Intelligent management platform for resort

Abstract

The invention relates to the technical field of intelligent management, and discloses a resort intelligent management platform, which comprises a monitoring acquisition module for acquiring a working area picture in real time and formulating a preliminary acquisition period duration, an information identification module for identifying a personnel face image in the preprocessed working area picture based on a face detection and identification combined model, extracting facial feature vectors; the method comprises the following steps: identifying personnel identity information based on Euclidean distance, determining whether the personnel is a worker or not, and identifying whether the dressing of the worker meets the standard or not when the personnel is identified, optimizing an acquisition period according to the statistical condition of identification data in an identification period, and performing early warning by an information early warning module when no worker appears in the acquisition period. And when the dressing information of the worker does not meet the standard, an alarm is given. According to the invention, the dynamic management of the personnel identity information is realized, and the actual operation management requirement is met.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology, and more specifically, to a smart management platform for resorts. Background Technology

[0002] As a vital venue for serving tourists, the operational and management efficiency of resorts directly impacts the tourist experience and overall service quality. Resorts typically comprise numerous functional areas, such as reception areas, recreational areas, accommodation areas, and security patrol areas, all requiring real-time monitoring and management by staff in various roles. To ensure the orderly, safe, and standardized operation of the resort, management needs to effectively supervise staff attendance, on-duty status, and adherence to regulations regarding the wearing of uniforms and safety equipment.

[0003] However, existing management methods remain relatively traditional, generally relying on manual patrols, attendance records, broadcast scheduling, and passive video surveillance. These methods have several problems in practical applications: on the one hand, manual methods suffer from low efficiency, high personnel costs, and the susceptibility of human subjective judgment to errors; on the other hand, while traditional video surveillance systems achieve real-time shooting and recording of scenes, they lack image recognition and automatic analysis capabilities, usually requiring manual supervision or post-event playback, and cannot achieve immediate identification and intelligent feedback.

[0004] Furthermore, existing video surveillance systems cannot automatically identify the identities of staff members in the footage, nor can they determine whether they are on duty or dressed appropriately. If vacancies occur, or staff are not wearing uniforms or are not dressed correctly, the system cannot automatically identify and issue warnings, which can easily lead to safety incidents or service problems.

[0005] Therefore, it is necessary to design and develop a smart management platform for resorts to solve the problems existing in current technologies. Summary of the Invention

[0006] In view of this, the present invention proposes a smart management platform for resorts, which aims to solve the problem that existing technologies mainly rely on manual patrols, attendance systems or traditional monitoring methods to determine whether staff are on duty, which cannot achieve automatic identification and monitoring of staff on duty status, resulting in low management efficiency and untimely response.

[0007] This invention proposes a smart management platform for resorts, comprising: The monitoring and acquisition module is used to collect images of the work area in real time and to determine the initial collection cycle duration based on the historical identification data of each work position. The information recognition module performs image preprocessing on the work area images collected in each acquisition cycle; it identifies the facial images of personnel in the preprocessed work area images based on a pre-established face detection and recognition joint model, and extracts facial feature vectors from the identified facial images; it compares the facial feature vectors with the staff feature vectors based on Euclidean distance to identify the personnel's identity information and confirm whether they are staff members; when identified as staff members, it identifies the staff member's clothing information and confirms whether the clothing information meets the standards. The information processing module confirms the recognition period and adaptively optimizes the acquisition period based on the statistical data of the recognition data within the recognition period; the recognition period includes multiple acquisition periods. The information early warning module determines the alarm type and issues an alarm when no staff members are present during the data collection period; it also determines the alarm type and issues an alarm when staff members are present during the data collection period and their attire does not meet the standards.

[0008] Furthermore, when performing image preprocessing on the work area images acquired in each acquisition cycle, the following steps are included: The image preprocessing includes image frame extraction, noise reduction, illumination compensation, and image enhancement.

[0009] Furthermore, the process of identifying facial images of individuals within the pre-processed work area image based on a pre-established joint model for face detection and recognition, and extracting facial feature vectors from the identified facial images, includes: Construct a historical image set based on facial images of historical individuals; The historical image set is sampled according to a preset ratio to obtain a training subset and a test subset; Obtain the pre-established joint face detection and recognition model, iteratively train the joint face detection and recognition model according to the training subset, evaluate the iteratively trained joint face detection and recognition model according to the test subset, and determine whether to stop iterative training based on the evaluation value.

