal to the number of reference images, the deck images to be analyzed corresponding to the deck image group to be analyzed are recorded as monitoring images, and then the monitoring images are obtained by summarizing them.

[0023] It should be noted that if the number of deck images to be analyzed corresponding to a certain deck image group to be analyzed is less than the number of reference images, it is judged that the deck images to be analyzed corresponding to the deck images to be analyzed in the deck image group to be analyzed do not meet the conditions for dangerous behavior analysis, and no dangerous behavior analysis is required.

[0024] In a preferred embodiment of the present invention, the analysis of the evaluation of the degree of personnel gathering of each monitoring image requires the construction of a personnel gathering degree evaluation index for each monitoring image, and the specific method is as follows: extract each monitoring image selected according to a predefined monitoring image selection principle, and locate the position of personnel in each monitoring image.

[0025] Select any personnel position as the benchmark personnel position, and record the positions of all personnel except the benchmark personnel position as other personnel positions, obtain the distance between the benchmark personnel position and each other personnel position, record it as the monitoring distance between the benchmark personnel and each other personnel, compare the monitoring distance between the benchmark personnel and each other personnel with the pre-set reference distance respectively to obtain the monitoring concentration of the benchmark personnel and each other personnel, and then perform the mean calculation to obtain the monitoring concentration corresponding to the benchmark personnel.

[0026] Similarly, the monitoring concentration corresponding to each other personnel position as the benchmark personnel position is obtained, and then the mean value is calculated to obtain the personnel concentration evaluation index of each monitoring image.

[0027] It should be supplemented that the analysis method for obtaining the monitoring concentration of the benchmark person and each other person is as follows: extracting the monitoring distance corresponding to the benchmark person and each other person, performing difference calculation with the preset reference distance respectively to obtain the monitoring distance deviation between the benchmark person and each other person, and then performing ratio calculation with the preset reference distance to obtain the monitoring distance abnormality between the benchmark person and each other person, and then taking the inverse to obtain the monitoring concentration of the benchmark person and each other person.

[0028] It should be noted that, because the positions of various personnel will not overlap in practice, the monitoring distance anomaly is not 0.

[0029] S2. Personnel behavior data collection: based on the predefined behavior reference image group selection principle, obtain the reference image group corresponding to each monitoring image, and obtain the personnel behavior data corresponding to each monitoring image, wherein the personnel behavior data includes the evaluation index of the change of the personnel torso inclination angle, the evaluation index of the movement speed of the personnel limbs and the evaluation index of the change of the personnel spacing.

[0030] It should be noted that the reason for selecting the evaluation index of the change in the inclination angle of the trunk, the evaluation index of the speed of movement of the limbs, and the evaluation index of the change in the distance between people as the behavior data is that the evaluation index of the change in the inclination angle of the trunk can capture the changes in body posture that are the precursors of a fight, the evaluation index of the speed of movement of the limbs can accurately reflect the intensity of attack and defense, and the evaluation index of the change in the distance between people can help to warn of the escalation of conflicts. These behavioral data together constitute a comprehensive framework for the analysis of dangerous fighting behaviors, providing an accurate and scientific basis for timely identification, warning, and taking effective measures to stop violent conflicts, which is of great significance for maintaining public safety and reducing injury incidents.

[0031] In a preferred embodiment of the present invention, the predefined behavior reference image group selection principle is specifically: a preset number of reference images located before the moment of collecting fragments of each monitoring image are recorded as the reference image group corresponding to the monitoring image, and the reference image group includes several reference images.

[0032] Exemplarily, the number of reference images is 6.

[0033] It needs to be explained that the reason for setting a limit on the number of reference images is to fully consider the actual progress of human behavior and the analysis of possible misjudgments in determining the existence of dangerous behavior. Misjudgments may occur in practice.

[0034] It should be noted that the present invention obtains the reference image group corresponding to each monitoring image based on the predefined behavior reference image group selection principle when collecting ship personnel behavior data, and then obtains the corresponding personnel behavior data of each monitoring image. This analysis method can accurately match the reference image group corresponding to each monitoring image, thereby efficiently and accurately obtaining the personnel behavior data in each monitoring image. This analysis method significantly improves the accuracy and efficiency of behavior data recognition, and provides strong support for ship safety monitoring and personnel behavior analysis.

