A personnel intrusion detection method based on an artificial intelligence loading system

Abstract

The present application relates to the technical field of process monitoring, in particular to a personnel intrusion detection method based on an artificial intelligence loading system, comprising the following steps: acquiring a target continuous motion image sequence, calculating a node included angle and a direction deflection to generate positioning information, comparing a work boundary to determine an intrusion and generate an identification state, combining a direction vector to determine a device response instruction, loading a trajectory in a three-dimensional scene and marking a risk area, and obtaining a personnel path intrusion calibration record. In the present application, through continuous extraction of a human body structure boundary box in an image frame and dynamic displacement analysis of a trajectory node, a target path with continuous gait characteristics is identified and static and device interference areas are eliminated, whether the target has crossed the boundary is determined in combination with a real-time loading work boundary, a minimum distance between a turning point and a boundary path is calculated to complete intrusion behavior identification, real-time discrimination of a behavior trajectory and efficient linkage of alarm direction control are realized, and the accurate identification and real-time reaction capability of personnel intrusion events are enhanced.

Description

Technical Field

[0001] This invention relates to the field of process monitoring technology, and in particular to a method for detecting unauthorized personnel entry into a loading system based on artificial intelligence. Background Technology

[0002] The field of process monitoring technology involves monitoring, managing, and scheduling key aspects of business processes in industries such as manufacturing, logistics, and logistics. Core aspects include real-time acquisition of process execution status, automatic identification of process anomalies, dynamic early warning of critical events, visualization, and traceability. This technology is widely used in automated operation systems, employing sensors, video recognition, artificial intelligence, and digital twins to provide comprehensive dynamic monitoring of personnel, equipment, and materials in the operating environment, ensuring process safety, improving operational efficiency, and achieving process controllability. Traditional personnel intrusion detection methods, used in train loading operations to identify and determine whether personnel have illegally entered the work area, primarily rely on fixed-position video surveillance equipment for image capture, combined with image recognition methods based on background subtraction or edge detection to analyze and identify human silhouettes appearing in the footage. The system uses a static camera to capture video and determines whether there is movement by observing changes between image frames. It then combines this with simple human feature comparison to identify personnel targets. However, the identification mechanism relies on preset image model parameters and lacks adaptability to dynamically changing scenes. It has low accuracy in complex situations such as sudden intrusion by personnel or re-entry after equipment malfunctions. This makes it difficult to meet the requirements for high-sensitivity identification and real-time response to personnel intrusion in unmanned loading scenarios.

[0003] Existing technologies use fixed video surveillance equipment for image acquisition, judging whether there is object movement by static changes between image frames, and combining simple human feature recognition for target identification. Because the recognition mechanism relies on preset image model parameters, it lacks the ability to adapt to dynamic changes in the scene and cannot accurately analyze behavioral features in continuous motion trajectories. In scenarios such as sudden intrusion by personnel or re-entry after changes in equipment status, image recognition is easily affected by interference from operating equipment or scene occlusion, resulting in a high misjudgment rate of personnel target outlines. It also has the limitation of not being able to distinguish between normal and abnormal entry paths, affecting the timeliness and accuracy of recognition, and making it difficult to meet the operational requirements of highly sensitive recognition and real-time response to sudden intrusion behavior in unmanned scenarios. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a method for detecting unauthorized personnel in a vehicle loading system based on artificial intelligence, comprising the following steps:

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting unauthorized personnel in a vehicle loading system based on artificial intelligence, comprising the following steps:

[0006] S1: Obtain the sequence of monitoring camera images of the loading operation area, extract the human body structure bounding box region in each frame, and generate a continuous moving target image sequence by comparing the position offset value of the same target in adjacent frames through the morphological structure points within the pixel region.

[0007] S2: Based on the image sequence of the continuously moving target, read the coordinate values ​​of the nodes and calculate the angle change formed by every three consecutive nodes. Determine whether there is a direction deviation phenomenon by the change amplitude of the angle and the change trend of the connecting direction vector, and generate path offset behavior positioning information.

[0008] S3: Based on the path offset behavior positioning information, read the coordinate range of the dynamic operation boundary line segment of the real-time loading equipment, perform spatial relationship judgment between the path nodes and the operation boundary range, calculate the nearest distance value between the turning point and the path segment within the boundary, and generate the intrusion target identification status within the operation range;

[0009] S4: Based on the intrusion target identification status within the operating range, read the image frame number corresponding to the intrusion target location node, combine the target motion direction vector in the image frame, calculate the angle range of the target motion direction vector relative to the audible and visual alarm device in the two-dimensional plane, and obtain the device response direction instruction set.

[0010] As a further aspect of the present invention, the continuously moving target image sequence includes a target identification identifier, a continuous path node sequence, and a time frame index; the path offset behavior positioning information includes an offset frame number, offset point coordinates, and offset angle value; the intrusion target identification status within the operating range includes an intrusion target number, intrusion node location, intrusion trigger time, and shortest path distance; and the device response direction instruction set includes a device number, response angle, rotation direction, and action range.

[0011] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0012] S101: Acquire the sequence of monitoring camera images of the loading operation area, perform structured processing on each frame, extract the bounding box region of the human structure defined by rectangles from the image pixel data, construct the bounding box coordinate matrix, and perform joint matching based on the gray-scale gradient value of the human structure outline edge in the bounding box and the color distribution histogram. Call the timestamp information to compare the spatial position information offset value of the bounding box target in adjacent frames, and generate the position offset measurement result.