[0010] Furthermore, the process of identifying facial images of individuals within the pre-processed work area image based on a pre-established joint model for face detection and recognition, and extracting facial feature vectors from the identified facial images, includes: The joint face detection and recognition model includes MTCNN and ArcFace; MTCNN consists of three cascaded convolutional neural networks: a P-Net candidate box generation network, an R-Net candidate box refinement network, and an O-Net output network. The P-Net candidate box generation network takes the working area image as input, integrates face detection and keypoint localization to detect face regions within the image, proposes candidate boxes for face regions, and outputs the candidate boxes and face probabilities. The R-Net candidate box refinement network takes the candidate boxes from the P-Net as input, filters false positive candidate boxes, and outputs the candidate boxes for face regions. The O-Net output network takes the face region candidate boxes processed by the R-Net as input, extracts and detects facial keypoints, and outputs the face image bounding box and facial keypoints. The face image bounding boxes output from MTCNN are input into ArcFace. The face image bounding boxes are processed through convolutional layers and residual modules to extract multi-level semantic features. Spatial information is compressed through global average pooling layers, and a fixed-dimensional facial feature vector is output through fully connected layers.

[0011] Furthermore, when outputting a fixed-dimensional facial feature vector through a fully connected layer, the following steps are included: The facial feature vectors are subjected to L2 normalization, which uniformly scales each dimension of the facial feature vectors. The length of the facial feature vectors is standardized to 1, and all facial feature vectors are mapped onto a unit sphere in high-dimensional space.

[0012] Furthermore, when identifying personnel identity information by comparing the facial feature vector with the staff feature vector based on Euclidean distance, the process includes: A distance threshold is pre-set, and the geometric distance between the facial feature vector output by the face detection and recognition joint model and the staff feature vector is calculated in the feature space. The identity information is determined by comparing the geometric distance value with the distance threshold.

[0013] Furthermore, when identifying staff attire information and confirming whether the attire information conforms to standards, this includes: The entire image corresponding to the staff is divided into multiple image semantic regions, and key clothing areas in the image semantic regions are identified. Feature extraction and category determination are performed on each key clothing area.

[0014] Furthermore, when performing feature extraction and category determination for each of the key clothing areas, the process includes: Image feature extraction is performed on each of the key clothing areas, including color features, texture features, and structural features. After feature extraction, the features of each key clothing area are compared with a standard template library to determine the category.

[0015] Furthermore, when pre-setting an identification period and adaptively optimizing the acquisition period based on the statistical data of the identification data within the identification period, the following steps are included: A time sliding window is established, the size of which is the duration of the acquisition period. The identification data within the acquisition period is identified through the time sliding window. Each acquisition period corresponds to one identification data. The identification data corresponding to each acquisition period within the identification period is statistically analyzed to determine the statistical data.

[0016] Furthermore, when optimizing the duration of the data collection period based on the aforementioned statistical data, the following steps are included: If the same staff member is identified multiple times in each collection cycle within the identification cycle, the collection cycle duration is shortened. If no staff member is identified in any collection cycle within the identification period, the collection cycle duration will be increased.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: Based on a joint model of face detection and recognition, it automatically extracts and identifies personnel facial images from the collected work area images, and combines Euclidean distance to compare the personnel feature vectors, thereby achieving rapid confirmation of personnel identity and solving the problem of the inability to automatically identify personnel identity in traditional methods; based on the identification of personnel identity, it further identifies their clothing information, including work clothes, safety helmets, name tags, etc., automatically judges whether they meet preset standards, and promptly detects behaviors such as not wearing work clothes or wearing them improperly, enhancing the accuracy of on-site supervision; by setting an identification cycle and statistically analyzing the identification data within the cycle, it dynamically shortens or widens the subsequent collection cycle based on the frequency and pattern of personnel appearance within the identification cycle, avoiding frequent invalid collection and enhancing the reliability and practicality of the early warning function; an information early warning module is set up so that when no personnel are identified within the collection cycle or the identified personnel are improperly dressed, it can automatically determine the alarm type and promptly issue an alarm message to prompt management personnel to handle the situation, improving the response speed to abnormal events. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a structural block diagram of a smart management platform for resorts provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In view of the above issues, please refer to Figure 1 As shown, this application proposes a smart management platform for resorts, comprising: The monitoring and acquisition module is used to collect images of the work area in real time and to determine the initial collection cycle duration based on the historical identification data of each work position. The information recognition module performs image preprocessing on the work area images collected in each acquisition cycle; it identifies the facial images of personnel in the preprocessed work area images based on a pre-established face detection and recognition joint model, and extracts facial feature vectors from the identified facial images; it compares the facial feature vectors with the staff feature vectors based on Euclidean distance to identify the personnel's identity information and confirm whether they are staff members; when identified as staff members, it identifies the staff member's clothing information and confirms whether the clothing information meets the standards. The information processing module confirms the recognition period and adaptively optimizes the acquisition period based on the statistical data of the recognition data within the recognition period; the recognition period includes multiple acquisition periods. The information early warning module determines the alarm type and issues an alarm when no staff members are present during the data collection period; it also determines the alarm type and issues an alarm when staff members are present during the data collection period and their attire does not meet the standards.