[0035] S3. Identification of dangerous behaviors of personnel: Analyze the evaluation of dangerous behaviors of personnel in each monitoring image based on the personnel behavior data of each monitoring image, and judge whether there is dangerous behavior of personnel at the collection time corresponding to each monitoring image based on the evaluation of the degree of personnel gathering. If so, issue a warning and record the monitoring image as a risk image.

[0036] In a preferred embodiment of the present invention, the analysis of the evaluation of dangerous behaviors of personnel in each monitoring image requires the construction of an evaluation index of dangerous behaviors of personnel in each monitoring image, and the specific method is as follows: extracting the evaluation index of the change of the inclination angle of the personnel trunk, the evaluation index of the movement speed of the personnel limbs and the evaluation index of the change of the personnel spacing corresponding to each monitoring image, and recording them as , , ,in Indicates the number of the monitoring image. , Indicates the number of monitored images.

[0037] It should be noted that the analysis method of the evaluation index of the change in the trunk inclination angle of a person is as follows: the trunk inclination angle of a person corresponding to each person in each monitoring image in the reference image group corresponding to each monitoring image is extracted, and then the mean calculation is performed to obtain the reference trunk inclination angle of a person corresponding to each person in the monitoring image, and the trunk inclination angle of a person corresponding to each person in the monitoring image is obtained, and the difference calculation of the trunk inclination angle of a person corresponding to each person in the monitoring image and the corresponding reference trunk inclination angle is performed to obtain the deviation of the trunk inclination angle of each person in the monitoring image, and the absolute value of the deviation of the trunk inclination angle of each person in the monitoring image is calculated by ratio with the corresponding reference trunk inclination angle to obtain the evaluation index of the change in the trunk inclination angle of each person in the monitoring image, and then the mean calculation is performed to obtain the evaluation index of the change in the trunk inclination angle of the person corresponding to the monitoring image.

[0038] It needs to be explained that the inclination angle of the person's torso refers to the angle between the direction of the person's upper body and the deck plane.

[0039] It should be noted that the analysis method of the evaluation index of the speed of movement of the limbs of a person and the evaluation index of the change in the distance between persons refers to the analysis method of the evaluation index of the change in the inclination angle of the torso of a person.

[0040] It needs to be further explained that the speed of movement of a person's limbs refers to the larger of the speeds of movement of the person's hands and feet. Specifically, if the speed of movement of a person's hands is less than the speed of movement of their feet, the speed of movement of their feet will be taken as the speed of movement of the person's limbs.

[0041] It needs further explanation that the spacing between people refers to the minimum spacing between people.

[0042] Using the formula Analyze and obtain the personnel risk behavior evaluation index of each monitoring image ,in They respectively represent the weight factors corresponding to the evaluation index of the change in the inclination angle of the personnel's torso, the evaluation index of the movement speed of the personnel's limbs, and the evaluation index of the change in the distance between personnel.

[0043] It should be noted that the weight factors corresponding to the evaluation index of the change in the inclination angle of the torso, the evaluation index of the speed of movement of the limbs, and the evaluation index of the change in the distance between people are set based on the following: the importance of each indicator for early warning and evaluation of fighting behavior needs to be carefully considered, while combining the actual performance characteristics of the data, the professional judgment of experts on such behavior, and the actual situation in past fighting incidents. By using quantitative methods such as factor analysis and correlation coefficient calculation, as well as qualitative methods such as the Delphi method and expert scoring, we can ensure that each evaluation index is assigned a scientific and reasonable weight, so as to more accurately identify the potential risks of fighting behavior, take timely measures to intervene, and effectively ensure public safety.

[0044] For example, .

[0045] In a preferred embodiment of the present invention, the specific process of judging whether there is any dangerous behavior of personnel at the acquisition moment corresponding to each monitoring image is as follows: extracting the dangerous behavior evaluation index of personnel and the personnel gathering degree evaluation index of each monitoring image, and then summing them up according to the weights to obtain the comprehensive evaluation index of dangerous behavior of personnel in each monitoring image.