[0013] S102: Call the position offset measurement results, reconstruct the path sequence according to the spatial displacement vector corresponding to the same bounding box identifier in the continuous frames, detect human targets in the path sequence whose displacement amplitude is greater than the displacement detection reference value for more than three consecutive frames, merge and cluster the human target displacement paths that meet the conditions, remove human target indexes and path data with insufficient motion amplitude, and obtain a continuous movement path identifier set.

[0014] S103: Call the continuous motion path identifier set, retrieve the bounding box image region in the corresponding original image frame sequence that matches the timestamp of the motion path node, extract the region pixels in sequence to form a sequence, and perform sequential numbering and inter-frame registration operations on the sequence image regions to generate a continuous motion target image sequence.

[0015] As a further aspect of the present invention, the displacement detection benchmark value is specifically limited to a similarity score threshold formed by jointly matching the grayscale change gradient value of the human body structure edge line in the bounding box with the color distribution histogram. The similarity score threshold is between 0.65 and 0.85. It is selected based on the principle of minimum mismatch rate by statistically analyzing the human target trajectory data under differentiated movement states in the monitoring camera image frame sequence of the batch loading operation area.

[0016] The spatial displacement vector is defined as the combination of coordinate increments of the center point of the bounding box coordinate matrix in adjacent image frames along the X and Y axes on the two-dimensional image plane. The coordinate increments must satisfy that the Euclidean distance is not less than the pixel distance value converted from the displacement detection reference value.

[0017] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0018] S201: Call the continuously moving target image sequence, read the coordinate values ​​of multiple path nodes in the image frame, arrange the coordinate points in chronological order, extract the coordinate value group of three consecutive nodes, calculate the included angle value of the path polyline, and generate the path polyline included angle sequence.

[0019] S202: Call the path polyline angle sequence, perform difference calculation on adjacent angle values, and combine the difference with the direction vector change value for judgment. The direction vector change value is calculated by the angle difference between the direction unit vectors formed by the connecting lines of two adjacent nodes. If both the angle difference and the direction vector change value exceed the angle deflection judgment benchmark value, it is determined that a direction deflection has occurred at the corresponding node. Record the frame number and node index that meet the deflection condition, and generate a direction deflection event record table.

[0020] S203: Call the direction deflection event record table, call the position data of the deflection frame number and node index, extract the corresponding coordinate point information from the original path node coordinate sequence, and combine the frame number, node position and direction change value to establish path offset behavior positioning information.

[0021] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0022] S301: Call the path offset behavior positioning information, read the coordinate range data of the dynamic operation boundary line segment of the real-time loading equipment, and make the coordinate values ​​of the direction deflection node consistent with the spatial position data of the boundary line segment. Determine whether the node position is inside the boundary by the spatial relationship between the coordinate point and the line segment area, and index and mark the nodes that meet the spatial relationship to generate a path node entry boundary identifier set.

[0023] S302: Call the path node inbound boundary identifier set, extract the coordinate values ​​of the turning points corresponding to the marked nodes, and retrieve the coordinate values ​​of the continuous nodes that are still within the boundary range. Calculate the shortest distance between the turning point and each path segment to generate a path turning to boundary segment distance sequence.

[0024] S303: Call the path to the boundary segment distance sequence, perform difference calculation and numerical comparison between the shortest distance value and the set intrusion identification threshold. If the shortest distance value is lower than the intrusion identification threshold, record the corresponding path identifier and node position as the intrusion target to obtain the intrusion target identification status within the operation range.

[0025] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0026] S401: Call the intrusion target identification status within the operation range, read the recorded image frame number and corresponding node coordinate information, extract the spatial position coordinates of the intrusion target in two consecutive frames, construct the two-dimensional motion direction vector of the target, calculate the angle between the motion direction vector and the installation position vector of the sound and light alarm device in the real-time image frame, establish the angle direction range in combination with the device plane projection direction information, and generate the relative angle interval of the target motion direction.

[0027] S402: Call the relative angle range of the target movement direction, read the direction vector of the offset line segment in the running path of the loading equipment, compare the regional overlap between the direction vector and the angle range, filter the path segments that intersect with the target movement direction at an angle, calculate the angle difference and deflection angle value between the path segment and the direction of the alarm device, and combine the time index to mark the direction number that needs to be intervened to obtain the equipment response angle reference value set;

[0028] S403: Call the device response angle reference value set, map the device response angle reference values ​​to the corresponding direction command numbers according to the device direction encoding order, and combine the direction numbers with the corresponding deflection angle parameters to generate the device response direction command set.

[0029] As a further aspect of the present invention, the process of generating the relative angle range of the target motion direction is specifically defined as follows: when the angle between the motion direction vector and the installation position vector of the audible and visual alarm device is in the range of 0° to 90°, the angle direction range is formed by expanding ±15° to both sides of the device plane projection direction information with the angle value as the center, and the range is used as the relative angle range of the target motion direction.

[0030] The calculation of the angle difference and deflection angle value specifically involves calculating the angle between the path segment direction vector and the installation position vector of the audible and visual alarm device, and the angle between the path segment direction vector and the motion direction vector. The absolute value of the difference between the two is taken as the deflection angle value. The direction number corresponding to the path segment with a deflection angle value greater than 20° is marked as the direction number that needs intervention.

[0031] As a further aspect of the present invention, the method further includes step S5:

[0032] S5: Call the device response direction instruction set, load the real-time target path trajectory in the three-dimensional loading scene, insert a risk mark layer representing turning behavior into the real-time target path trajectory, draw the intrusion judgment area using red warning color, and assign event numbers to the mark layer to obtain the personnel path intrusion calibration record.

