Elevator monitoring image data synthesis method and system based on monitoring video

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an elevator monitoring image data synthesis method and system based on a monitoring video. The method comprises the steps that target elevator structure parameters and a real elevator empty image are obtained, and a virtual elevator empty image and a corresponding virtual camera are constructed according to the target elevator structure parameters and the real elevator empty image; extracting an edge line draft of the virtual elevator empty image as a layout constraint, coding the scene description text as a text constraint, and extracting a real elevator empty image feature as a reference vector; inputting the constraints into a denoising network, and generating foreground-free synthetic elevator empty images in batches; and finally, obtaining each main body reference image based on a virtual camera, and generating an elevator monitoring image data set by taking the foreground-free synthesized elevator empty image as a base background. According to the method, high-quality and high-fidelity elevator monitoring images can be automatically generated, high-quality training data can be provided for a deep learning model, and the recognition and early warning capability of an intelligent monitoring system can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method and system for synthesizing elevator monitoring image data based on surveillance video. Background Technology

[0002] Elevators, as an indispensable vertical transportation tool in modern urban life, are widely used but also bring various safety hazards. With the rapid development of artificial intelligence and computer vision technologies, intelligent monitoring using deep learning models has gradually become a trend. Although this intelligent monitoring technology has improved the scientific nature and efficiency of elevator monitoring to some extent, many problems still exist.

[0003] Specifically, the performance of the aforementioned deep learning models largely depends on the quality and quantity of training image data. However, data collection in elevator monitoring scenarios is often constrained by factors such as privacy protection, limited collection conditions, and scarce samples, resulting in a severe shortage of image data available for model training. This greatly limits the performance optimization and practical application effects of intelligent monitoring models.

[0004] To overcome data bottlenecks, engineers often use AIGC (Artificial Intelligence Generated Content) technology to synthesize training image data. However, mainstream image generation models show significant inadequacy when facing high-precision engineering scenarios such as elevator monitoring. Specifically, there are two core challenges: First, the model struggles to generate the slender structures commonly found in elevator scenes (such as pet leashes) in a coherent and realistic manner. It is prone to breakage, segmentation, or morphological distortion, failing to reproduce the "slender and continuous" physical structural characteristics of the leash. This results in the generated leash lacking geometric consistency, breakage, or morphological distortion in the real scene. Second, in complex scenarios involving multiple subjects (such as people, pets, and leashes), existing technologies cannot effectively coordinate the physical rationality and structural integrity of the generated content. This often results in interaction relationships that violate physical logic or incomplete main components, causing the generated images to fail to meet the standards for high-precision model training.

[0005] Furthermore, even if existing technologies can generate high-quality images, they cannot accurately extract pixel-level label information for slender objects (such as traction ropes). Due to the insufficient accuracy of traditional segmentation models on such targets, the effectiveness of using synthetic data for training deep learning models is significantly affected. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for synthesizing elevator monitoring image data based on surveillance video. This method can automatically generate high-quality, high-fidelity elevator monitoring images, providing excellent training image data for deep learning models; it can accurately identify and segment slender objects, systematically improving the identification and early warning capabilities of intelligent monitoring methods.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention provides a method for synthesizing elevator monitoring image data based on surveillance video, the method comprising:

[0009] S01, acquire the structural parameters of the target elevator and the actual empty elevator image captured by the monitoring camera, construct a virtual empty elevator image of the target elevator based on the structural parameters; set a virtual camera corresponding to the virtual empty elevator image based on the camera parameters of the monitoring camera.

[0010] S02, perform image edge extraction on the virtual elevator empty image to obtain a line drawing image, and use the line drawing image as a layout constraint;

[0011] S03, Based on the text encoder, feature extraction is performed on the scene description text of the virtual elevator empty image to obtain a text vector, and the text vector is used as a text constraint;

[0012] S04, Based on the image visual language model, extract the image features from the real elevator empty image as the corresponding reference vector;

[0013] S05, input the layout constraints, the text constraints and the reference vector into the denoising network model to generate different foreground-free synthetic elevator empty images in batches;

[0014] S06, Based on the virtual camera, obtain individual reference images corresponding to various subjects in the reference image, and use the foreground-free synthesized empty elevator image as the base background to generate an elevator monitoring image dataset;

[0015] As a preferred embodiment of the present invention, the denoising network model is constructed based on the open-source image generation models Flux-canny and Flux-fill; the Flux-canny model is the background core generation model, and the Flux-fill model is the foreground core generation model.

[0016] As a preferred embodiment of the present invention, step S06, before generating the elevator monitoring image dataset, further includes a sub-step, specifically:

[0017] The masking area between various subjects is determined based on the masking control method;

[0018] In a virtual elevator simulation scenario, various subject models driven by a skeleton are built, the perspective of the virtual camera is matched with the spatial dimension of the virtual elevator simulation scenario, and individual images corresponding to various subjects are exported.

[0019] The Canny edge extraction algorithm is used to process individual images, outputting the corresponding shape and structural features and generating corresponding structured line drawings;

[0020] A virtual foreground image corresponding to a subject is used as the reference input of the Redux model. Based on the preset detailed content, it is encoded in the Redux model and the detailed constraints are output.

[0021] The masked area and structured line drawing are input into the ControlNet network, and the detail constraints are input into the Flux-fill model to output a complete elevator monitoring image.

[0022] As a preferred embodiment of the present invention, the virtual elevator empty image is constructed based on Blender software combined with elevator standard setting scene physical rules.

[0023] As a preferred embodiment of the present invention, the camera parameters include the field of view, distortion parameters, and image resolution of a real camera.