[0046] For example, the weights corresponding to the personnel risk behavior evaluation index and the personnel gathering degree evaluation index are: .

[0047] The comprehensive evaluation index of dangerous behavior of personnel in each monitoring image is compared with the preset threshold of comprehensive evaluation index of dangerous behavior of personnel. If the comprehensive evaluation index of dangerous behavior of personnel in a monitoring image is greater than the threshold of comprehensive evaluation index of dangerous behavior of personnel, it is judged that dangerous behavior of personnel exists in the monitoring image.

[0048] It should be noted that the thresholds of the comprehensive evaluation index of dangerous behaviors of personnel are set based on the following criteria: The thresholds of the comprehensive evaluation index of dangerous behaviors of personnel are set based on behavioral characteristics analysis, historical data comparison, expert experience judgment, public safety standards and other aspects. Through in-depth analysis of key behavioral data such as the inclination angle of the torso, the speed of movement of the limbs, and the distance between people, combined with actual data from past dangerous events, and referring to the professional opinions of experts in related fields and general standards in the field of public safety, we can scientifically and reasonably set the thresholds of various evaluation indexes. These thresholds can effectively identify and warn of potential dangerous behaviors, providing an important basis for timely intervention measures and ensuring personnel safety.

[0049] For example, the threshold value of the comprehensive evaluation index of dangerous behavior of personnel is .

[0050] The comprehensive evaluation index of dangerous behavior of personnel in each reference image in the reference image group corresponding to each monitoring image is extracted, and then it is determined whether there is dangerous behavior of personnel in each reference image.

[0051] If a monitoring image contains dangerous behavior of a person and each reference image in the reference image group corresponding to the monitoring image contains dangerous behavior of a person, it is determined that dangerous behavior of a person exists at the acquisition moment corresponding to the monitoring image.

[0052] S4. Ship environmental hazard analysis: collect safety data on the target ship deck, including rainfall, heave distance, heave period and center offset distance, and analyze the environmental hazard of the target ship at the time of collection of each risk image.

[0053] It should be noted that the reasons for selecting rainfall, heave distance, heave cycle and center offset distance as the safety data for the target ship deck are as follows: 1. Rainfall: Rainfall will make the deck slippery, increasing the risk of slipping when people are walking and handling cargo. It may also affect the electrical equipment on the deck and increase safety hazards such as leakage. 2. Heave distance and heave cycle: These two parameters reflect the movement state of the ship in the waves. A large heave distance and a short cycle mean that the ship is more severely impacted, which may cause damage to the deck structure or displacement of cargo, affecting the stability of the ship. 3. Center offset distance: This reflects the stability of the ship's posture during navigation. A large center offset distance means that the ship may lose balance, leading to serious consequences such as capsizing.

[0054] In a preferred embodiment of the present invention, the analysis of the environmental hazard level of the target ship at the time of collection of each risk image requires constructing the environmental hazard level of the target ship at the time of collection of each risk image. The specific method is as follows: extract the rainfall, fluctuation distance, fluctuation period and center offset distance of the deck of the target ship at the time of collection of each risk image, and record them as , , , ,in The number of the risk image. , Indicates the number of risky images.

[0055] It should be noted that the rainfall on the deck of the target ship can be obtained using a rain gauge, and the fluctuation distance refers to the difference between the horizontal height at the corresponding acquisition time of the current monitoring image and the horizontal height at the corresponding acquisition time of the previous monitoring image. The horizontal height can be obtained using a spirit level.

[0056] It should be noted that the fluctuation period refers to the movement time between the highest position and the lowest position of the target ship, which can be analyzed by collecting data from the level meter. The center offset distance refers to the horizontal distance between the center of the ship corresponding to the current monitoring image and the center of the ship corresponding to the previous monitoring image, which can be measured using radar ranging equipment.

[0057] Based on the heaving distance and center offset distance of the target ship deck at the corresponding acquisition time of each risk image, the stability evaluation index of the target ship deck at the corresponding acquisition time of each risk image is obtained. .

[0058] Using the formula The environmental danger degree of the target ship at the time of acquisition of each risk image is obtained by analysis ,in Indicates the preset reference rainfall. Indicates the preset reference fluctuation period.

