A method for analyzing user behavior of a MaaS platform based on face recognition

Abstract

The application provides a face recognition-based MaaS platform user behavior analysis method, which comprises the following steps: separating a user from carried articles by using an image segmentation technology, generating a user posture image and an article contour image, performing edge sharpening processing on the article contour image, judging whether a structure conforming to the characteristics of a luggage case exists according to a luggage shape template comparison, if the structure conforming to the characteristics of the luggage case exists, classifying and identifying the volume and quantity of the luggage case through a deep learning convolution network, combining the stop position distribution and the stop duration to determine the luggage burden level of the user, applying a path adjustment scheme, if it is found that the elevator waiting time exceeds a preset threshold, combining the historical selection preference in the user identity file, the stop position distribution and the stop duration to adjust the guidance suggestion and update the path planning information.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to a MaaS platform user behavior analysis method based on face recognition. BACKGROUND

[0002] Under the background of modern urban traffic and service integration, in-depth study of the influence of user behavior on intelligent travel platform has important theoretical value and practical significance. This research field not only relates to the improvement of user experience, but also directly relates to the optimization of urban traffic efficiency and public resource allocation. With the deep application of artificial intelligence and big data technology in the field of transportation, intelligent services based on accurate identification of user behavior have become a key factor to improve the core competitiveness of the platform. However, current research and technical solutions for user luggage status are often limited to superficial behavior observation, lacking deep mining of user dynamic needs and environmental interaction. Most of the existing methods fail to fully consider the adaptability of user behavior changes under different luggage burden conditions and the spatial environment, and ignore how the subtle behavior adjustment of users during movement due to luggage systematically affects overall travel efficiency. The increase of user luggage burden will significantly change their behavior patterns and decision preferences, among which the type, quantity and weight of luggage become key factors affecting user space selection and movement path. The increase of luggage burden often leads to frequent pauses in user movement to adjust the state, which further enhances the user's dependence on auxiliary facilities such as elevators, escalators, and significantly changes their preference for walking distance and passage selection. However, existing computer vision recognition technology often fails to accurately capture the complex details of these dynamic changes and environmental interactions, especially in the presence of diverse luggage types and complex scene conditions. Traditional methods usually use static feature extraction and fixed threshold judgment to identify luggage status, but this method is not capable of dealing with the dynamic nature of user behavior and the variability of the environment, resulting in insufficient prediction of user's real needs and difficulty in providing reliable support for subsequent service optimization and resource allocation. Therefore, how to improve and optimize the technology for identifying user luggage burden status, accurately capture user behavior adjustment patterns in different scenarios, and improve the intelligent level of service recommendation and resource allocation based on this, has become a key technical problem that needs to be broken through. SUMMARY

[0003] The present application provides a MaaS platform user behavior analysis method based on face recognition, mainly including:

[0004] Collecting dynamic images of users, extracting carrying object contour features and body posture features, locking target users carrying luggage with face recognition, establishing user identity files and obtaining initial user image data sets;

[0005] Separating the user from the carrying article, generating a user posture image and an article contour image, performing edge processing on the article contour image, comparing according to a luggage shape template, and judging whether there is a luggage feature structure;

[0006] If there is a luggage feature structure, classify the luggage volume and quantity, combine the stop position distribution and the stop duration, and determine the user luggage burden level;

[0007] According to the luggage burden level, analyze the moving speed and direction features in the user posture image, combine the stop position distribution and the stop duration, identify the user's access facility selection preference, and determine whether to prefer the elevator path by the difference in the time spent at the stairway and the elevator;

[0008] If the elevator path is preferred, obtain the elevator usage frequency and waiting time data and the passage flow data, determine the elevator congestion level and the potential congestion area distribution, and generate the elevator and luggage association degree;

[0009] Through face recognition and comparison with the user identity file, combine the luggage burden level, the elevator and luggage association degree data, and the potential congestion area distribution, analyze the user's elevator selection preference and usage scenario behavior pattern, and generate elevator usage guidance suggestions;

[0010] According to the elevator usage guidance suggestions, the elevator congestion level and the passage flow data, dynamically adjust the path planning information, and generate a path adjustment scheme suitable for the user;

[0011] According to the path adjustment scheme, combine the stop position distribution and the elevator path data, push the adjusted elevator usage scheme and alternative facility recommendation to the dynamic guidance system, track the actual effect of the user executing the path adjustment scheme through face recognition, and determine the target congestion prevention guide.

[0012] Further, the user dynamic image is collected, the carrying article contour features and body posture features are extracted, the target user carrying the luggage is locked by combining face recognition, the user identity file is established, and the initial user image data set is obtained, including:

[0013] Collecting a video stream of a person entering a detection area, extracting motion target contour information from the video stream, using key point detection to extract shoulder, hip and knee key point coordinates as body posture features to form a target feature set; performing morphological processing on the contour information in the target feature set to identify a luggage case rectangular contour, and if a rectangular contour conforming to the size of the luggage case is detected and there is a stable positional relationship with the human body contour, determining that the person is a target user carrying a luggage case, cropping a face region image and a body image; extracting a face feature vector from the face region image, comparing it with a user face feature database, matching a user identity or assigning a new user ID, and storing the face feature vector; integrating the user ID, rectangular contour parameters, key point coordinates and body image to construct a structured user identity file, and storing it in the initial user image dataset.

[0014] Further, the separated user and the carried article generate a user posture image and an article contour image, and the article contour image is subjected to edge processing to determine whether there is a luggage case feature structure according to a luggage shape template comparison, including:

[0015] A probability map of whether a pixel point belongs to a human body or an article is generated by an instance segmentation algorithm, a user posture image is formed by extracting human body region pixels according to the probability map, and a non-human body region pixel forms an initial contour image; an edge detection operator is applied to the initial contour image to mark edge points and perform morphological operations to obtain a continuous article contour boundary; the aspect ratio, rectangularity and corner point number of the article contour boundary are calculated to form a shape feature vector, which is compared with a standard feature range of a luggage case to determine whether there is a luggage case feature structure.

[0016] Further, the determination of the user luggage burden level includes:

[0017] Features are extracted from the luggage case feature structure by a convolutional neural network, the luggage case is classified as small, medium or large according to the pixel size of the bounding box, the number of luggage cases is determined by calculating the intersection over union of the bounding box, a volume level sequence and a total number are generated; the coordinates and the number of continuous frames of the user's pause are recorded by a tracking algorithm to form a spatiotemporal record set; the pause point area distribution density and the total pause time are calculated according to the spatiotemporal record set, and the luggage burden index is calculated by combining the volume level sequence and the number weighted, and the user luggage burden level is determined according to the luggage burden index.

[0018] Further, according to the luggage burden level, the moving speed and direction features in the user posture image are analyzed to identify the user's preference for the selection of a transit facility, and whether the user prefers an elevator path is determined by the difference in the time spent at a staircase and an elevator entrance, including:

[0019] Extract the user center point coordinate change from the user posture image, calculate the instantaneous moving speed and the main moving direction, and generate the motion feature dataset; track the user trajectory according to the motion feature dataset, and record the stay duration and position coordinates at the stairway and elevator port area; calculate the ratio of the stay duration at the elevator port to the stay duration at the stairway and the stay position distribution density, and determine whether the user is inclined to the elevator path.

[0020] Further, according to the luggage burden level, the moving speed and direction features in the user posture image are analyzed to identify the user's selection preference for the passageway facilities, including:

[0021] According to the luggage burden level, the instantaneous speed and speed change rate of the user position coordinates in the user posture image are calculated, the turning angle at the passageway branch point is recorded as the direction selection data, the stay duration and line of sight orientation angle at the stairway area are detected, and the behavior feature vector containing the speed change rate, turning angle, stay duration and line of sight orientation angle is generated; the frequency of the user approaching the elevator port, waiting posture parameters and stay duration difference in the behavior feature vector are counted, and the user's selection preference for the passageway facilities is determined by weighted calculation combined with the number of arm movements and step adjustment indicators.