[0021] Specifically, the monitoring and acquisition module is deployed in each work area, enabling continuous monitoring of on-site images or timed acquisition according to a predetermined rhythm. It combines historical recognition data to statistically analyze the activity patterns of each work position and sets the initial acquisition cycle duration. The information recognition module preprocesses the image data acquired within each acquisition cycle. The processed images are input into a preset face detection and recognition joint model to detect and identify faces appearing in the images. Simultaneously, it extracts the facial feature vectors of the identified faces and compares them with the feature vectors of registered staff in the platform using Euclidean distance to determine if the person is a formal employee. If confirmed as an employee, the system further identifies their clothing area based on image segmentation to determine if they are wearing compliant work clothes, name tags, and other identification items to assist in determining whether the person is in a compliant state. The information processing module comprises multiple collection cycles within each identification cycle. At the end of each cycle, it performs data statistics and analysis on all identification results, such as identification success rate, staff on-duty frequency, number of instances of non-compliant attire, and the total number of staff present. Based on this statistical information, it assesses the rationality of the current collection cycle setting and dynamically optimizes the duration or frequency of subsequent collection cycles. For example, if one identification cycle contains five collection cycles, performing data statistics and analysis on all identification results within those five collection cycles means performing data statistics and analysis on all identification results within those five collection cycles. The information early warning module automatically triggers an early warning mechanism when the system detects an anomaly. If no staff are detected within a collection cycle, the system determines it as a potential vacancy, automatically identifies the appropriate alarm type, and issues an alarm. If staff are detected but their attire does not meet regulations, such as not wearing uniforms or name tags, a violation alarm will also be automatically generated. Early warning information can be pushed to on-duty personnel in real time via the console, management backend, or mobile terminals, allowing management personnel to understand the situation on-site and take intervention measures immediately.

[0022] Understandably, by constructing an intelligent monitoring system comprised of a monitoring and data acquisition module, an information identification module, an information processing module, and an information early warning module, the platform achieves fully automated identification and anomaly alerts for staff on-duty status and attire. This platform employs a joint face detection and recognition model, combined with Euclidean distance comparison, to accurately extract and identify the identity information of personnel appearing in the footage, further identifying their attire and using intelligent algorithms to determine compliance with dress codes. This achieves dual identification of "on-duty status" and "compliance," enhancing the accuracy and reliability of management. The information processing module, by statistically analyzing the identification results within the recognition period, achieves adaptive optimization of the data acquisition cycle, enabling the system's data acquisition strategy to dynamically adjust according to the actual on-site conditions, improving the system's intelligence and adaptability. The platform's built-in information early warning mechanism automatically determines the alarm type and issues an immediate alert when it detects unattended positions or abnormal attire, achieving closed-loop control from anomaly detection to proactive response, thus improving the resort's operational management efficiency.

[0023] In some embodiments of this application, image preprocessing of the work area images acquired in each acquisition cycle includes: The image preprocessing includes image frame extraction, noise reduction, illumination compensation, and image enhancement.

[0024] Specifically, within each acquisition cycle, the system first performs image preprocessing on the acquired working area image. The preprocessing process includes four steps: image frame extraction, denoising, illumination compensation, and image enhancement. Image frame extraction selects keyframes or multiple frames from the video stream to ensure the analyzed images are representative and timely. Denoising employs spatial or temporal filtering algorithms to effectively remove random noise and interference from the image, improving image quality. Illumination compensation automatically adjusts the image's brightness and contrast to correct for the effects of ambient light changes, ensuring the clarity of faces and clothing areas. Image enhancement further improves image detail through techniques such as sharpening, edge enhancement, or color correction.