[0033] The personnel path intrusion calibration record includes a 3D scene path, a risk marker layer, red area numbers, and event identification tags.

[0034] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0035] S501: Call the device response direction instruction set, synchronously load the target path trajectory data in the real-time image frame sequence in the three-dimensional loading scene, perform state matching on the node positions including the direction deflection event in the path trajectory, extract the three-dimensional coordinate points of the steering behavior node, and insert the three-dimensional coordinate points into the layer data structure of the three-dimensional path trajectory to obtain the path trajectory steering layer data.

[0036] S502: Call the path trajectory turning layer data, read the node information marked as directional deflection and determined to have intrusion behavior, take the work boundary range coordinate group corresponding to the node's area as the drawing reference, use the red code value in the warning color configuration parameter, and attach it to the layer structure of the target path trajectory to obtain the three-dimensional scene intrusion layer area set.

[0037] S503: Call the three-dimensional scene intrusion layer area set, assign event numbers to the red marked areas, and bind the event numbers with the corresponding image frame numbers, path node indexes and target identifiers to construct a four-field structure of calibration information, which is arranged in chronological order to obtain the personnel path intrusion calibration record.

[0038] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0039] In this invention, by continuously extracting the bounding box of the human body structure in the image frame and analyzing the dynamic displacement of the trajectory nodes, the target path with continuous gait characteristics is identified and static and equipment interference areas are eliminated. The path deflection behavior position is marked by the directional angle formed by the path nodes and the vector change trend. Combined with the real-time loading operation boundary, it is determined whether the target has crossed the boundary and the minimum distance between the turning point and the boundary path is calculated to complete the intrusion behavior identification. Based on the angular relationship between the target path direction and the equipment warning direction, an instruction set is generated to guide the response action. The trajectory is dynamically loaded in the three-dimensional scene and the risk area is marked, realizing the real-time discrimination of complex behavior trajectories and the efficient linkage of alarm direction control, enhancing the accurate identification and real-time response capability of personnel intrusion events. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the steps of the present invention;

[0042] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0043] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0044] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0045] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0046] Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0048] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0049] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0050] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0051] Please see Figure 1 This invention provides a method for detecting unauthorized personnel entry based on an artificial intelligence-based loading system, comprising the following steps:

[0052] S1: Obtain the monitoring camera image frame sequence of the loading operation area, extract the human body structure bounding box region in each frame, compare the position offset value of the same target in adjacent frames by the morphological structure points in the pixel region, filter the target trajectory with continuous gait characteristics, eliminate the interference of operating equipment and static object areas, call the continuous movement path nodes of personnel images, and generate a continuous motion target image sequence.

[0053] S2: Based on the image sequence of a continuously moving target, read the coordinate values ​​of the nodes and calculate the angle change formed by every three consecutive nodes. Determine whether there is a direction deflection by the change in the angle and the trend of the change in the direction vector of the connecting line. Mark the image frame number and node position of each time a deflection occurs, and generate path offset behavior positioning information.

[0054] S3: Based on path offset behavior positioning information, read the coordinate range of the dynamic operation boundary line segment of the loading equipment in real time, and perform spatial relationship judgment between the path node and the operation boundary range. If any node is within the operation boundary, calculate the nearest distance value between the turning point and the path segment within the boundary. If the nearest distance value is less than the set intrusion identification threshold, it is determined that the target has intrusion behavior and generates the intrusion target identification status within the operation range.

[0055] S4: Based on the intrusion target identification status within the work area, read the image frame number corresponding to the intrusion target location node, combine the target motion direction vector in the image frame, calculate the angle direction range of the target motion direction vector relative to the sound and light alarm device in the two-dimensional plane, combine the offset line segment of the loading equipment's running path, determine the rotation angle and response range that the intervention device needs to respond to, and obtain the device response direction instruction set;

[0056] S5: Call the device response direction instruction set, load the real-time target path trajectory in the 3D loading scene, insert a risk marker layer representing turning behavior into the real-time target path trajectory, draw the intrusion judgment area using red warning color, and assign event numbers to the marker layer to obtain the personnel path intrusion calibration record.

[0057] The continuous moving target image sequence includes target identification, continuous path node sequence, and time frame index; the path offset behavior positioning information includes offset frame number, offset point coordinates, and offset angle value; the intrusion target identification status within the operation range includes intrusion target number, intrusion node location, intrusion trigger time, and shortest path distance; the equipment response direction instruction set includes equipment number, response angle, rotation direction, and action range; and the personnel path intrusion calibration record includes 3D scene path, risk marker layer, red area number, and event identification label.

[0058] Please see Figure 2 The specific steps of S1 are as follows:

[0059] S101: Acquire the sequence of monitoring camera images of the loading operation area, perform structured processing on each frame, extract the bounding box region of the human structure defined by rectangles from the image pixel data, construct the bounding box coordinate matrix, and perform joint matching based on the gray-scale gradient value of the human structure outline edge in the bounding box and the color distribution histogram. Call the timestamp information to compare the spatial position information offset value of the bounding box target in adjacent frames, and generate the position offset measurement result.