[0022] Further, the elevator usage frequency and waiting time data and passageway flow data are obtained to determine the elevator congestion level and potential congestion area distribution, and the elevator-luggage correlation degree is generated, including:

[0023] The number of door opening and closing of the elevator is obtained through the elevator controller, the number of door opening and closing is calculated as the usage frequency, and the time interval from pressing the button to entering the elevator is recorded as the waiting time data; the people flow density in the elevator hall and passageway area is detected through the monitoring device, when the usage frequency and people flow density exceed the threshold value, the elevator is determined to be in high congestion level, and the elevator position and surrounding people flow density distribution are recorded; the usage frequency proportion of the user carrying luggage is extracted, and the correlation coefficient with the waiting time data is calculated as the elevator-luggage correlation degree.

[0024] Further, the path adjustment scheme suitable for the user is generated, including:

[0025] The face feature vector is extracted through face recognition, the user identity is confirmed by comparing with the user identity archive, the luggage burden level and historical behavior data are obtained; the distance from the user position to each elevator is calculated, the user elevator selection frequency difference is analyzed combined with the waiting number and congestion level, and the preference feature vector is formed; the recommended elevator is selected according to the preference feature vector and the elevator congestion level score, the waiting opportunity and luggage placement position are determined, and the guidance scheme is formed; the length and passageway time difference of the guidance scheme path and the original path are calculated to determine the path adjustment scheme.

[0026] Further, the dynamic adjustment of the path planning information according to the elevator use guidance suggestion, the elevator congestion level and the passage flow data comprises:

[0027] Obtain the number of elevator cars and waiting passengers, calculate the remaining carrying capacity, and mark the high load state; extract the traffic mode preference and the distribution of the resident activity area in the user identity profile; if the elevator is in a high load state, calculate the expected passage time of the escalator and the standby passage, combine the traffic mode preference to select the optimal path, generate guidance information including the main path and the alternative path, and update the path planning information.

[0028] Further, the dynamic adjustment of the path planning information according to the elevator use guidance suggestion, the elevator congestion level and the passage flow data comprises:

[0029] According to the path adjustment scheme, the stop trigger factor and the elevator operation load are extracted to generate a facility use priority sequence; the face feature vector is extracted through face recognition to confirm the user identity, and the real-time position coordinate sequence is recorded; the distance deviation of the real-time position coordinate sequence from the priority sequence recommended path is calculated, the deviation frequency and the crowd density are counted, the congestion node is selected, the shunting period and the alternative path are generated, and the target congestion prevention guide is formed.

[0030] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0031] The application discloses a MaaS platform user behavior analysis method based on face recognition, which collects user dynamic images through multi-angle cameras, extracts carrying article outlines and body posture features, locks target users in combination with face recognition, and establishes an identity profile. Image segmentation technology is used to separate users and articles, determine luggage features, and identify luggage burden levels. The user moving speed, direction and stop behavior are analyzed to determine the passage facility selection preference. According to the elevator use frequency, waiting time and passage flow, the elevator congestion level and potential congestion area distribution are generated. In combination with the user behavior mode, elevator use guidance suggestions are generated, including optimal elevator selection, waiting opportunity and luggage placement guidance. The path planning is dynamically adjusted, the elevator use scheme and alternative facility recommendation are pushed, and the target congestion prevention is realized. The application can effectively alleviate the elevator congestion problem and improve the user passage efficiency and experience. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flowchart of a MaaS platform user behavior analysis method based on face recognition.

[0033] Figure 2 A schematic diagram of a MaaS platform user behavior analysis method based on face recognition. DETAILED DESCRIPTION

[0034] For a further understanding of the present application, reference will be made to the following detailed description of the application taken in conjunction with the accompanying drawings. The following detailed description of the application is provided as an example to explain the application and should not be considered as limiting the application. It is also to be understood that the detailed description of the application is not to be taken as limiting the application. In addition, it is to be understood that only the relevant parts of the drawings are shown for the purpose of explanation.

[0035] As Figure 1 -2, the embodiment of the MaaS platform user behavior analysis method based on face recognition can specifically include:

[0036] Step S101, dynamic image acquisition is performed on the user through a multi-angle camera device, the outline feature and the body posture feature of the carried article are extracted, the target user carrying the luggage is locked by combining the face recognition technology, the user identity file is established, and the initial user image data set is obtained.

[0037] At least three high-definition cameras of different angles are deployed at the entrance of the airport or the station, real-time video stream acquisition is performed on the personnel entering the detection area, the outline information of the moving target is extracted from the continuous video frames by the frame difference method, the coordinate positions of a total of 14 key points of the shoulder, the hip and the knee are extracted as the body posture feature by using the human body key point detection method, and the target feature set containing the outline boundary coordinates and the key point coordinates is generated. The outline information in the target feature set is subjected to morphological processing, the rectangular outline feature of the luggage is identified by connected domain analysis, if a rectangular outline meeting the size range of the luggage is detected and the outline has a stable relative position relationship with the human body outline, it is judged that the personnel is the target user carrying the luggage, and the face region image and the complete body image of the target user are cropped from the corresponding video frame. For the face region image cropped, a 128-dimensional face feature vector is extracted by using a face recognition network, identity matching is performed by calculating the cosine similarity of the feature vector and each vector in the pre-established user face feature database, if the maximum similarity exceeds a preset threshold, the corresponding user identity is matched, otherwise a new user ID is assigned and the face feature vector is stored in the user face feature database. According to the determined user ID, the face feature vector, the rectangular outline parameter, the 14 key point coordinates, the complete body image and the timestamp information of the user are integrated to construct a structured user identity file, and the complete body image containing the user ID label is stored in the initial user image data set, the identity file establishment of the target user carrying the luggage and the initial image data acquisition are completed.

[0038] Specifically, the deployment of the multi-angle camera needs to fully consider the spatial layout of the detection area and the characteristics of the human flow.

[0039] In one possible implementation, three high-definition cameras are installed at the positions of directly above the entrance, 45 degrees left front and 45 degrees right front, respectively, to form a stereoscopic visual angle coverage. The camera directly above mainly captures the top profile of the luggage and the overall position relationship with the human body, and the side cameras can clearly record the side features and dragging action of the luggage. This layout ensures that at least one angle can obtain a clear target image even in a dense flow of people. The frame difference method plays a key role in motion target extraction.

[0040] Specifically, by calculating the gray difference of the corresponding pixel points of two adjacent frames of images, when the difference exceeds a preset threshold, the pixel point is marked as a motion area. After superimposing the difference results of multiple consecutive frames, a complete motion profile can be outlined. This method is particularly effective for objects such as suitcases with regular shapes, as the suitcase will form a stable rectangular profile feature during movement. The human key point detection uses a deep learning method to achieve accurate positioning.

[0041] In one embodiment, 14 key parts of the human body are identified by a convolutional neural network, including the head, shoulders, elbows, hands, hips, knees, and ankles. The spatial position relationship of these key points can accurately reflect the human posture, such as when the right hand is detected to be drooping and coinciding with the position of the suitcase handle, the system determines that the user is dragging the luggage. The angle and distance information between the key points can also distinguish between the different action patterns of pushing a luggage cart and pulling a suitcase. Connected component analysis plays a filtering role in luggage identification.

[0042] Illustratively, the motion profile image after binarization is labeled with eight-neighbor connected components, and the area, aspect ratio, and rectangularity of each connected region are calculated. A suitcase usually appears as a connected region with an area within a certain range, an aspect ratio close to the standard suitcase ratio, and a rectangularity greater than 0.85. Through the comprehensive judgment of these morphological characteristics, the suitcase can be effectively distinguished from other personal items such as backpacks and handbags. The 128-dimensional feature vector extracted by the face recognition network has a high degree of discrimination.