[0025] Understandably, systematic image preprocessing of the acquired work area images improves the quality and clarity of the input images, reduces the interference of environmental noise and lighting changes on subsequent face and clothing recognition algorithms, and improves recognition accuracy. Image frame extraction ensures the representativeness of the analyzed images, avoiding resource waste caused by invalid frames; denoising eliminates random noise in the images, making facial and clothing features clearer; lighting compensation technology balances image brightness under different lighting conditions, improving recognition capabilities in complex environments; image enhancement further highlights image details and enhances feature distinguishability. The optimization of the overall preprocessing workflow effectively improves the accuracy of detecting staff identity and clothing.

[0026] In some embodiments of this application, the process of identifying facial images of people within a pre-processed work area image based on a pre-established joint model for face detection and recognition, and extracting facial feature vectors from the identified facial images, includes: Construct a historical image set based on facial images of historical individuals; The historical image set is sampled according to a preset ratio to obtain a training subset and a test subset; Obtain the pre-established joint face detection and recognition model, iteratively train the joint face detection and recognition model according to the training subset, evaluate the iteratively trained joint face detection and recognition model according to the test subset, and determine whether to stop iterative training based on the evaluation value.

[0027] Specifically, based on facial images detected in the work area, a historical image set is continuously built. This set includes facial image samples collected under different lighting conditions, angles, and poses, ensuring sample diversity and data quality. After construction, the system divides the historical image set according to a preset ratio (e.g., 8:2 or 7:3) to generate training and testing subsets. The training subset is used to update model parameters, while the testing subset is used to verify the model's generalization ability. Subsequently, the system loads a preset joint face detection and recognition model and uses the training subset as input data to perform multiple rounds of iterative training, continuously optimizing its feature extraction capabilities and recognition accuracy. After each round of iterative training, the system uses the testing subset to evaluate the current model's performance. Evaluation metrics may include recognition accuracy, recall, and loss function value. Based on the evaluation results and set termination conditions (e.g., the evaluation value reaches a threshold or there is no significant improvement after several consecutive rounds), the system determines whether to stop iterative training.

[0028] Understandably, the continuous construction of historical image sets ensures the comprehensiveness and diversity of training samples, enabling the model to learn facial features adaptable to different lighting, angles, expressions, and occlusion conditions. Through reasonable sample partitioning and iterative training mechanisms, the model can learn fully while its generalization performance is monitored in real time using test subsets, avoiding overfitting or undertraining. The evaluation and feedback mechanism during model training allows the system to dynamically determine when to terminate training based on recognition results, improving overall training efficiency.

[0029] In some embodiments of this application, the process of identifying facial images of people within a pre-processed work area image based on a pre-established joint model for face detection and recognition, and extracting facial feature vectors from the identified facial images, includes: The joint face detection and recognition model includes MTCNN and ArcFace; MTCNN consists of three cascaded convolutional neural networks: a P-Net candidate box generation network, an R-Net candidate box refinement network, and an O-Net output network. The P-Net candidate box generation network takes the working area image as input, integrates face detection and keypoint localization to detect face regions within the image, proposes candidate boxes for face regions, and outputs the candidate boxes and face probabilities. The R-Net candidate box refinement network takes the candidate boxes from the P-Net as input, filters false positive candidate boxes, and outputs the candidate boxes for face regions. The O-Net output network takes the face region candidate boxes processed by the R-Net as input, extracts and detects facial keypoints, and outputs the face image bounding box and facial keypoints. The face image bounding boxes output from MTCNN are input into ArcFace. The face image bounding boxes are processed through convolutional layers and residual modules to extract multi-level semantic features. Spatial information is compressed through global average pooling layers, and a fixed-dimensional facial feature vector is output through fully connected layers.