[0060] Each frame of the image needs to be processed in a structured manner. The main process includes image cleaning, recognition, feature extraction and matching. By calling the interface of an industrial camera or fixed monitoring equipment, an image stream is acquired at a reasonable frame rate (e.g., 25 frames per second). Then, image framing technology is used to extract a static frame image sequence. After each frame is uniformly resized and denoised, a deep learning human body recognition model is used to complete the detection of regional targets, obtain the bounding box data of the human body in the image, and manage the bounding boxes by number. The pixel area of ​​the bounding box is cropped into a rectangular area image. The grayscale image in the area is used to extract the contour boundary information. The brightness change at the human body contour is observed through image processing methods, and color information is extracted from the RGB image to construct statistical data on color distribution. The above-mentioned grayscale gradient change of the edge line and color statistical feature data together constitute the image feature template of the human target. For human bounding boxes with the same number in two consecutive frames, the similarity of the feature templates is compared for matching confirmation. After matching confirmation, the displacement change is calculated using the center position coordinates of the bounding boxes in the previous and next frames to generate the position offset measurement result.

[0061] S102: Call the position offset measurement results, reconstruct the path sequence based on the spatial displacement vector corresponding to the same bounding box identifier in consecutive frames, detect human targets in the path sequence whose displacement amplitude is greater than the displacement detection benchmark value for more than three consecutive frames, merge and cluster the human target displacement paths that meet the conditions, remove human target indexes and path data with insufficient motion amplitude, and obtain a set of continuous moving path identifiers.

[0062] This study focuses on the spatial changes of human bodies with the same target number across multiple consecutive image frames. It extracts all bounding box information detected for each human target in the image sequence and sorts them chronologically to form a dataset of movement trajectories for each single-numbered target. The spatial displacement amplitude is calculated for the center position coordinates of the human body in each frame. Target path records with consistently high displacement amplitudes across three or more consecutive time segments are identified. The "high level" threshold is determined empirically and manually adjusted based on the actual size of the human body in the actual loading area monitoring footage and the camera's field of view. After identifying targets that meet the criteria, a similar path merging principle is used to cluster target number paths that are spatially repetitive or temporally continuous. Clustering considers the geographical proximity of the path's starting and ending points, as well as the continuity of the numbering. After path aggregation, paths with generally small displacement amplitudes or insignificant fluctuations within consecutive frames are removed, considered as stationary or misidentified targets not related to loading personnel. This results in a set of continuous movement path identifiers.

[0063] S103: Call the continuous motion path identifier set, retrieve the bounding box image region in the corresponding original image frame sequence that matches the timestamp of the motion path node, extract the region pixels in sequence to form a sequence, and perform sequential numbering and inter-frame registration operations on the sequence image regions to generate a continuous motion target image sequence.

[0064] Image data related to each path is retrieved from the original image frame sequence and normalized. The timestamp sequence corresponding to each path number is read, and the frame image region corresponding to the path node is accurately extracted from the monitoring image frame according to the timestamp. In each frame, the bounding box position under the image frame is indexed by the path number, and the image region where the human body is located is cropped and formed into a continuous image sequence. The continuous image sequence is numbered in chronological order to form a target image set with clear numbering and consistent order. In order to deal with camera shake or viewpoint shift, the extracted image regions need to be registered to ensure that the relative position of the person in the image remains consistent in each frame. This registration is achieved by image key point detection technology, using the significant pixel features in the region as reference points to perform inter-frame alignment operations and generate a continuous motion target image sequence.

[0065] Please see Figure 3 The specific steps of S2 are as follows:

[0066] S201: Call the image sequence of the continuously moving target, read the coordinate values ​​of multiple path nodes in the image frame, arrange the coordinate points in chronological order, extract the coordinate value group of three consecutive nodes, calculate the included angle value of the path polyline, and generate the path polyline included angle sequence.

[0067] The node coordinates corresponding to each path are extracted. These coordinates represent the spatial position of the human target in each image frame. During the process, a mapping table is established between the image frame order and the path number. The coordinate values ​​of the human center point in each image frame are read and arranged in ascending order according to the image frame timestamp, constructing a time-continuous human displacement trajectory. In each path trajectory, three consecutive coordinate points are extracted sequentially to form a coordinate group. By reading the position information of each group of three points, two polylines are formed, and the degree of inflection of the polylines is analyzed. If the three points are A, B, and C... If the corresponding frame sequence is t1, t2, t3, then the angle between the two paths from A to B and from B to C is analyzed to identify the curvature trend of the movement path. This operation is performed through image geometric analysis. The purpose is to identify the angle changes in the direction of movement of the person in the path. The angle values ​​formed by each set of three points are recorded and archived to provide a basis for directional deflection analysis. In the loading operation scenario, if a porter moves from the loading area to the truck opening, the information on the change of the broken line direction at multiple corner points will be completely recorded and used as a data source for judging the characteristics of the movement behavior, generating a path broken line angle sequence.

[0068] S202: Call the path polyline angle sequence, perform difference calculation on adjacent angle values, and combine the difference with the direction vector change value for judgment. The direction vector change value is calculated by the angle difference between the direction unit vectors formed by the lines connecting two adjacent nodes. If both the angle difference and the direction vector change value exceed the angle deflection judgment benchmark value, it is determined that a direction deflection has occurred at the corresponding node. Record the frame number and node index that meet the deflection condition, and generate a direction deflection event record table.

[0069] Difference analysis is performed on adjacent angles in the sequence to clarify the range of change in the human target's movement path over different time periods. During processing, every two consecutive values ​​in the angle sequence are traversed, and the numerical difference is calculated to assess whether the broken path has undergone a significant turn. At the same time, the change in the path direction vector is also analyzed. The change value of the direction vector is obtained by extracting the unit direction vector formed by two adjacent broken path segments and calculating the angle between them. The change value reflects the rotation trend of the direction. When the angle difference and the change value of the direction vector at a certain node both exceed the preset angle deflection judgment benchmark value (set to 30 degrees), it is considered that a direction deflection has occurred at this node. This type of judgment is applicable to detecting the dynamic changes of the handling personnel when turning corners, making U-turns, or turning during loading. Nodes that meet the conditions are recorded, including the frame number where the deflection occurred, the index number of the node position in the path, and other information, generating a direction deflection event record table.