[0043] It should be noted that this feature vector compresses information such as texture and contour of the face into a fixed-length numerical sequence through multiple layers of convolution and pooling operations. When calculating the cosine similarity, the smaller the angle between the two vectors, the higher the similarity. When the similarity reaches 0.9 or above, it can be basically determined as the same person. This method still maintains a high recognition accuracy under conditions of light changes and slight occlusions. The establishment of a structured archive realizes the organic integration of multi-dimensional information.

[0044] Preferably, the archive is stored in JSON format, containing a timestamp field to record the detection time of each time, so as to track the user's travel trajectory. The association of the complete body image with the user ID is stored, providing a reliable data basis for subsequent luggage tracking and abnormal behavior detection. This archiving management method greatly improves the security management efficiency of airports and stations.

[0045] In step S102, the user and the carrying articles are separated by using image segmentation technology to generate a user posture image and an article contour image, and the article contour image is subjected to edge sharpening processing, and it is judged whether there is a structure conforming to the characteristics of a luggage case according to the comparison of the luggage shape template.

[0046] The original image is subjected to instance segmentation by using MaskR-CNN algorithm, a convolutional neural network is used to generate a probability map of each pixel belonging to a human body or articles, and a user posture image is formed by extracting human body region pixels from the original image according to the segmentation mask with a probability threshold greater than 0.5, and an initial contour image containing all articles is formed by extracting non-human body region pixels. The Laplace operator is applied to the initial contour image, and the weighted sum of the gray values of the center pixel and the surrounding 8 pixels is calculated, wherein the center weight is -8 and the surrounding weight is 1, to obtain an edge response value, and when the absolute value of the response value exceeds a preset threshold, it is marked as an edge point, and morphological closing operation is performed on all edge points to obtain a continuous article contour boundary. For the article contour boundary, the length and width of the contour are calculated by the minimum enclosing rectangle algorithm, the aspect ratio is calculated, and the ratio of the contour area to the area of the enclosing rectangle is calculated as the rectangularity, and the number of corner points on the contour is obtained by Harris corner point detection, forming a shape feature vector containing the aspect ratio, rectangularity and number of corner points. The shape feature vector is compared with the preset standard feature range of the luggage case, and if the aspect ratio is between 1.2 and 2.5, the rectangularity is greater than 0.85, and the number of corner points is 4 to 8, it is judged that the article contour conforms to the characteristic structure of the luggage case, and the correspondence between the user posture image and the judgment result is saved.

[0047] Specifically, the MaskR-CNN algorithm exhibits strong target recognition capability in instance segmentation.

[0048] Specifically, the algorithm generates candidate boxes through a region proposal network, and then classifies the pixels within each candidate box to determine whether they belong to the foreground or the background. In the airport security scene, when a passenger carries multiple pieces of luggage, the algorithm can accurately distinguish between the human body contour and each piece of luggage. The generation process of the probability map involves multiple convolution operations, and each pixel point will obtain a probability value between 0 and 1, indicating the likelihood of the pixel belonging to a specific class. When the probability value exceeds 0.5, the pixel is classified as the corresponding class, thereby achieving accurate image segmentation. The Laplace operator, as a classic edge detection method, highlights the areas of sharp gray level changes in the image.

[0049] In a possible implementation, the operator slides on the image through a 3x3 convolution kernel, the center pixel is assigned a weight of -8, and the surrounding 8 pixels are each assigned a weight of 1. This weight distribution makes the calculation result close to zero in a gray uniform area, and a larger response value is generated at an edge position due to a significant gray difference. The edges of the luggage case usually present clear linear features, and after being processed by the Laplace operator, these edges are significantly enhanced to form a continuous contour line. The morphological closing operation plays a key role in maintaining the integrity of the contour.

[0050] It should be noted that due to image noise or uneven illumination, the initially detected edges can have breaks. The closing operation can fill in these small breaks by first expanding and then eroding, so that the contour remains continuous. For regular objects such as luggage cases, the closing operation is particularly effective because it can maintain the integrity of the rectangular contour while smoothing out the fine jagged edges. The minimum enclosing rectangle algorithm determines the optimal bounding box of the contour by a rotating caliper method.

[0051] For example, the algorithm first finds all the vertices of the contour, and then finds the minimum enclosing rectangle by rotating a rectangular frame at different angles. The length and width information of the rectangle is recorded in this process. The aspect ratio of a standard luggage case is usually between 1.2 and 2.5, which covers various specifications from small carry-on luggage to large checked luggage. The calculation of the rectangularity reflects the regularity of the shape of the object.

[0052] Preferably, the rectangular similarity of the contour can be quantified by calculating the ratio of the actual contour area to the minimum enclosing rectangle area. The rectangularity of a luggage case as a standard rectangular object is usually greater than 0.85, while the rectangularity of a backpack or other irregular object is usually less than 0.7. This difference provides a reliable basis for accurate identification. The Harris corner detection can identify significant turning points on the contour.

[0053] In an embodiment, the algorithm finds points with significant changes in multiple directions by calculating the gray level change rate around each pixel point. A standard luggage case has 4 obvious corner points, while a luggage case with a pull rod and wheels may detect 6 to 8 corner points. The statistics of the number of such corner points provides an additional basis for luggage identification, improving the accuracy and robustness of the overall identification.

[0054] In step S103, if there is a structure that meets the characteristics of the luggage case, the volume and number of the luggage case are classified and identified through a deep learning convolution network, and the user's luggage burden level is determined in combination with the stop position distribution and the stop duration.

[0055] The identified luggage image is extracted by a ResNet convolutional neural network, the actual size is calculated according to the pixel size of the luggage bounding box and the proportion of the standard reference, the length less than 50 cm is classified as small, 50 to 70 cm is classified as medium, and more than 70 cm is classified as large, and the intersection over union of the detected multiple bounding boxes is calculated, when the intersection over union is more than 0.5, the bounding box with higher confidence is reserved, the volume level sequence and total number of each user carrying luggage are obtained. According to the volume level sequence and total number of the luggage, a Kalman filter is used to continuously track the user carrying the luggage, and the coordinate position of the user center point in each frame is recorded, when the position change in 5 consecutive frames is less than the preset pixel threshold, it is determined as a pause state, the starting coordinates and the number of continuous frames of the pause are recorded, and a space-time record set containing multiple pause events is formed. The Euclidean distance between adjacent pause points is calculated for the space-time record set, the pause points with a distance less than a preset threshold are classified into the same region, the pause frequency in each region is counted as the distribution density value, and the total pause time is accumulated. The distribution density value, the total pause time and the number of each level in the volume level sequence are weighted and summed, the weight of the large luggage is 3, the weight of the medium luggage is 2, and the weight of the small luggage is 1, and the luggage burden index is calculated. According to the numerical range of the luggage burden index, the luggage burden level of the user is determined, if the index is greater than the first preset threshold, it is determined as high burden level, if the index is less than the first preset threshold and greater than the second preset threshold, it is determined as medium burden level, and if the index is less than the second preset threshold, it is determined as low burden level, and the determination of the luggage burden level of the user is completed.

[0056] Specifically, the ResNet convolutional neural network plays a key role in luggage size identification.

[0057] Specifically, the network solves the gradient disappearance problem of deep network through residual connection, so that the network can learn the subtle feature difference of the luggage. In practical application, the network first extracts the edge, texture and shape features of the luggage, and then outputs the accurate coordinates of the bounding box through the fully connected layer. These coordinate information is compared with the standard reference placed in the scene, so as to calculate the actual size of the luggage. The standard reference usually selects the ground marking line or fixed size tile, whose size is known and easy to identify in the image. The intersection over union calculation plays a filtering role when processing overlapping bounding boxes.

[0058] In one possible implementation, when the ratio of the overlapping area of two bounding boxes to the union area exceeds 0.5, the system considers that the two boxes mark the same luggage. At this time, the bounding box with a higher confidence score is retained, and the box with a lower confidence score is deleted. This mechanism effectively avoids the problem of repeated counting of the same luggage, especially in the case of partial occlusion or angle change of the luggage. The Kalman filter provides stable trajectory prediction capability in user tracking.