[0030] Specifically, to achieve accurate recognition and efficient feature extraction of facial images of people within the preprocessed work area, a joint deep learning-based face detection and recognition model was adopted, including the MTCNN face detection model and the ArcFace face recognition model. This joint model can efficiently complete the entire process from locating face regions and extracting key points from the original image to generating high-discrimination facial feature vectors. MTCNN (Multi-task Cascaded Convolutional Neural Network) is a multi-task cascaded convolutional neural network with a three-stage network structure: a candidate box generation network P-Net, a candidate box refinement network R-Net, and an output network O-Net. P-Net, as the initial detection network, receives the input work area image and uses lightweight convolutional layers to quickly generate multiple candidate boxes that may contain faces, while predicting the probability of face presence and the location of key points for each candidate box. R-Net further filters and refines the candidate boxes output by P-Net, eliminating a large number of false positive regions and retaining face regions with higher confidence, generating more accurate face candidate boxes. O-Net performs fine-tuning on the retained high-confidence candidate boxes, further extracting five key points (e.g., corners of the eyes, tip of the nose, and corners of the mouth) and outputting the final face image bounding box and facial key point coordinates, providing a precise foundation for subsequent face alignment and recognition. The face image bounding box output by MTCNN is then fed as input into the ArcFace recognition network. ArcFace employs a deep convolutional neural network architecture, with its front end including multiple convolutional layers and residual modules for multi-scale semantic feature extraction of face images, preserving detailed texture and structural information of the face; subsequently, a global average pooling layer is used to compress spatial features to reduce dimensionality and computational cost, and finally, a fully connected layer outputs a high-discrimination facial feature vector with a fixed dimension (e.g., 512 dimensions or other set dimensions). This feature vector has good discriminative ability and can be used to compare with the feature vector of staff members using Euclidean distance to achieve high-precision personnel identification.

[0031] Understandably, MTCNN, as the front-end detection module, employs a three-stage cascaded structure. Through P-Net, R-Net, and O-Net, it progressively filters and refines face candidate regions, effectively filtering background interference and false positive candidate boxes in complex scenes. This ensures that the final output face image bounding boxes have high localization accuracy and keypoint accuracy, providing a precise foundation for subsequent facial feature extraction. ArcFace, as the recognition module, can extract multi-layer semantic features based on high-quality aligned face images through deep convolutional networks and residual modules. It then generates highly discriminative facial feature vectors through global average pooling and fully connected structures. These facial feature vectors not only maintain good stability and discriminative power under different lighting, angles, and expression changes, but also facilitate subsequent identity verification using Euclidean distance, resulting in more accurate and reliable recognition results.

[0032] The joint model achieves efficient synergy in both detection and recognition: MTCNN enables rapid localization, while ArcFace facilitates deep recognition, making the system more adaptable to practical applications, especially suitable for complex environments such as resort areas with frequent changes in lighting and diverse populations.

[0033] In some embodiments of this application, when outputting a fixed-dimensional facial feature vector through a fully connected layer, the following steps are included: The facial feature vectors are subjected to L2 normalization, which uniformly scales each dimension of the facial feature vectors. The length of the facial feature vectors is standardized to 1, and all facial feature vectors are mapped onto a unit sphere in high-dimensional space.

[0034] Specifically, each dimension of the facial feature vector is uniformly scaled, normalizing its overall vector length to 1, essentially projecting the vector onto a unit sphere. This normalization operation ensures that feature vectors extracted from different face images have the same magnitude, thus eliminating numerical amplitude differences caused by factors such as image illumination, scale, and contrast, focusing on directional differences between vectors for effective comparison. In this way, all facial feature vectors are mapped onto a unit sphere in a high-dimensional space, making subsequent identity comparison using Euclidean distance more stable and accurate.

[0035] Understandably, L2 normalization unifies the magnitude of all facial feature vectors, allowing the comparison of facial features to focus only on the directional differences of the vectors, eliminating variations in feature amplitude caused by factors such as image brightness, contrast, and facial pose, thereby enhancing recognition accuracy. The feature distribution structure on a unit sphere is more standardized, facilitating efficient distance measurement in high-dimensional spaces, especially suitable for commonly used similarity assessment methods such as Euclidean distance, effectively reducing the risk of false positives and false negatives.

[0036] In some embodiments of this application, when comparing the facial feature vector with the staff feature vector based on Euclidean distance to identify staff identity information, the following steps are included: A distance threshold is pre-set, and the geometric distance between the facial feature vector output by the face detection and recognition joint model and the staff feature vector is calculated in the feature space. The identity information is determined by comparing the geometric distance value with the distance threshold.