[0070] S203: Call the direction deflection event record table, call the position data of the deflection frame number and node index, extract the corresponding coordinate point information from the original path node coordinate sequence, and combine the frame number, node position and direction change value to establish path offset behavior positioning information;

[0071] Based on the recorded frame number and node index, the original coordinate sequence of the path nodes is traced back to extract the precise spatial location data of the nodes in the corresponding frame images. When reading the node coordinate information, a data structure combination operation is performed on each deflection event, encapsulating the image frame number, node position coordinate value, and corresponding direction change value to form a complete record entry. This entry is used to clearly identify the location point where the deflection behavior occurs during continuous movement. For example, in a loading path, if the target path node corresponding to the 45th frame changes direction, the frame number, node index, position information, and turning angle will be recorded and classified as a direction deflection event. Spatial and temporal labels for such events will be marked. The distribution of personnel and path repetition in the work area can be analyzed by path deflection frequency and location clustering, providing basic data support for route planning and personnel scheduling, and establishing path deflection behavior positioning information.

[0072] Please see Figure 4 The specific steps of S3 are as follows:

[0073] S301: Call the path offset behavior positioning information, read the coordinate range data of the dynamic operation boundary line segment of the real-time loading equipment, and make the coordinate values ​​of the direction deflection node consistent with the spatial position data of the boundary line segment. Determine whether the node position is inside the boundary by the spatial relationship between the coordinate point and the line segment area, index and mark the nodes that meet the spatial relationship, and generate a path node entry boundary identifier set.

[0074] The system loads real-time dynamic operation boundary segment data of the loading equipment. The boundary data comes from the virtual boundary settings or preset operation ranges within the site. Each boundary segment consists of start and end coordinates, forming a rectangular or polygonal operation area. A unified reference system for coordinate data is used to ensure that the coordinate values ​​of the direction deflection nodes have a consistent spatial dimension expression with the boundary segment coordinates. For example, a unified pixel coordinate system with the upper left corner of the image as the origin is used for conversion. After processing, the spatial relationship between the coordinates of each node and the boundary segment is judged. The inclusion detection principle from point to rectangular area is adopted to determine whether the node falls inside a certain operation boundary segment. During the judgment process, it is considered whether there is an intersection or inclusion relationship between the spatial position of the current path node and the area formed by the operation boundary. When the node coordinates fall into the closed area formed by a certain boundary, the node is marked as having entered the boundary state and added to the path node entry boundary identifier set. The operation can help identify which deflection behaviors are related to approaching the loading equipment. For example, if a worker suddenly turns or stops when approaching the equipment boundary, the behavior node is recorded as a key point for generating risk and a path node entry boundary identifier set is generated.

[0075] S302: Call the path node inbound boundary identifier set, extract the coordinate values ​​of the turning points corresponding to the marked nodes, and retrieve the coordinate values ​​of the continuous nodes that are still within the boundary range. Calculate the shortest distance between the turning point and each path segment to generate a sequence of path turning distances to the boundary segment.

[0076] The coordinates of the marked turning points are extracted, and these coordinates are used as the starting point to search for continuous nodes within the boundary region along the path, constructing a complete "path segment within the boundary." The distance between the turning points and the boundary segments is evaluated, focusing on determining the shortest straight-line distance from the turning point to the boundary segment. The distance calculation uses each node in the path as a test point, calculating the shortest spatial projection distance to each boundary segment. This is represented on the image plane as the vertical or oblique distance from the point to the line. A node-by-node scanning method is used, ensuring the turning points serve as a reference, to evaluate the proximity of each node in the motion trajectory to the boundary. If multiple consecutive nodes remain within the boundary region, the entire segment is identified as a boundary-approaching behavior trajectory, and a one-to-one correspondence is established with each boundary segment. The minimum distance value from the path turning point to the boundary segment is recorded to identify which motion trajectories rapidly approach or even touch the device boundary after deflection. This is a crucial basis for determining whether there are potential risky behaviors, generating a path turning-to-boundary-segment distance sequence.

[0077] S303: Call the path to the boundary segment distance sequence, perform difference calculation and numerical comparison between the shortest distance value and the set intrusion identification threshold. If the shortest distance value is lower than the intrusion identification threshold, the corresponding path identifier and node position are recorded as intrusion targets to obtain the intrusion target identification status within the operation range.

[0078] The minimum distance value of each path segment is compared and analyzed with a set intrusion identification threshold. The intrusion identification threshold is determined by manually set conditions such as equipment operation safety distance, human-machine collaboration boundary, and area warning parameters. For example, it can be set to 50 pixels or 0.5 meters. During the judgment process, the difference between each set of minimum distance values ​​is calculated and compared with the threshold. If a minimum distance value is lower than the intrusion identification threshold, it means that the personnel corresponding to the path segment have approached the equipment operation boundary too closely, entered the warning area, or even created an interaction risk. The unique identifier of the path, the location of the node where the intrusion occurred, the corresponding frame number, and other information are packaged and recorded to establish an intrusion target identification status table. The table clearly indicates the path location and time point of the suspected intrusion behavior. The identification results can be transmitted to monitoring management or warning equipment in real time, or used as one of the input data for behavior analysis to analyze and classify the behavior of personnel in the current work area. This operation is particularly suitable for application scenarios such as automated terminals and intelligent logistics loading sites that require real-time identification of boundary intrusion events, ensuring the continuity of equipment operation and the safety of personnel in the work area, and obtaining the intrusion target identification status within the work area.