[0059] It should be noted that the filter predicts the user's position at the next time by establishing a state space model of user motion, combining historical position information and current observation value. In the airport or station environment, the moving speed of the user carrying luggage is relatively slow and regular, and the Kalman filter can smooth the noise points in the trajectory and accurately identify the real pause behavior. When the user is queuing at the ticket window or staying in the rest area, the system judges the pause state through the position change of 5 consecutive frames, and the time window is about 0.2 seconds, which can capture the short pause and filter out the instantaneous pause in normal walking. The Euclidean distance calculation is used to analyze the spatial distribution characteristics of the pause points.

[0060] For example, in the waiting hall, the user may pause at multiple places such as ticket gates, seat areas, and toilet entrances. By calculating the straight-line distance between these pause points, when the distance is less than 5 meters, the system classifies them into the same functional area. This area division reflects the user's behavior pattern, and frequent pausing between multiple areas often means physical exertion caused by carrying heavy luggage. The weighted calculation of the luggage burden index reflects the actual burden difference caused by different volumes of luggage.

[0061] In one embodiment, a user carrying two large luggage has a volume weight contribution of 6 points, and if the user pauses 8 times in 10 minutes, distributed in 3 different areas, the pause density is 2.67. Combined with the total pause time of 120 seconds, the final burden index is obtained by linear combination. This quantitative method enables the system to objectively evaluate the actual burden of the user, providing data support for subsequent personalized services. High-burden-level users may need luggage handling assistance, while low-burden-level users can quickly pass through the security channel.

[0062] Step S104, according to the luggage burden level, analyze the moving speed and direction features in the user posture image, combine the pause position distribution and duration, identify the user's choice preference for the transit facility, and determine whether to prefer the path containing the elevator by analyzing the difference in stay time at the stairway and elevator entrance.

[0063] According to the determined luggage burden level, the coordinate changes of the user center point between adjacent frames in the user posture image sequence are extracted, the coordinate difference value is calculated to obtain the displacement distance, the displacement distance is divided by the inter-frame time interval to obtain the instantaneous moving speed, the average value of the direction angles of the displacement vectors of the continuous multiple frames is calculated to determine the main moving direction, and the speed sequence and the direction sequence corresponding to different burden levels are combined to form a motion feature data set. The motion feature data set is used to track the user moving trajectory. When a trajectory point enters a preset stairwell area or a boundary of an elevator area, timing starts, and when the trajectory point exits the area, timing ends. The total stay time of the user in all stairwell areas and the total stay time of the user in all elevator areas are accumulated respectively, and the position coordinates of each stay event are recorded. According to the total stay time, the ratio of the elevator area stay time to the stairwell area stay time is calculated, and the spatial distribution density of the stay position coordinates is counted. If the ratio is greater than a preset threshold and the stay density near the elevator area is higher than the stay density near the stairwell area, it is determined that the user prefers to select a path containing an elevator, otherwise it is determined that the user does not prefer to select a path containing an elevator.

[0064] Specifically, there is a close correlation between the luggage burden level and the user moving speed.

[0065] In a possible implementation, the average moving speed of a user with a high burden level is usually 0.5 to 0.8 meters per second due to carrying multiple large pieces of luggage, while the moving speed of a user with a low burden level can reach 1.2 to 1.5 meters per second. Such speed difference is not only reflected in straight-line walking, but also more obvious when turning and avoiding other pedestrians. By continuously collecting the user center point coordinates, the system can accurately calculate the instantaneous speed at each moment, and then analyze the speed change rule. The direction angle calculation of the displacement vector involves coordinate system conversion and angle normalization processing.

[0066] Specifically, the system first determines the coordinates of the user in the nth frame as (x1, y1) and the coordinates of the user in the (n+1)th frame as (x2, y2), and then calculates the direction angle by using an arctangent function. Considering that the user will have a small left-right swing in actual movement, the system uses a sliding window method to average the direction angles of the continuous 10 frames, so as to obtain a stable main moving direction. This smoothing method can effectively filter out the direction jitter caused by the change of walking posture. The definition of the stairwell area and the elevator area needs to be accurately marked in combination with the actual site layout.

[0067] It should be noted that the staircase entrance area is usually set as a square area of 3m x 3m in front of the staircase entrance, and the elevator entrance area includes a rectangular area of 2m x 4m in front of the elevator door and a waiting area. When the user trajectory point first enters these preset areas, the system starts a timer to record the entry time; when the trajectory point completely leaves the area boundary, the timer stops, and the difference between the two times is the single stay duration. The accumulation process of the stay duration reflects the user's hesitation and selection behavior in front of different facilities.

[0068] For example, a user carrying heavy luggage may only briefly stop and observe at the staircase entrance, with an average stay time of about 3 to 5 seconds, and then continue to look for an elevator; the stay time at the elevator entrance is significantly longer, including the time waiting for the elevator to arrive, usually 30 seconds to 2 minutes. This duration difference provides an important basis for judging user preferences. The calculation of spatial distribution density is realized by using the kernel density estimation method.

[0069] In one embodiment, the system divides the monitoring area into a 1m x 1m grid and counts the number of stay events in each grid. The grids near the elevator entrance usually show high density clustering characteristics, while the density near the staircase entrance is relatively low. By calculating the average stay density within a 5m range around the elevator entrance and the density ratio of the same range around the staircase entrance, the user's preference for different passage facilities can be quantified. The setting of the ratio threshold needs to consider multiple factors.

[0070] Preferably, when the ratio of the stay duration at the elevator entrance to the stay duration at the staircase entrance exceeds 3.0, and the stay density near the elevator entrance is at least 2 times that of the staircase entrance, the system determines that the user has a clear preference for the path containing the elevator. This double judgment mechanism avoids misjudgment caused by a single indicator, improving the accuracy and reliability of preference recognition.

[0071] Based on the luggage load level data, the user's movement speed trend is extracted, the user's direction selection preference at the passage branching point is captured, the user's wandering time and line of sight stay direction in the staircase entrance area are recorded, the approach frequency and waiting posture features in the elevator entrance area are monitored, and the differences in user's pause behavior in front of the two types of passage facilities are compared and analyzed. By statistically analyzing the frequency of step adjustment towards the elevator direction and the tendency of body movements approaching the elevator button, the path exploration mode of users carrying heavy luggage is identified, and the user's dependence on the elevator passage method and selection priority are determined.

[0072] Based on the luggage burden level data, the instantaneous speed is obtained by calculating the Euclidean distance of the position coordinates of the user at adjacent time points divided by the time interval, the derivative of the speed with respect to time is calculated by using the five-point difference method to obtain the speed change rate, when the user approaches the channel branching point, the angle change value of the moving track from straight line to turning is recorded as the direction selection data, and the event that the user continuously stays at the stairway for more than a preset time threshold is detected, the nose tip orientation angle is obtained as the line of sight direction by face key point detection, and the user behavior feature vector containing the speed change rate sequence, the turning angle value, the stay time and the line of sight orientation angle is generated. For the user behavior feature vector, the cumulative number of times that the user enters the elevator port within a radius of 5 meters is counted as the approach frequency value, the user skeleton key points are extracted by the OpenPose algorithm, the double foot spacing change and the shoulder inclination angle are calculated as the waiting posture parameters, the difference value between the stairway stay time and the elevator port stay time is calculated, and the line of sight orientation angle is combined with the duration of pointing to the elevator to construct a comprehensive feature matrix containing the approach frequency value, the posture parameter, the stay time difference value and the line of sight duration. According to the comprehensive feature matrix, the standard deviation of the adjacent step spacing during the user moving towards the elevator is extracted as the step adjustment index, the number of actions that the height of the raised arm exceeds the shoulder and points to the position of the elevator button is identified, the approach frequency value, the posture parameter mean value, the step adjustment index and the action number are weighted according to the weight of 3:2:2:3, and if the weighted result exceeds the first preset threshold, it is determined as the elevator strong dependence type path exploration mode, if it is between the first and second preset thresholds, it is determined as the elevator preference type path exploration mode, and the dependence degree and selection priority of the user on the elevator passing mode are determined.