[0059] For example, , .

[0060] In a preferred embodiment of the present invention, the specific method of analyzing and obtaining the stability evaluation index of the target ship deck at the time of acquisition of each risk image is as follows: extracting the ups and downs distance of the target ship deck at the time of acquisition of each risk image and center offset distance .

[0061] Using the formula The stability evaluation index of the target ship deck at the time of acquisition of each risk image is obtained by analysis. ,in Indicates the preset ups and downs distance threshold, Indicates the preset center offset distance threshold.

[0062] S5. Identification of severity level of dangerous behaviors of personnel: Based on the evaluation of dangerous behaviors of personnel in each risk image and the environmental hazard level of the target ship at the corresponding collection time, the corrected evaluation of dangerous behaviors of personnel at the corresponding collection time of each risk image is analyzed, and the severity level of dangerous behaviors of personnel is identified.

[0063] In a preferred embodiment of the present invention, the analysis of the corrected personnel dangerous behavior evaluation situation at the corresponding collection moment of each risk image requires the construction of a corrected personnel dangerous behavior evaluation index at the corresponding collection moment of each risk image, and the specific method is as follows: extracting the comprehensive evaluation index of the personnel dangerous behavior of each risk image and the environmental hazard degree of the target ship at the corresponding collection moment.

[0064] The comprehensive evaluation index of personnel dangerous behavior of each risk image is multiplied by the environmental hazard degree of the target ship at the corresponding collection time to obtain the correction amount of the comprehensive evaluation index of personnel dangerous behavior of each risk image, and then the correction amount of the comprehensive evaluation index of personnel dangerous behavior corresponding to each risk image is summed up to obtain the corrected personnel dangerous behavior evaluation index of each risk image at the corresponding collection time.

[0065] It should be noted that when identifying the severity level of dangerous behaviors of personnel, the present invention conducts analysis based on the evaluation of dangerous behaviors of personnel in each risk image and the environmental hazard situation of the target ship at the corresponding acquisition time. This analysis method can more comprehensively evaluate the actual degree of harm of dangerous behaviors, improve the accuracy and scientificity of identification, and provide a more accurate and reliable basis for ship safety management and emergency response.

[0066] In a preferred embodiment of the present invention, the specific method of identifying the severity level of dangerous behavior of personnel is as follows: extracting the modified dangerous behavior evaluation index of each risk image corresponding to the acquisition time, recorded as .

[0067] The modified personnel risk behavior evaluation index of each risk image corresponding to the collection time Substitute personnel dangerous behavior severity evaluation model Get the severity level evaluation index of dangerous behavior of personnel at the time of collection of each risk image ,in Indicates the preset first-level threshold of the personnel dangerous behavior evaluation index. Indicates the preset secondary threshold of the personnel dangerous behavior evaluation index.

[0068] It should be noted that the first-level threshold and the second-level threshold of the personnel dangerous behavior evaluation index are set based on many factors, including behavioral feature analysis, historical data comparison, expert experience judgment, public safety standards, and data volatility analysis. These bases together form the basis for threshold setting and ensure the rationality and scientificity of the threshold. In the present invention, the first-level threshold of the personnel dangerous behavior evaluation index is higher than the second-level threshold of the personnel dangerous behavior evaluation index.

[0069] For example, , .

[0070] like Indicates that the severity level of the dangerous behavior of the personnel at the time of collection of the risk image is level one. Indicates that the severity level of the dangerous behavior of the personnel at the time of collection of the risk image is level 2. It indicates that the severity level of the dangerous behavior of the personnel at the time of collection of the risk image is level three.

[0071] It needs to be explained that among the severity levels corresponding to level one, level two and level three, level one is the most severe, followed by level two, and level three is the lightest.

[0072] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they shall all fall within the protection scope of the present invention.