[0079] Please see Figure 5 The specific steps of S4 are as follows:

[0080] S401: Call the intrusion target identification status within the operation range, read the recorded image frame number and corresponding node coordinate information, extract the spatial position coordinates of the intrusion target in two consecutive frames, construct the two-dimensional motion direction vector of the target, calculate the angle between the motion direction vector and the installation position vector of the sound and light alarm device in the real-time image frame, establish the angle direction range by combining the device plane projection direction information, and generate the relative angle interval of the target motion direction.

[0081] The relevant image frame numbers and the coordinate information of the corresponding intrusion target nodes are extracted, including the coordinate positions of the same target in the image in frames t and t+1. Based on the coordinate difference between these two frames, a two-dimensional motion direction vector of the target is constructed. The two-dimensional motion direction vector represents the direction and magnitude of the target's movement from the previous frame to the next frame. On this basis, the installation coordinate position of the audible and visual alarm device in the current image frame is read synchronously. A spatial vector is formed by the coordinates and the current position of the target, and a geometric relationship analysis is performed with the aforementioned target motion direction vector. The angle between the two vectors is used as the angular reference between the target's motion direction and the device's relative position. At the same time, the planar projection direction information of the audible and visual alarm device is called. The projection direction can be regarded as the preset angle range of the device's monitoring or pointing. By comparing the angle segment between the motion direction vector and the device's orientation, the area where the motion direction is located is defined as the relative angle interval of the target's motion direction. For example, if the target moves to the left front of the device, the direction is divided into a range of 90° to 135°, and the angle interval is used as the basis for subsequent matching to generate the relative angle interval of the target's motion direction.

[0082] S402: Call the relative angle range of the target movement direction, read the direction vector of the offset line segment in the running path of the loading equipment, compare the regional overlap between the direction vector and the angle range, filter the path segments that intersect with the target movement direction at an angle, calculate the angle difference and deflection angle value between the path segment and the direction of the alarm device, and combine the time index to mark the direction number that needs to be intervened to obtain the set of equipment response angle reference values;

[0083] Directional matching analysis needs to be performed on the marked offset path segments in the loading equipment's operating path. The direction vector of each offset path segment is retrieved. The direction vector represents the forward direction of each path segment during the equipment's movement. The path direction represented by each direction vector is mapped to an angle value and compared with the relative angle range of the target's movement direction to determine whether there is an intersection or overlap. When the direction angle of a path segment enters the target angle range or overlaps with the range, it is considered that the path segment has a potential intersection with the target's current movement direction. It is necessary to assess whether there will be a need for equipment intervention. The directional angle and deflection angle between the path segments that meet the conditions and the alarm device are calculated. The focus is on identifying scenarios with obvious direction changes. By associating the timestamp with the specific frame number, the angle direction number of each device that needs to respond is marked to obtain the device response angle reference value set.

[0084] S403: Call the device response angle reference value set, map the device response angle reference values ​​to the corresponding direction command numbers according to the device direction coding order, combine the direction numbers with the corresponding deflection angle parameters, and generate the device response direction command set.

[0085] The dataset is mapped according to the current direction coding order of the devices. Each alarm device has several preset response direction codes, corresponding to specific angle ranges that the audible and visual devices can respond to. The direction numbers in the response angle reference value set are compared with the device's responsive direction codes to establish a one-to-one correspondence. The numbers are combined with the deflection angle data to form executable command parameters. For example, a direction number corresponds to a deflection angle of 60 degrees, which will be mapped to direction code 03 in the device's internal control logic. The generated direction number command combination is integrated and organized to realize the angle response configuration of the audible and visual alarm, ensuring that the device can issue a warning signal in the preset direction. The content includes information such as direction code, deflection angle, and response timing, which is used by the device scheduler for quick command invocation to realize directional response control of intrusion targets approaching the work boundary and generate a set of device response direction commands.

[0086] Please see Figure 6 The specific steps of S5 are as follows:

[0087] S501: Call the device response direction instruction set, synchronously load the target path trajectory data in the real-time image frame sequence in the three-dimensional loading scene, perform state matching on the node positions including the direction deflection event in the path trajectory, extract the three-dimensional coordinate points of the steering behavior node, and insert the three-dimensional coordinate points into the layer data structure of the three-dimensional path trajectory to obtain the path trajectory steering layer data.

[0088] Entering the 3D loading scene processing stage, the loading scene data in the current 3D environment model is loaded, including the path trajectory information of moving targets in the real-time image frame sequence. Combining the path node positions in each frame image, the path trajectory data is embedded into the 3D spatial coordinate system in chronological order. It is necessary to convert the 2D image coordinates into 3D world coordinates and achieve projection mapping through the calibration matrix to obtain the precise position of each node in the 3D scene. For nodes in the path trajectory that have been marked as directional deflection, the spatial position is extracted and converted into 3D coordinate points, which are then inserted into the 3D layer data structure to which the path trajectory belongs. This allows the 3D path trajectory to indicate which nodes have directional deflection behavior. The layer structure includes path segments, time node markers, and behavior type fields. When adding new turning node information, the continuity of the original trajectory data and the synchronization of the time index must be maintained to achieve the structured and behavioral overlay of the path trajectory, which is convenient for event detection and 3D graphic visualization, and to obtain the path trajectory turning layer data.