[0073] Specifically, the five-point difference method has good precision and stability in the calculation of the speed change rate.

[0074] Specifically, the method uses the speed values of the current time and the two times before and after the current time to calculate the speed change rate by weighted difference. When the user carries heavy luggage and walks in the corridor, the speed will show periodic fluctuations, especially when approaching the branching point, the speed change rate will have a significant peak. This peak reflects the user's hesitation and deceleration behavior at the decision point, providing an important basis for subsequent path selection analysis. The direction selection record of the channel branching point involves accurate measurement of the track angle.

[0075] In one possible implementation, the system first determines the straight line travel direction of the user 10 meters before the branching point as the reference direction, and then tracks the actual travel direction of the user after passing through the branching point. The included angle between the two direction vectors is the turning angle value. Users carrying large luggage often choose paths with larger turning radii, and their turning angles are usually within 30 degrees, while lightly equipped users may choose sharper turns with angles up to 45 degrees or more. Face key point detection plays a key role in line of sight direction judgment.

[0076] It should be noted that by detecting the positions of 68 key points such as the corners of the eyes, the tip of the nose and the corners of the mouth, a three-dimensional model of the face is constructed, and the angle of orientation of the tip of the nose can accurately reflect the gaze direction of the user. In the stairwell area, the user's line of sight will switch between the stairs and the surrounding environment, and when the line of sight continuously points upwards for more than 3 seconds, it indicates that the user is evaluating the difficulty of climbing the stairs. The OpenPose algorithm realizes real-time extraction of the human skeleton through a deep learning network. This algorithm can identify 25 key points including the head, torso and limbs, forming a complete human skeleton structure. When waiting at the elevator entrance, the user's posture will change characteristically: the distance between the feet changes from dynamic changes during walking to relative stability, usually maintaining a distance of 30-50 centimeters; the shoulders will tilt towards the direction of the luggage, and the tilt angle is proportional to the weight of the luggage, and heavy luggage can cause a tilt of 10-15 degrees. The calculation of the pace adjustment index reflects the changes in the user's behavior when approaching the target.

[0077] For example, when the user determines to use the elevator, the pace will gradually adjust to align with the center position of the elevator door. The standard deviation of the adjacent pace interval can quantify this adjustment process, and the larger the standard deviation, the more frequent the user's pace adjustment, which usually occurs when carrying multiple pieces of luggage that need to be rebalanced. The weight setting in the weighted calculation is based on the degree of influence of each index on the final decision.

[0078] In one embodiment, the proximity frequency and the number of actions are given a higher weight of 3 because they directly reflect the user's intention to act; the posture parameters and the pace adjustment index are given a weight of 2 as auxiliary judgment basis. Through this weighting mechanism, multi-dimensional information can be integrated to accurately identify elevator-dependent users, who usually repeatedly approach the elevator area and exhibit obvious waiting postures, providing data support for the planning of barrier-free facilities in airports and stations.

[0079] Step S105, if the user's preference for the path containing the elevator is displayed, the elevator usage frequency and waiting time data are obtained, the channel flow data are obtained, the elevator congestion level and the position distribution of potential congestion areas are determined, and the elevator-luggage correlation degree is generated through historical data analysis.

[0080] If the passage facility selection preference shows that the user prefers to select a path containing an elevator, the number of door openings and the running period data of each elevator are obtained in real time through the elevator controller interface. The number of door openings per unit time is taken as the elevator usage frequency, and the time interval from pressing the elevator button to entering the elevator is recorded to calculate the average waiting time in each period, and the running state record containing the elevator usage frequency value and the waiting time data is generated. According to the usage frequency value and the waiting time data in the running state record, the personnel target in the elevator hall and the passage area is detected by the monitoring device using the YOLO algorithm, and the number of personnel in a unit area is counted to obtain the crowd density value. When the elevator usage frequency exceeds the first preset threshold and the crowd density in the area where the elevator is located exceeds the second preset threshold, the elevator is determined to be in a high congestion level, and the floor position and the surrounding crowd density distribution of the high congestion level elevator are recorded to form the congestion state evaluation result. The elevator position and congestion level information in the congestion state evaluation result are used to extract the usage record of each elevator in a preset time period from the stored user behavior record, and the usage times of users carrying luggage are selected to calculate the proportion of the usage times of users carrying luggage in the total usage times. The correlation degree between the proportion value and the waiting time in the corresponding period is calculated by the Pearson correlation coefficient, and the calculated correlation coefficient is taken as the association degree of the elevator and the luggage.

[0081] Specifically, the elevator controller interface provides a key channel for obtaining elevator operation data.

[0082] In one possible implementation, the system establishes a connection with the controller through a standard elevator monitoring protocol and reads the running state data of the elevator at fixed time intervals. The number of door openings involves continuous monitoring of the door state changes of the elevator. When the door changes from a closed state to an open state, the counter increases by 1. The running period records the time taken by the elevator to complete a complete up-and-down trip. These raw data provide a basis for subsequent usage frequency calculation. The measurement of waiting time requires accurate capture of key time points of user behavior.

[0083] Specifically, when a user presses an elevator button, the system records the time as the waiting start time. Through sensors or image recognition technology inside the elevator, the moment when the user enters the elevator car is detected and recorded as the waiting end time. The difference between the two time points is the single waiting time. In peak periods, the waiting time can reach several minutes, while in non-peak periods, it is usually within 30 seconds. The YOLO algorithm exhibits excellent real-time performance in crowd density detection.

[0084] It should be noted that YOLO divides the entire image into grids, and each grid is responsible for predicting the target whose center point falls within the grid. Through the convolutional neural network, the position and category information of all targets are output at one time, avoiding the repeated calculation of the traditional sliding window method. In the elevator hall scene, the algorithm can detect multiple waiting users at the same time, and can maintain a high detection accuracy even if there is partial occlusion. The calculation of the crowd density needs to consider the unevenness of the spatial distribution.

[0085] For example, the personnel density within 2 meters in front of the elevator door is usually the highest, while the density on both sides of the corridor is relatively low. The system divides the monitoring area into 1 square meter grid units, and counts the number of people detected in each grid. When the crowd density of a certain area exceeds 4 people per square meter, it indicates that the area is already quite crowded, affecting normal passage. The Pearson correlation coefficient plays a quantitative role in analyzing the correlation between elevators and luggage.

[0086] In one embodiment, the system extracts two sets of data for each time period: the proportion of users carrying luggage and the average waiting time. The calculation of the correlation coefficient is based on the ratio of the covariance of the two sets of data to the respective standard deviations. When the correlation coefficient is close to 1, it indicates that the higher the proportion of users carrying luggage, the longer the elevator waiting time, and the two show strong positive correlation. This correlation reflects the impact of luggage on elevator transportation efficiency. The formation of the congestion state evaluation result integrates multi-dimensional information.

[0087] Preferably, the system not only records the congestion level of the elevator, but also includes a crowd distribution heat map of the surrounding area. An elevator with a high congestion level is often accompanied by a crowd gathering in the surrounding corridor, forming a chain effect. By identifying these congestion patterns, management personnel can take early relief measures, such as temporarily opening the freight elevator for passenger use, or guiding some passengers to use elevators on other floors, thereby improving the overall passage efficiency.

[0088] Step S106, when the system re-monitors users in different areas, it associates and compares with the user identity archives through face recognition technology, analyzes the user's behavior patterns in the elevator selection preference and elevator use scenarios according to the location distribution of potential congestion areas, combines the luggage burden level and the degree of elevator and luggage correlation, generates elevator use guidance suggestions, and obtains a path adjustment scheme suitable for the user.