[0037] Specifically, after the face detection and recognition joint model (MTCNN+ArcFace) outputs the facial feature vector from the currently acquired image, it calculates the Euclidean geometric distance between this vector and multiple facial feature vectors stored in the historical database in the feature space. This Euclidean distance reflects the similarity between the current face image and historically registered face images in the feature dimension space; the smaller the distance value, the closer the two are, indicating greater similarity in facial features. The currently calculated distance value is compared with a set distance threshold: if the distance between a certain employee's feature vector and the current feature vector is less than or equal to the threshold, the current identified object can be determined to be the employee corresponding to that historical identity; if all comparison results are greater than the threshold, it is considered that the facial feature cannot be effectively matched with historical data, that is, it cannot be confirmed that the person is a registered employee. The specific value of the distance threshold is determined based on the geometric distance value obtained from the actual measurement when determining whether someone is a staff member. The distance threshold needs to be determined through multiple measurements and multiple sets of identification data to determine the specific value of each set of geometric distance values. By determining the distance threshold through multiple sets of geometric distance values, it is ensured that when someone is identified as a staff member after the distance threshold is determined, the identification data corresponds to only one staff member, thus achieving accurate identification of each staff member.

[0038] Understandably, using Euclidean distance as a similarity metric between feature vectors enables rapid comparison of different facial features in high-dimensional space with low computational cost and fast convergence, meeting the needs of real-time processing applications. By presetting a reasonable distance threshold, the system can automatically determine whether the current person is a registered staff member based on the distance value during the comparison process, avoiding recognition ambiguity caused by error accumulation or image interference, and reducing the false recognition rate and false negative rate.

[0039] In some embodiments of this application, identifying staff attire information and confirming whether the attire information conforms to standards includes: The entire image corresponding to the staff is divided into multiple image semantic regions, and key clothing areas in the image semantic regions are identified. Feature extraction and category determination are performed on each key clothing area.

[0040] Specifically, an image segmentation algorithm is used to semantically divide the image of the worker, separating it into multiple image regions with specific semantic meanings, such as the head region, upper body region, lower body region, and shoulder and neck region. Based on predefined key clothing region definitions (e.g., the upper garment region for identifying work clothes, the chest region for identifying name tags, etc.), the corresponding clothing region images are extracted from the semantic regions of the image, and feature extraction and image classification processing are performed on them. Each key clothing region is determined to be either "compliant with standards" or "non-compliant with standards." For example, based on the uniform color and style standards of work clothes, the upper body region is compared and identified; the color and outline of the lower garment are identified and judged; and the existence and location features of the name tag are confirmed.

[0041] Understandably, semantic segmentation technology can effectively isolate the target clothing area, reduce interference from background and non-critical information, and improve the accuracy and stability of clothing recognition. By extracting rich features such as color, texture, and shape from key areas, accurate judgment of different clothing categories is achieved, ensuring effective verification that clothing meets job standards.

[0042] In some embodiments of this application, feature extraction and category determination are performed on each of the key clothing areas, including: Image feature extraction is performed on each of the key clothing areas, including color features, texture features, and structural features. After feature extraction, the features of each key clothing area are compared with a standard template library to determine the category.

[0043] Specifically, to accurately identify and classify key clothing areas, the system performs image feature extraction on each key clothing area. Feature extraction mainly includes color features, texture features, and structural features. Color features are extracted by statistically analyzing the color distribution of pixels within the area, extracting color histograms, and identifying the dominant color to determine the uniform color of work clothes. Texture features utilize methods such as Local Binary Pattern (LBP) and Gray-Level Co-occurrence Matrix (GLCM) to capture fabric texture and pattern features within the area, aiding in distinguishing different materials and styles. Structural features are extracted using edge detection and shape descriptors to extract the contour shape and geometric structure information of the area, identifying the presence and specific form of work badges. After extracting multi-dimensional features, the system compares the obtained features with feature samples of corresponding key clothing areas in a pre-established standard template library, using methods such as distance metrics or similarity calculations to determine the clothing category corresponding to the current area. This category determination process effectively identifies whether workers are wearing standard work clothes and work badges, ensuring the standardization and consistency of attire.