[0089] S502: Call the path trajectory turning layer data, read the node information marked as direction deflection and determined to have intrusion behavior, take the work boundary range coordinate group corresponding to the node's area as the drawing reference, use the red code value in the warning color configuration parameter, and attach it to the layer structure of the target path trajectory to obtain the three-dimensional scene intrusion layer area set.

[0090] Nodes identified as having "direction deviation" and "intrusion behavior" are extracted individually to generate a visual warning layer. During extraction, target nodes are identified based on their marker attributes in the path layer, and their positions in 3D space are mapped and matched with the work boundary area. The work boundary area data describes the boundary contour with a set of coordinate points, which are used as the drawing reference for the spatial range during the drawing process. Preset warning color configuration parameters are called, especially the red code value representing a high-risk state. The color is used as a graphic identifier and superimposed on the corresponding node position in the target path trajectory layer. In the layer structure, the red identifier is not only a visual expression, but also embeds a status data field bound to the node to describe the spatial location and time information of the intrusion event. If multiple nodes are concentrated in the same area, a continuous red area surface or line segment is drawn in the layer to obtain a 3D scene intrusion layer area set.

[0091] S503: Call the 3D scene intrusion layer region set, assign event numbers to the red marked regions, and bind the event numbers with the corresponding image frame numbers, path node indexes and target identifiers to construct a four-field structure of calibration information, which is arranged in chronological order to obtain the personnel path intrusion calibration record;

[0092] Each red-marked area needs to be numbered to form an event labeling record. A unique event number is assigned to each red area, and the nodes in the path trajectory are encoded using time series to ensure the uniqueness and traceability of the number. The event number is bound to the corresponding image frame number, and the index value of the node in the path sequence and the corresponding target identification code are recorded synchronously. The above four data constitute a complete set of event labeling information fields, namely event number, frame number, node index, and target ID. The four-field structure is written into the event labeling record list and arranged in chronological order, providing a data interface for subsequent automatic alarm strategies, video evidence collection, and scene retrospection. This enables the personnel behavior data of the entire loading site to have structured storage, classification management, and rapid retrieval capabilities, effectively supporting the functional expansion of the intelligent loading monitoring platform and obtaining personnel path intrusion labeling records.

[0093] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting unauthorized entry into a loading system based on artificial intelligence, characterized in that, Includes the following steps: S1: Acquire the sequence of surveillance camera images of the loading operation area, extract the human body structure bounding box region in each frame, and generate a sequence of continuously moving target images by comparing the positional offset values ​​of the same target in adjacent frames through the morphological structure points within the pixel region; the specific steps of S1 are as follows: S101: Acquire the sequence of monitoring camera images of the loading operation area, perform structured processing on each frame, extract the bounding box region of the human structure defined by rectangles from the image pixel data, construct the bounding box coordinate matrix, and perform joint matching based on the gray-scale gradient value of the human structure outline edge in the bounding box and the color distribution histogram. Call the timestamp information to compare the spatial position information offset value of the bounding box target in adjacent frames, and generate the position offset measurement result. S102: Call the position offset measurement results, reconstruct the path sequence according to the spatial displacement vector corresponding to the same bounding box identifier in the continuous frames, detect human targets in the path sequence whose displacement amplitude is greater than the displacement detection reference value for more than three consecutive frames, merge and cluster the human target displacement paths that meet the conditions, remove human target indexes and path data with insufficient motion amplitude, and obtain a continuous movement path identifier set. S103: Call the continuous moving path identifier set, retrieve the bounding box image region in the corresponding original image frame sequence that matches the timestamp of the moving path node, extract the region pixels in sequence to form a sequence, and perform sequential numbering and inter-frame registration operations on the sequence image regions to generate a continuous moving target image sequence. S2: Based on the image sequence of the continuously moving target, read the coordinate values ​​of the nodes and calculate the angle change formed by every three consecutive nodes. Determine whether there is a direction deviation phenomenon by the change amplitude of the angle and the change trend of the connecting direction vector, and generate path offset behavior positioning information. S3: Based on the path offset behavior positioning information, read the coordinate range of the dynamic operation boundary line segment of the real-time loading equipment, perform spatial relationship judgment between the path nodes and the operation boundary range, calculate the nearest distance value between the turning point and the path segment within the boundary, and generate the intrusion target identification status within the operation range; S4: Based on the intrusion target identification status within the operating range, read the image frame number corresponding to the intrusion target location node, combine the target motion direction vector in the image frame, calculate the angle range of the target motion direction vector relative to the audible and visual alarm device in the two-dimensional plane, and obtain the device response direction instruction set.

2. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 1, characterized in that, The continuously moving target image sequence includes target identification, continuous path node sequence, and time frame index. The path offset behavior positioning information includes offset frame number, offset point coordinates, and offset angle value. The intrusion target identification status within the working range includes intrusion target number, intrusion node position, intrusion trigger time, and shortest path distance. The device response direction instruction set includes device number, response angle, rotation direction, and action range.

3. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 1, characterized in that, The displacement detection benchmark value is specifically defined as a similarity score threshold formed by jointly matching the gray-scale change gradient value of the human body structure edge line in the bounding box with the color distribution histogram. The similarity score threshold is between 0.65 and 0.