[0089] When the monitoring device re-captures the user image in different areas, the 128-dimensional face feature vector is extracted by the FaceNet algorithm, and the cosine similarity calculation is performed with the feature vector in the pre-stored user identity file. If the similarity exceeds the pre-set threshold, the user identity is confirmed, and the luggage burden level data, historical elevator use time period record and average waiting time data of the user are obtained from the file to generate an identity association record containing the user identity, current area coordinates, burden level value and historical behavior data. According to the current area coordinates in the identity association record, the stored elevator and luggage association degree data and potential congestion area distribution map are called, the physical distance from the user's location to each elevator is calculated by the Euclidean distance formula, and the waiting time is estimated by combining the real-time waiting number of each elevator. The frequency difference between the user's selection of long-distance and low-congestion elevators during peak hours and the selection of short-distance elevators during off-peak hours is analyzed, and the distance preference value, time period preference value and congestion tolerance are combined into a three-dimensional preference feature vector. Using the component values of the three-dimensional preference feature vector, combined with the congestion level of each elevator and the luggage burden level of the user, a comprehensive score is made according to the distance weight 0.3, the waiting time weight 0.5 and the congestion degree weight 0.2. The elevator with the highest score is selected as the recommended elevator, the average waiting time of the elevator during the same period is determined to determine the recommended waiting time as the current time or delayed by a pre-set number of minutes, and the corner space occupancy rate of the elevator car is used to generate a position suggestion for placing luggage in the front corner or the rear corner, forming a complete guidance scheme containing the recommended elevator number, waiting time and placement position. According to the recommended elevator position in the complete guidance scheme, a path planning algorithm is used to calculate the shortest path from the user's current position to the elevator, and the new path length is compared with the direct path length of the user's original target position. If the path length increases by less than a pre-set distance threshold and the total travel time after adding the waiting time is still less than the expected time of the original path, the new path is determined as the path adjustment scheme suitable for the user.

[0090] Specifically, the FaceNet algorithm has excellent performance in face recognition.

[0091] Specifically, the algorithm maps the face image to a 128-dimensional feature space through a deep convolutional neural network, so that different photos of the same person are close in this space, while photos of different people are far apart. In practical applications in airports or stations, even if the user moves for a long time and the face appears tired, or the lighting conditions change, FaceNet can still accurately extract stable feature vectors. The calculation of cosine similarity is realized by the product of the inner product of the vector and the modulus of the vector. When the similarity reaches 0.85 or above, it can be confirmed that it is the same user. The construction of the three-dimensional preference feature vector reflects the user's comprehensive behavior tendency.

[0092] In one possible implementation, the distance preference value is quantified by the statistical distance difference between the user's selected elevator and the nearest elevator. Users carrying heavy luggage tend to walk further to use an idle elevator, and their distance preference value is usually higher. The time period preference value records the user's elevator selection pattern at different times, and users may prefer to wait during the morning rush hour and pursue fast passage during off-peak hours. The crowdedness tolerance is derived by analyzing the user's stay time in different density crowds, reflecting the user's acceptance of crowded environments. The weight setting of the comprehensive scoring mechanism is based on statistical analysis of a large amount of user behavior data.

[0093] It should be noted that the waiting time is given the highest weight of 0.5, because reducing the waiting time is the primary consideration for users carrying luggage. The distance weight of 0.3 reflects the user's acceptance of additional walking distance, and the crowdedness weight of 0.2 reflects the consideration of comfort. This weight distribution enables the system to find a balance between efficiency and comfort. The analysis of elevator car space utilization involves a precise understanding of the internal layout.

[0094] For example, the four corners of a standard elevator car are usually ideal locations for placing large luggage, neither affecting the entry and exit of other passengers nor fully utilizing the space. The system calculates the available space in each corner by analyzing the distribution of people inside the car and the location of existing luggage. When the current corner space occupancy is less than 30%, it is recommended to place the luggage in the front corner; when the back corner is more free, the back corner position is recommended. This fine-grained space management greatly improves the carrying efficiency of the elevator. The path planning algorithm needs to consider multiple constraint conditions when generating adjustment schemes.

[0095] In one embodiment, the algorithm not only calculates the path with the shortest geometric distance, but also considers the actual passage capacity of the passage. Although the spacious main passage may have a longer path, it has a faster passage speed; while the shortcut may be longer in time due to its narrowness or dense flow of people. By controlling the path length increase value within a reasonable range while ensuring the reduction of total passage time, the path adjustment scheme generated by the system is both practical and efficient. The determination of waiting opportunity combines real-time data and historical regularity.

[0096] Preferably, the system analyzes the historical average waiting pattern of a specific elevator during the current time period and identifies the fluctuation period of the waiting time. When it is detected that it is about to enter the trough period, the user is advised to go immediately; if it is in the peak period, it is advised to delay for 3-5 minutes. This dynamic adjustment mechanism effectively avoids the user's ineffective waiting at the elevator door, improving the overall passage experience and facility utilization efficiency.

[0097] Step S107, an application path adjustment scheme is used, if it is found that the elevator waiting time exceeds the preset threshold, the historical selection preference, the stop position distribution and the stop duration in the user identity file are combined to adjust the guidance suggestion, and the path planning information is updated.

[0098] The application path adjustment scheme is used to obtain the number of people in the elevator car and the number of people waiting at each floor of the elevator shaft, calculate the remaining carrying capacity according to the rated carrying capacity of the elevator, obtain the number of people that can be accommodated by dividing the remaining carrying capacity by the standard value of the space occupied by each person recorded in the building management system, and if the number of people waiting exceeds the product of the number of people that can be accommodated and a preset congestion coefficient, the current elevator congestion level is marked as a high load state. The historical vertical transportation selection records in the user identity file are extracted, the number of elevator uses and the number of escalator uses are counted, and the ratio of the two is calculated as the user's transportation mode preference value. The stop position coordinates and the corresponding stop duration data of the user at different floors are obtained, and the user's permanent activity area distribution is determined according to the position coordinates whose stop duration exceeds a preset stay threshold. If the current elevator congestion level is in a high load state, the real-time running data of the escalators in the building and the people flow density data collected by the people flow monitoring equipment in each passage are called, the predicted travel time of the escalator path is calculated according to the distance from the user's current position to the permanent activity area, the running speed of the escalator and the people flow density, the predicted travel time of the standby passage path is calculated according to the length of the passage and the people flow travel speed, and the two times are weighted according to the user's transportation mode preference value. According to the predicted travel time after weighting, the optimal path scheme is selected, the node sequence of the path and the historical congestion data of the corresponding period are obtained, and if there is a congestion risk in the predicted arrival period, guidance information including the main path, the alternative path and the recommended travel period is generated, and the updated path planning information is pushed to the user terminal.

[0099] Specifically, the determination of the elevator congestion level involves comprehensive evaluation of multidimensional data.

[0100] In one possible implementation, the number of people in the car is obtained in real time by a weight sensor installed on the top of the car, which can sense the total weight change in the car, and the current number of passengers can be obtained by subtracting the self-weight of the car from the total weight and dividing by the standard value of the weight per person. The number of people waiting at each floor of the elevator shaft is obtained by video analysis of the waiting hall by a camera, and the number of people in the waiting area is identified by using target detection technology. The calculation of the remaining carrying capacity needs to consider the difference between the rated carrying capacity and the current carrying capacity, and the standard value of the space occupied by each person is usually set to 0.2 square meters per person, which is derived from historical statistical data in the building management system. The preset congestion coefficient is dynamically adjusted according to different time periods, and is set to 1.5 during peak hours and 2.0 during off-peak hours, which can more accurately reflect the actual congestion of the elevator. The historical data mining in the user identity file provides an important basis for personalized path recommendation.