Claims

1. A real-time analysis and dynamic warning method for dangerous behaviors based on deep learning, characterized in that: include: S1. Analysis of personnel gathering: Use a high-definition camera to collect the target ship deck video, and then select a number of monitoring images according to the predefined monitoring image selection principle, and analyze the personnel gathering evaluation of each monitoring image; S2, personnel behavior data collection: based on the predefined behavior reference image group selection principle, obtain the reference image group corresponding to each monitoring image, and obtain the personnel behavior data corresponding to each monitoring image, wherein the personnel behavior data includes the personnel trunk tilt angle change evaluation index, the personnel limb movement speed evaluation index and the personnel spacing change evaluation index; S3. Identification of dangerous behaviors of personnel: Analyze the evaluation of dangerous behaviors of personnel in each monitoring image based on the personnel behavior data of each monitoring image, and judge whether there is dangerous behavior of personnel at the acquisition time corresponding to each monitoring image based on the evaluation of the degree of personnel gathering. If there is, issue an early warning and record the monitoring image as a risk image; S4. Ship environmental risk analysis: collect safety data on the target ship deck, including rainfall, heave distance, heave period and center offset distance, and analyze the environmental risk of the target ship at the time of collection of each risk image; The analysis of the environmental risk level of the target ship at the time of collection of each risk image requires the construction of the environmental risk level of the target ship at the time of collection of each risk image, and the specific method is as follows: The rainfall, heave distance, heave period and center offset distance of the target ship deck at the time of acquisition of each risk image are extracted and recorded as , , , ,in The number of the risk image. , Indicates the number of risky images; Based on the heaving distance and center offset distance of the target ship deck at the corresponding acquisition time of each risk image, the stability evaluation index of the target ship deck at the corresponding acquisition time of each risk image is obtained. ; Using the formula The environmental hazard degree of the target ship at the time of acquisition of each risk image is obtained by analysis ,in Indicates the preset reference rainfall. Indicates the preset reference fluctuation period; The specific method of analyzing and obtaining the stability evaluation index of the target ship deck at the time of acquisition of each risk image is as follows: Extract the ups and downs distance of the target ship deck at the time of acquisition of each risk image and center offset distance ; Using the formula The stability evaluation index of the target ship deck at the time of acquisition of each risk image is obtained by analysis. ,in Indicates the preset ups and downs distance threshold, Indicates the preset center offset distance threshold; S5. Identification of severity level of dangerous behaviors of personnel: Based on the evaluation of dangerous behaviors of personnel in each risk image and the environmental hazard level of the target ship at the corresponding collection time, the corrected evaluation of dangerous behaviors of personnel at the corresponding collection time of each risk image is analyzed, and the severity level of dangerous behaviors of personnel is identified.

2. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 1, characterized in that: The predefined monitoring image selection principle is specifically as follows: Extract the target ship deck video collected by the high-definition camera, and then extract the image according to the preset reference interval frame number to obtain a number of target ship deck images, and then judge the presence of personnel in each target ship deck image to obtain the number of personnel in each target ship deck image. If the number of personnel on the deck of a target ship is greater than the preset number of monitoring personnel, the target ship deck image is recorded as the ship deck image to be analyzed; Arrange the ship deck images to be analyzed in chronological order, and divide them into several ship deck image groups to be analyzed according to continuity. It is determined whether the number of deck images to be analyzed corresponding to each deck image group to be analyzed is greater than the number of preset reference images. If the number of deck images to be analyzed corresponding to a deck image group to be analyzed is greater than or equal to the number of reference images, the deck images to be analyzed corresponding to the deck image group to be analyzed are recorded as monitoring images, and then the monitoring images are obtained by summarizing them.

3. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 1, characterized in that: The analysis of the personnel gathering degree evaluation of each monitoring image requires the construction of a personnel gathering degree evaluation index for each monitoring image, and the specific method is as follows: Extracting each monitoring image selected according to a predefined monitoring image selection principle, and locating the position of the personnel in each monitoring image; Select any person's position as the benchmark person's position, and record the positions of all persons except the benchmark person's position as other person's positions, obtain the distance between the benchmark person's position and the positions of all other persons, record it as the monitoring distance between the benchmark person and all other persons, compare the monitoring distance between the benchmark person and all other persons with the preset reference distance respectively to obtain the monitoring concentration of the benchmark person and all other persons, and then perform mean calculation to obtain the monitoring concentration corresponding to the benchmark person; Similarly, the monitoring concentration corresponding to the positions of other personnel is obtained as the benchmark personnel position, and then the mean value is calculated to obtain the personnel concentration evaluation index of each monitoring image.

4. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 1, characterized in that: The specific principle for selecting the predefined behavior reference image group is: A preset number of reference images located before the moment of collecting the fragments of each monitoring image is recorded as a reference image group corresponding to the monitoring image, and the reference image group includes a plurality of reference images.

5. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 3, characterized in that: The analysis of the evaluation of dangerous behaviors of personnel in each monitoring image requires the construction of an evaluation index of dangerous behaviors of personnel in each monitoring image, and the specific method is as follows: The evaluation index of the change of the trunk tilt angle, the evaluation index of the movement speed of the limbs and the evaluation index of the change of the spacing between people corresponding to each monitoring image are extracted and recorded as , , ,in Indicates the number of the monitoring image. , Indicates the number of monitored images; Using the formula Analyze and obtain the personnel risk behavior evaluation index of each monitoring image ,in They respectively represent the weight factors corresponding to the evaluation index of the change in the inclination angle of the personnel's torso, the evaluation index of the movement speed of the personnel's limbs, and the evaluation index of the change in the distance between personnel.

6. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 5, characterized in that: The specific process of determining whether there is a dangerous behavior of a person at the time of collection corresponding to each monitoring image is as follows: Extract the personnel dangerous behavior evaluation index and personnel gathering degree evaluation index of each monitoring image, and then sum them up according to the weight to obtain the comprehensive evaluation index of personnel dangerous behavior of each monitoring image; Compare the comprehensive evaluation index of dangerous behaviors of personnel in each monitoring image with the preset threshold value of comprehensive evaluation index of dangerous behaviors of personnel. If the comprehensive evaluation index of dangerous behaviors of personnel in a monitoring image is greater than the threshold value of comprehensive evaluation index of dangerous behaviors of personnel, it is judged that dangerous behaviors of personnel exist in the monitoring image. Extracting the comprehensive evaluation index of dangerous behaviors of personnel in each reference image in the reference image group corresponding to each monitoring image, and then judging whether there is dangerous behaviors of personnel in each reference image; If a monitoring image contains dangerous behavior of a person and each reference image in the reference image group corresponding to the monitoring image contains dangerous behavior of a person, it is determined that dangerous behavior of a person exists at the acquisition moment corresponding to the monitoring image.

7. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 1, characterized in that: The analysis of the evaluation of the corrected personnel dangerous behavior at the time of collection of each risk image requires the construction of the corrected personnel dangerous behavior evaluation index at the time of collection of each risk image, and the specific method is as follows: Extract the comprehensive evaluation index of dangerous behaviors of personnel in each risk image and the environmental risk degree of the target ship at the corresponding acquisition time; The comprehensive evaluation index of personnel dangerous behavior of each risk image is multiplied by the environmental hazard degree of the target ship at the corresponding collection time to obtain the correction amount of the comprehensive evaluation index of personnel dangerous behavior of each risk image, and then the correction amount of the comprehensive evaluation index of personnel dangerous behavior corresponding to each risk image is summed up to obtain the corrected personnel dangerous behavior evaluation index of each risk image at the corresponding collection time.

8. The method for real-time analysis and dynamic warning of dangerous behaviors based on deep learning as claimed in claim 7, characterized in that: The specific method for identifying the severity level of dangerous behavior of personnel is as follows: Extract the modified personnel risk behavior evaluation index corresponding to the acquisition time of each risk image, denoted as ; The modified personnel risk behavior evaluation index of each risk image corresponding to the collection time Substitute personnel dangerous behavior severity evaluation model Get the severity level evaluation index of dangerous behavior of personnel at the time of collection of each risk image ,in Indicates the preset first-level threshold of the personnel dangerous behavior evaluation index. Indicates the preset secondary threshold of the personnel dangerous behavior evaluation index; like Indicates that the severity level of the dangerous behavior of the personnel at the time of collection of the risk image is level one. Indicates that the severity level of the dangerous behavior of the personnel at the time of collection of the risk image is level 2. It indicates that the severity level of the dangerous behavior of the personnel at the time of collection of the risk image is level three.

Images