[0044] In some embodiments of this application, when a recognition period is pre-established and the collection period is adaptively optimized based on the statistical data of the recognition data within the recognition period, the following steps are included: A time sliding window is established, the size of which is the duration of the acquisition period. The identification data within the acquisition period is identified through the time sliding window. Each acquisition period corresponds to one identification data. The identification data corresponding to each acquisition period within the identification period is statistically analyzed to determine the statistical data.

[0045] Specifically, a time-sliding window is established, with a window size equal to the duration of the current data collection period, to cover and capture the recognition data within that period. Within each collection period, the system collects and generates corresponding recognition data, such as recognition success rate, frequency of staff on duty, number of instances of non-compliant attire, and number of staff present. The recognition period comprises multiple consecutive collection periods. The recognition data from all collection periods within a given period are statistically summarized to form overall statistical data. By analyzing this statistical data, the system can accurately grasp the recognition effect and on-site personnel status of the current collection period. Based on fluctuations and trends in the recognition data, the system dynamically adjusts the duration of the collection period, achieving adaptive optimization of the collection cycle.

[0046] Understandably, the sliding window design ensures the continuity and real-time nature of data statistics, enabling the system to monitor the recognition effect and the changing trends of on-site personnel status in real time, and promptly detect anomalies. Based on the comprehensive statistics of recognition data from multiple collection cycles, the rationality of the current collection cycle can be scientifically assessed, avoiding unreasonable collection cycle settings due to excessively long or short cycles, and effectively balancing system performance and resource consumption. This dynamic adjustment mechanism enhances the system's adaptability to dynamic changes in the on-site environment and personnel, improves the accuracy and completeness of recognition data, and ensures timely detection and feedback of staff identity and attire status.

[0047] In some embodiments of this application, optimizing the duration of the data collection period based on the statistical data includes: If the same staff member is identified multiple times in each collection cycle within the identification cycle, the collection cycle duration is shortened. If no staff member is identified in any collection cycle within the identification period, the collection cycle duration is increased.

[0048] Specifically, if the same worker is identified multiple times in each collection cycle within the identification period, it indicates that the collection cycle is set too long, leading to repeated collection of the same target and wasting resources. Based on this, the system determines that the current collection cycle is too long and needs to be appropriately shortened to reduce the generation of redundant data and improve real-time performance and monitoring accuracy. Conversely, if no worker is identified in any collection cycle within the identification period, it indicates that the current collection cycle is too short, failing to obtain effective target information, resulting in wasted system resources and increased processing burden. Other identification situations within the identification period do not require adjustment of the collection cycle.

[0049] In summary, the joint model based on face detection and recognition automatically extracts and identifies facial images of personnel from the collected work area footage. By combining Euclidean distance with the personnel's feature vectors, it achieves rapid identification of personnel, solving the problem of traditional methods' inability to automatically identify personnel. Beyond identifying personnel, it further identifies their attire information, including work clothes, safety helmets, and name tags, automatically determining whether they meet preset standards and promptly detecting behaviors such as not wearing work clothes or improper attire, enhancing the accuracy of on-site supervision. By establishing an identification cycle and statistically analyzing the identification data within that cycle, the subsequent collection cycle can be dynamically shortened or extended based on the frequency and patterns of personnel appearance within that cycle, avoiding frequent invalid data collection and enhancing the reliability and practicality of the early warning function. An information early warning module is established; when no personnel are identified within the collection cycle or when identified personnel are improperly dressed, the alarm type can be automatically determined and an alarm message can be issued promptly, prompting management personnel to take action, improving the response speed to abnormal events.

[0050] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A smart management platform for a holiday resort, characterized in that, The method comprises the following steps: A monitoring acquisition module is used to acquire real-time working area pictures, and a preliminary acquisition cycle length is determined according to historical identification data of each working post; An information identification module is used to perform image preprocessing on the working area pictures acquired in each acquisition cycle; a pre-established face detection and recognition joint model is used to identify personnel face images in the preprocessed working area pictures, and a face feature vector is extracted from the identified personnel face images; the face feature vector is compared with a staff feature vector based on the Euclidean distance to identify personnel identity information and confirm whether the personnel are staff; when the personnel are identified as staff, staff clothing information is identified, and it is confirmed whether the clothing information meets the standard; An information processing module is used to confirm an identification cycle, and the acquisition cycle is adaptively optimized according to the statistical situation of the identification data in the identification cycle; the identification cycle comprises a plurality of acquisition cycles; An information early warning module is used to determine an alarm type and issue an alarm when no staff appear in the acquisition cycle; when staff appear in the acquisition cycle and the staff clothing information does not meet the standard, the alarm type is determined and an alarm is issued.