[0101] Specifically, the vertical traffic selection record contains the user's moving trajectory in the building each time, records the time of selecting the elevator or escalator, the start and end floors, and other information. The calculation of the traffic mode preference value not only considers the simple ratio of the number of uses, but also combines factors such as waiting time and congestion level for weighted processing. The stop position coordinates are obtained through indoor positioning technology. When the user stays at a certain position for more than a preset stay threshold, the system records the coordinate information of the position. The determination of the distribution of the resident activity area adopts a clustering analysis method to perform spatial clustering on multiple position points with a long stay duration, forming a user's main activity area map. These areas usually include office, conference room, restaurant and other functional spaces. The prediction calculation of path length integrates multi-source real-time data.

[0102] It should be noted that although the running speed of the escalator is relatively fixed, the actual passing time is also affected by the personnel density on the escalator. When the personnel density exceeds a certain threshold, passengers need to wait stationary on the escalator, resulting in an extension of the actual passing time. The people flow monitoring device includes an infrared sensor, a laser radar, etc., which can collect the personnel distribution density in the channel in real time. The calculation of the expected passing time of the standby channel path needs to consider the physical length of the channel, the current people flow density, and the personnel walking speed, among which the walking speed will be dynamically adjusted according to the people flow density, and the greater the density, the slower the speed. In the weighted processing, if the user's historical preference shows a greater tendency to use the elevator, the time length weight of the escalator path will be appropriately increased, and vice versa. This personalized weighting method can better meet the user's actual selection habits. The generation and pushing of path planning information form a complete service closed loop.

[0103] In one embodiment, the historical congestion data is obtained by statistically analyzing the people flow in the same period in the past, forming a three-dimensional data matrix of time period-position-congestion level. When congestion is predicted to occur in a certain period, the system will automatically generate a time window suggestion, recommending the user to travel in advance or delay. The generation of the guidance information not only includes the optimal main path, but also provides at least one alternative path, ensuring that the user has other choices when the main path suddenly congests. The information pushed to the user terminal adopts a combination of text and graphics, including path map, expected time length, congestion warning, etc. Through this multi-dimensional information pushing, the user can make more reasonable travel decisions, effectively avoid congestion areas, and improve the passing efficiency in the building.

[0104] Step S108, according to the updated path planning information, combining the stop trigger factor and the elevator path planning data, pushing the adjusted elevator use scheme and alternative facility recommendation to the dynamic guidance system, while continuously tracking the actual effect of the user executing the guidance suggestion to determine the target congestion prevention guide.

[0105] According to the node sequence and the expected passing time length in the updated path planning information, records of the user historical movement data in which the stay frequency and the stay time length at a specific position exceed a preset threshold are obtained as the pause trigger factors, and current period elevator operation load data is combined to sort the main elevator group, the standby elevator group and the escalator facility according to the load from low to high, to generate a facility use priority sequence. The facility use priority sequence is sent to the user terminal through the message push module of the building management platform, and user face image data is collected from the cameras in the building, a 128-dimensional face feature vector is extracted by using a convolutional neural network, and a cosine similarity is calculated with the pre-stored feature vector in the user identity archive. When the similarity exceeds a preset identity confirmation threshold, the user identity is determined, and the real-time position coordinate sequence of the user is recorded. The actual movement path of the user is calculated according to the real-time position coordinate sequence, the Euclidean distance between each node of the actual path and the corresponding node of the recommended path of the facility use priority sequence is obtained, and if the Euclidean distance of a plurality of continuous nodes exceeds a preset deviation threshold, the deviation starting position, the deviation path segment and the people flow density value of the period are recorded to form a path deviation data set. The deviation times of each position node are counted from the path deviation data set, the nodes whose deviation times exceed a preset frequency threshold and whose people flow density values exceed a preset density threshold are selected as congestion nodes, a low-density period is extracted from the historical data as a shunting period for each congestion node, a path bypassing the node is selected from the path planning information as a replacement path, and a suboptimal facility is selected from the facility use priority sequence as a switching suggestion, and the target congestion prevention guide is formed by integration.

[0106] Specifically, the identification of the pause trigger factors depends on the deep mining of the user historical behavior patterns.

[0107] In a possible implementation, the stay frequency of the user at a certain position is determined by counting the number of occurrences of the position coordinate within a preset radius range. When the user is located in the same area at a plurality of continuous sampling points and the duration exceeds a preset threshold, an effective stay is recorded. Such pause often occurs in elevator halls, corners or passage intersections, reflecting the decision hesitation or waiting behavior of the user at these positions. The elevator operation load data is obtained by real-time monitoring of the passenger carrying rate, the running direction and the number of stop floors of each elevator. The higher the load is, the longer the waiting time of the elevator is. The generation of the facility use priority sequence considers the load condition and the position of the user, and preferentially recommends the vertical transportation facility with lower load and closer distance to the user. The application of the convolutional neural network in face recognition realizes the accurate confirmation of the user identity.

[0108] Specifically, the facial images captured by the camera are first pre-processed, including face detection, alignment, and normalization, to ensure the standardization of the input images. The convolutional neural network gradually extracts facial features from low-level to high-level through multiple layers of convolution and pooling operations, and finally outputs a 128-dimensional feature vector in the fully connected layer. This feature vector contains key information about the face, such as the numerical representation of geometric features such as eye spacing, nose shape, and mouth contour. The calculation of cosine similarity is obtained by performing the dot product operation on two feature vectors and dividing by the product of the vector magnitudes, with a value range of -1 to 1. The closer to 1, the more similar the two faces. The preset identity confirmation threshold is usually set above 0.8 to ensure the accuracy of identity recognition. The collection of real-time position coordinate sequences and path deviation detection form the core mechanism of user behavior monitoring.

[0109] It should be noted that the calculation of Euclidean distance involves the straight-line distance between the actual path nodes and the corresponding nodes of the recommended path on the two-dimensional plane, which is obtained by taking the square root of the sum of the coordinate differences. When the user's actual walking path deviates from the recommended path, the system does not immediately determine it as abnormal, but requires that multiple consecutive nodes exceed the deviation threshold. This design avoids false positives caused by temporary avoidance of others or short-term stays. The deviation starting position records the specific location where the user starts to change the path, and the deviation path segment records the actual walking trajectory of the user. Combined with the crowd density value in this period, the environmental factors that cause the user to change the path can be analyzed. The identification of congestion nodes and the generation of prevention guidelines embody the complete process from data analysis to decision output.

[0110] In one embodiment, the sliding time window method is used to count the number of deviations, only counting the deviation events in the recent period of time to ensure the timeliness of the data. The double determination standard of congestion nodes requires both a high number of deviations and a large crowd density, accurately locating the key positions that truly affect the traffic efficiency. The selection of shunting periods analyzes the crowd density distribution of the node at different time periods in the historical data to find the periods with significantly lower density than the average. The generation of alternative paths uses graph search algorithms to find feasible paths that bypass the congestion nodes in the building topology. The facility switching suggestion selects the facility with the next optimal ranking but can avoid the congestion node from the generated facility usage priority sequence. These information integrations form the target congestion prevention guidelines, providing operational and specific solutions for crowd management in buildings.