85. It is selected based on the principle of minimum mismatch rate by statistically analyzing the human target trajectory data under different movement states in the monitoring camera image frame sequence of the batch loading operation area. The spatial displacement vector is defined as the combination of coordinate increments of the center point of the bounding box coordinate matrix in adjacent image frames along the X and Y axes on the two-dimensional image plane. The coordinate increments must satisfy that the Euclidean distance is not less than the pixel distance value converted from the displacement detection reference value.

4. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Call the continuously moving target image sequence, read the coordinate values ​​of multiple path nodes in the image frame, arrange the coordinate points in chronological order, extract the coordinate value group of three consecutive nodes, calculate the included angle value of the path polyline, and generate the path polyline included angle sequence. S202: Call the path polyline angle sequence, perform difference calculation on adjacent angle values, and combine the difference with the direction vector change value for judgment. The direction vector change value is calculated by the angle difference between the direction unit vectors formed by the connecting lines of two adjacent nodes. If both the angle difference and the direction vector change value exceed the angle deflection judgment benchmark value, it is determined that a direction deflection has occurred at the corresponding node. Record the frame number and node index that meet the deflection condition, and generate a direction deflection event record table. S203: Call the direction deflection event record table, call the position data of the deflection frame number and node index, extract the corresponding coordinate point information from the original path node coordinate sequence, and combine the frame number, node position and direction change value to establish path offset behavior positioning information.

5. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Call the path offset behavior positioning information, read the coordinate range data of the dynamic operation boundary line segment of the real-time loading equipment, and make the coordinate values ​​of the direction deflection node consistent with the spatial position data of the boundary line segment. Determine whether the node position is inside the boundary by the spatial relationship between the coordinate point and the line segment area, and index and mark the nodes that meet the spatial relationship to generate a path node entry boundary identifier set. S302: Call the path node inbound boundary identifier set, extract the coordinate values ​​of the turning points corresponding to the marked nodes, and retrieve the coordinate values ​​of the continuous nodes that are still within the boundary range. Calculate the shortest distance between the turning point and each path segment to generate a path turning to boundary segment distance sequence. S303: Call the path to the boundary segment distance sequence, perform difference calculation and numerical comparison between the shortest distance value and the set intrusion identification threshold. If the shortest distance value is lower than the intrusion identification threshold, record the corresponding path identifier and node position as the intrusion target to obtain the intrusion target identification status within the operation range.

6. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Call the intrusion target identification status within the operation range, read the recorded image frame number and corresponding node coordinate information, extract the spatial position coordinates of the intrusion target in two consecutive frames, construct the two-dimensional motion direction vector of the target, calculate the angle between the motion direction vector and the installation position vector of the sound and light alarm device in the real-time image frame, establish the angle direction range in combination with the device plane projection direction information, and generate the relative angle interval of the target motion direction. S402: Call the relative angle range of the target movement direction, read the direction vector of the offset line segment in the running path of the loading equipment, compare the regional overlap between the direction vector and the angle range, filter the path segments that intersect with the target movement direction at an angle, calculate the angle difference and deflection angle value between the path segment and the direction of the alarm device, and combine the time index to mark the direction number that needs to be intervened to obtain the equipment response angle reference value set; S403: Call the device response angle reference value set, map the device response angle reference values ​​to the corresponding direction command numbers according to the device direction encoding order, and combine the direction numbers with the corresponding deflection angle parameters to generate the device response direction command set.

7. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 6, characterized in that, The process of generating the relative angle range of the target motion direction is specifically defined as follows: when the angle between the motion direction vector and the installation position vector of the audible and visual alarm device is within the range of 0° to 90°, the angle direction range is formed by expanding ±15° to both sides of the device plane projection direction information with the angle value as the center, and the range is used as the relative angle range of the target motion direction. The calculation of the angle difference and deflection angle value specifically involves calculating the angle between the direction vector of the path segment and the installation position vector of the audible and visual alarm device, and the angle between the direction vector of the path segment and the motion direction vector. The absolute value of the difference between the two is taken as the deflection angle value. The direction number corresponding to the path segment with a deflection angle value greater than 20° is marked as the direction number that needs intervention.

8. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 1, characterized in that, The method further includes step S5: S5: Call the device response direction instruction set, load the real-time target path trajectory in the three-dimensional loading scene, insert a risk mark layer representing turning behavior into the real-time target path trajectory, draw the intrusion judgment area using red warning color, and assign event numbers to the mark layer to obtain the personnel path intrusion calibration record. The personnel path intrusion calibration record includes a 3D scene path, a risk marker layer, red area numbers, and event identification tags.

9. The personnel intrusion detection method based on the artificial intelligence loading system according to claim 8, characterized in that, The specific steps of S5 are as follows: S501: Call the device response direction instruction set, synchronously load the target path trajectory data in the real-time image frame sequence in the three-dimensional loading scene, perform state matching on the node positions including the direction deflection event in the path trajectory, extract the three-dimensional coordinate points of the steering behavior node, and insert the three-dimensional coordinate points into the layer data structure of the three-dimensional path trajectory to obtain the path trajectory steering layer data. S502: Call the path trajectory turning layer data, read the node information marked as directional deflection and determined to have intrusion behavior, take the work boundary range coordinate group corresponding to the node's area as the drawing reference, use the red code value in the warning color configuration parameter, and attach it to the layer structure of the target path trajectory to obtain the three-dimensional scene intrusion layer area set. S503: Call the three-dimensional scene intrusion layer area set, assign event numbers to the red marked areas, and bind the event numbers with the corresponding image frame numbers, path node indexes and target identifiers to construct a four-field structure of calibration information, which is arranged in chronological order to obtain the personnel path intrusion calibration record.

Images