2. The intelligent management platform for holiday resort according to claim 1, characterized in that, When the working area pictures acquired in each acquisition cycle are preprocessed, the image preprocessing comprises image frame extraction, denoising processing, illumination compensation and image enhancement. When the preprocessed working area pictures are identified based on the pre-established face detection and recognition joint model, and a face feature vector is extracted from the identified personnel face images, the following steps are included:

3. The intelligent management platform for holiday resort according to claim 2, characterized in that, A historical image set is constructed according to historical personnel face images; The historical image set is sampled according to a preset proportion to obtain a training subset and a test subset; The pre-established face detection and recognition joint model is obtained, and the face detection and recognition joint model is iteratively trained according to the training subset; the face detection and recognition joint model after iterative training is evaluated according to the test subset, and it is determined whether to stop iterative training according to the evaluation value. When the preprocessed working area pictures are identified based on the pre-established face detection and recognition joint model, and a face feature vector is extracted from the identified personnel face images, the following steps are included:

4. The intelligent management platform for holiday resort according to claim 3, characterized in that, The face detection and recognition joint model comprises MTCNN and ArcFace; MTCNN is composed of three cascaded convolutional neural networks, including a P-Net candidate frame generation network, an R-Net candidate frame refining network and an O-Net output network; the P-Net candidate frame generation network is used to input the working area picture, detect the face region contained in the image by combining face detection and key point positioning, propose a face region candidate frame, and output the face region candidate frame and a face probability; the R-Net candidate frame refining network is used to input the face region candidate frame of the P-Net, filter false positive candidate frames, and output a face region candidate frame; the O-Net output network is used to input the face region candidate frame processed by the R-Net, extract face key points for detection, and output a face image frame and face key points; ​ The face image frame output in the MTCNN is input into the ArcFace, and the face image frame is subjected to a convolution layer and a residual module to extract multi-level semantic features, compress spatial information through a global average pooling layer, and output a fixed-dimensional face feature vector through a fully connected layer.

5. The intelligent management platform for holiday resort according to claim 4, characterized in that, When outputting a fixed-dimensional face feature vector through a fully connected layer, the method comprises: The face feature vector is subjected to L2 normalization processing, each dimension of the face feature vector is uniformly scaled, the length of the face feature vector is standardized to 1, and all face feature vectors are mapped on a unit sphere in a high-dimensional space.

6. The intelligent management platform for a resort according to claim 5, wherein, When identifying personnel identity information based on the Euclidean distance comparison between the face feature vector and the staff feature vector, the method comprises: A distance threshold is set in advance, the geometric distance value in the feature space between the face feature vector output by the face detection and recognition joint model and the staff feature vector is calculated, and the size of the geometric distance value and the distance threshold is compared to determine the identity information.

7. The intelligent management platform for a resort according to claim 6, wherein, When identifying staff clothing information and confirming whether the clothing information meets the standard, the method comprises: The entire image corresponding to the staff is divided into multiple image semantic regions, key clothing regions in the image semantic regions are identified, and feature extraction and category judgment are performed on each key clothing region.

8. The intelligent management platform for a resort according to claim 7, wherein, When performing feature extraction and category judgment on each key clothing region, the method comprises: Image feature extraction is performed on each key clothing region, and the extraction content includes color features, texture features, and structure features; after feature extraction is completed, the features of each key clothing region are compared with a standard template library to perform category judgment.

9. The intelligent management platform for a resort according to claim 8, wherein, When confirming the recognition period and adaptively optimizing the collection period according to the statistical situation of the recognition data in the recognition period, the method comprises: A time sliding window is established, the size of the time sliding window is the length of the collection period, the recognition data in the collection period is recognized through the time sliding window; each collection period corresponds to the recognition data, the statistical situation data is determined by counting the recognition data corresponding to each collection period in the recognition period.

10. The intelligent management platform for holiday resort according to claim 9, characterized in that, When optimizing the length of the collection period according to the statistical situation data, the method comprises: When the same staff is identified multiple times in each collection period in the recognition period, it is determined to shorten the length of the collection period; When no staff is identified in each collection period in the recognition period, it is determined to increase the length of the collection period.

Images