[0111] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for user behavior analysis on a MaaS platform based on face recognition, characterized in that, The method includes: Collect dynamic images of users, extract the outline features of carried items and body posture features, combine facial recognition to locate target users carrying suitcases, establish user identity profiles and obtain initial user image datasets; Separate the user from their belongings, generate a user pose image and an item outline image, perform edge processing on the item outline image, and determine whether there are suitcase feature structures by comparing with a luggage shape template. If luggage features are present, the luggage volume and quantity are classified. Combining the distribution of pause locations and pause duration, the user's luggage burden level is determined. Specifically, a ResNet convolutional neural network is used to extract features from the identified luggage images. The actual size is calculated based on the ratio of the pixel size of the luggage bounding box to a standard reference object. Luggage less than 50 cm in length is classified as small, 50-70 cm as medium, and greater than 70 cm as large. Simultaneously, the intersection-union ratio (IUR) is calculated for multiple detected bounding boxes. When the IUR exceeds 0.5, the bounding box with higher confidence is retained, resulting in a luggage volume level sequence and total quantity for each user. Based on the luggage volume level sequence and total quantity, a Kalman filter is used to continuously track users carrying luggage, recording the coordinates of the user's center point in each frame. When the position change within 5 consecutive frames is less than a preset pixel threshold, a pause is determined. The starting coordinates and duration of each pause are recorded to form a spatiotemporal record set containing multiple pause events. The Euclidean distance between adjacent pause points is calculated for the spatiotemporal record set. Pause points with a distance less than a preset threshold are grouped into the same region. The number of pauses in each region is counted as the distribution density value. At the same time, the duration of all pause events is accumulated to obtain the total pause duration. The distribution density value, the total pause duration, and the number of each category in the luggage size category sequence are weighted and summed, with large luggage having a weight of 3, medium luggage having a weight of 2, and small luggage having a weight of 1. The luggage burden index is calculated. The user's luggage burden level is determined based on the range of the luggage burden index. If the index is greater than the first preset threshold, it is determined to be a high burden level. If the index is less than the first preset threshold but greater than the second preset threshold, it is determined to be a medium burden level. If the index is less than the second preset threshold, it is determined to be a low burden level. The user's luggage burden level is thus determined. Based on the luggage burden level, analyze the movement speed and direction features in the user's posture image, and combine the pause location distribution and pause duration to identify the user's access facility selection preference. Determine whether the user prefers the elevator path by the difference in dwell time at the stairwell and elevator entrance. If the focus is on elevator routes, data on elevator usage frequency and waiting time, as well as channel traffic data, are obtained to determine the elevator congestion level and the distribution of potential congestion areas, and to generate the degree of association between elevators and luggage. By comparing facial recognition with the user's identity profile, and combining the luggage burden level, the degree of correlation between elevators and luggage, and the distribution of potential congestion areas, the system analyzes the user's elevator selection preferences and usage behavior patterns to generate elevator usage guidance suggestions. Based on the elevator usage guidance suggestions, the elevator congestion level, and the passage flow data, the route planning information is dynamically adjusted to generate a route adjustment plan suitable for the user. Based on the route adjustment plan, combined with the distribution of stopping positions and elevator route data, the adjusted elevator usage plan and alternative facility recommendations are pushed to the dynamic guidance system. The actual effect of users implementing the route adjustment plan is tracked through facial recognition to determine the target congestion prevention guidance.

2. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The process of collecting user dynamic images, extracting the outline features of carried items and body posture features, combining facial recognition to locate the target user carrying a suitcase, establishing a user identity profile, and obtaining an initial user image dataset includes: Video streams of people entering the detection area are collected. Moving target contour information is extracted from the video stream. Keypoint detection is used to extract the coordinates of key points on the shoulders, hips, and knees as body posture features, forming a target feature set. Morphological processing is performed on the contour information in the target feature set to identify the rectangular contour of a suitcase. If a rectangular contour matching the suitcase size is detected and has a stable positional relationship with the human body contour, the person is determined to be the target user carrying the suitcase. The facial region image and body image are then cropped. Facial feature vectors are extracted from the facial region image and compared with a user facial feature database to match user identity or assign a new user ID, which is then stored in the facial feature vector. The user ID, rectangular contour parameters, keypoint coordinates, and body image are integrated to construct a structured user identity profile, which is then stored in the initial user image dataset.

3. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The process of separating the user and their belongings, generating a user pose image and an item outline image, performing edge processing on the item outline image, and comparing it with a luggage shape template to determine whether there are luggage feature structures includes: A probability map is generated by an instance segmentation algorithm to indicate whether a pixel belongs to a human body or an object. Based on the probability map, pixels in the human body region are extracted to form a user pose image, and pixels in the non-human body region are used to form an initial contour image. An edge detection operator is applied to the initial contour image to mark edge points and perform morphological operations to obtain a continuous object contour boundary. The aspect ratio, rectangularity, and number of corner points of the object contour boundary are calculated to form a shape feature vector, which is compared with the standard feature range of a suitcase to determine whether a suitcase feature structure exists.

4. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The step of analyzing the user's movement speed and direction features in the user's posture image based on the luggage burden level, identifying the user's preference for access facilities, and determining whether the user prefers the elevator route based on the difference in dwell time at stairwells and elevator entrances includes: The user's center point coordinate changes are extracted from the user's posture image, and the instantaneous movement speed and main movement direction are calculated to generate a motion feature dataset. The user's trajectory is tracked based on the motion feature dataset, and the dwell time and position coordinates in the stairwell and elevator areas are recorded. The ratio of the dwell time at the elevator entrance to the dwell time at the stairwell entrance and the density of the dwell position distribution are calculated to determine whether the user is biased towards the elevator path.

5. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The step of analyzing the user's movement speed and direction features in the user's posture image based on the baggage burden level to identify the user's preference for access facilities includes: Based on the luggage burden level, the instantaneous velocity and velocity change rate of the user's position coordinates in the user posture image are calculated. The turning angle at the bifurcation point of the passage is recorded as direction selection data. The dwell time and line of sight angle are detected in the stairwell area, and a behavioral feature vector containing the velocity change rate, turning angle, dwell time, and line of sight angle is generated. The frequency of the user approaching the elevator, the waiting posture parameters, and the difference in dwell time in the behavioral feature vector are statistically analyzed. Combined with the number of arm movements and the step adjustment index, a weighted calculation is performed to determine the user's access facility selection preference.

6. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The process of acquiring elevator usage frequency and waiting time data, as well as passageway traffic data, to determine elevator congestion levels and potential congestion area distribution, and to generate the correlation between elevators and luggage includes: The elevator controller obtains the number of elevator door openings and closings and the operating cycle. The number of door openings and closings is calculated as the usage frequency, and the time interval from when a user presses the button to when they enter the elevator is recorded as the waiting time data. The monitoring equipment detects the flow density of people in the elevator hall and corridor areas. When the usage frequency and flow density exceed the threshold, the elevator is determined to be in a high-crowding level, and the elevator location and the distribution of flow density around it are recorded. The proportion of users carrying luggage is extracted, and the correlation coefficient between the usage frequency and the waiting time data is calculated as the degree of association between the elevator and luggage.

7. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The generation of a path adjustment scheme suitable for the user includes: Facial feature vectors are extracted using facial recognition and compared with user identity profiles to confirm user identity, obtain luggage burden level and historical behavior data; the distance from the user's location to each elevator is calculated, and combined with the number of waiting people and congestion level, the differences in user elevator selection frequency are analyzed to form a preference feature vector; based on the preference feature vector and elevator congestion level score, recommended elevators are selected to determine waiting time and luggage placement location, forming a guidance plan; the difference in length and travel time between the guidance plan path and the original path is calculated to determine the path adjustment plan.

8. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The dynamic adjustment of route planning information based on the elevator usage guidance suggestions, elevator congestion level, and passageway traffic data includes: Obtain the number of passengers in the elevator car and waiting area, calculate the remaining carrying capacity, and mark the high-load state; extract the transportation mode preference and the distribution of the user's usual activity area from the user identity profile; if the elevator is in a high-load state, calculate the estimated travel time of the escalator and the backup passage, combine the transportation mode preference to select the optimal path, generate guidance information including the main path and alternative paths, and update the path planning information.

9. The method for user behavior analysis of a MaaS platform based on face recognition according to claim 1, characterized in that, The process of pushing adjusted elevator usage plans and alternative facility recommendations to the dynamic guidance system, tracking the actual effects of users implementing the route adjustment plans through facial recognition, and determining target congestion prevention guidelines includes: Based on the path adjustment scheme, the factors triggering the pauses and the elevator operating load are extracted to generate a facility usage priority sequence; facial feature vectors are extracted through facial recognition to confirm the user's identity and record the real-time location coordinate sequence; the distance deviation between the real-time location coordinate sequence and the recommended path of the priority sequence is calculated, the number of deviations and the flow density are counted, congestion nodes are selected, diversion periods and alternative paths are generated, and a target congestion prevention guide is formed.

Images