[0024] As a preferred embodiment of the present invention, the denoising network model uses a ControlNet network and a VAE encoder to perform feature encoding on the line drawing image to generate spatial structural feature constraints.

[0025] As a preferred embodiment of the present invention, the subject includes a person, a non-human creature, and related objects; the generation order of various subjects in the generation process of the virtual elevator empty image is as follows: first, a person model is generated; second, a non-human creature model is generated; and finally, a related object model between the person model and the non-human creature model is generated.

[0026] As a preferred embodiment of the present invention, in S03, the step of generating the text vector specifically includes:

[0027] Input a real-world empty elevator image into a deployed large language model to generate scene description text;

[0028] The scene description text is semantically transformed using a dual CLIP model, and the corresponding text vector is generated.

[0029] Secondly, the present invention provides an elevator monitoring image data synthesis system based on surveillance video, the system comprising:

[0030] The data acquisition and modeling module is used to acquire the structural parameters of the target elevator and the real empty elevator image captured by the monitoring camera, and to construct a virtual empty elevator image of the target elevator based on the structural parameters; it is also used to set a virtual camera corresponding to the virtual empty elevator image based on the camera parameters of the monitoring camera.

[0031] The layout constraint generation module is used to extract image edges from the virtual elevator empty image to obtain a line drawing image, and use the line drawing image as a layout constraint.

[0032] The text constraint generation module is used to extract features from the scene description text of the virtual elevator empty image based on the text encoder, obtain a text vector, and use the text vector as a text constraint.

[0033] The reference vector generation module is used to extract image features from the real empty elevator image based on the image visual language model, and use them as the corresponding reference vector.

[0034] An image generation module, which has a built-in denoising network model, is used to receive the layout constraints, text constraints and reference vectors, and generate different foreground-free synthetic elevator empty images in batches.

[0035] The dataset construction module is used to acquire individual reference images corresponding to various subjects in the reference image based on the virtual camera, and use the foreground-free synthesized elevator empty image as the base background to synthesize elevator monitoring images and construct an elevator monitoring image dataset.

[0036] Thirdly, the present invention also provides a computer device, including a processor and a memory, wherein the memory stores computer program instructions, which are executed by the processor to implement the steps of the above-described method for synthesizing elevator monitoring image data based on surveillance video.

[0037] In summary, the present invention has the following beneficial effects:

[0038] 1. This invention establishes an automated data synthesis process by integrating Blender 3D modeling and AIGC generation technologies. It can generate elevator monitoring images with high fidelity and diversity in batches, providing a sufficient and reliable source of training data for deep learning models, and fundamentally supporting the optimization and application of intelligent monitoring models.

[0039] 2. This invention introduces an innovative technology chain of "Blender prior constraints + Flux-Redux model precise control + iterative editing generation", which successfully solves the problems of easy breakage of thin leashes, incongruity of multi-subject (human-pet-leash) interaction or structural defects in existing technologies; the generated images can well conform to the physical and geometric characteristics of the real world at both the microscopic and macroscopic levels, and can be used to train high-precision recognition models.

[0040] 3. This invention transforms prior knowledge such as elevator structural parameters and camera parameters into quantifiable generation constraints (layout, text, reference vectors), making the entire generation process highly controllable and interpretable. This structured generation paradigm not only ensures data consistency but also enables customized production for specific needs (such as different elevator models or different monitoring angles). Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a system structure block diagram of the present invention;

[0044] Figure 3 This is a schematic diagram of the segmentation model in an embodiment of the present invention. Detailed Implementation

[0045] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0046] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0047] Example 1

[0048] like Figure 1 As shown in the figure, this embodiment presents a specific implementation step example of an elevator monitoring image data synthesis method based on surveillance video.

[0049] To construct a physically accurate elevator simulation scene based on Blender and obtain reference images, the physical rules of the scene must first be set in accordance with the GB / T 7025.1-2023 elevator standard, including the car dimensions of 1.6m (width) × 1.4m (depth) × 2.5m (height) and a double-door structure. Next, the main body of the scene is modeled. The human model is constructed using the Human Generator 3D plugin (a detailed skeleton with a height of 1.75m and 206 bones; the skeleton binding is verified in pose mode after exporting to Blender in FBX format). The pet dog is built using basic geometric shapes such as spheres and cubes, with details optimized using a subdivision modifier and short fur material applied. Both ends are bound to the human wrist joint bones (parent-child relationship) and the dog collar (fixed constraint), respectively. The three are then integrated into the car with the human body centered, the pet dog on the left, and a spacing ≥0.1m. Finally, a virtual camera is configured with a 35mm focal length and a 1024×1024 resolution.

[0050] Individual reference images of various subjects acquired through a virtual camera are considered foreground reference images; empty virtual elevator images without foreground targets are considered background reference images.

[0051] The process involves generating a composite image of an empty elevator, using the open-source image generation model Flux-canny as the core, and combining it with the visual language model Flux-redux to achieve cross-modal semantic alignment.

[0052] The input layer incorporates a reference image (providing a visual baseline) and a black-and-white line drawing (strengthening structural constraints) as control information, which, combined with contextualized text, are injected into the sampler (denoising network). Through iterative denoising optimization, high-fidelity composite images of vacuum elevators in multiple scenes are finally generated, balancing generation diversity with structural accuracy.

[0053] First, a real image that meets the requirements of the scenario needs to be identified as a reference benchmark (such as a real-life image of an empty elevator with a specific style, providing visual anchors such as materials and structure). This image is then simultaneously input into the Redux model and the visual language model. The two work together for feature processing: the visual language model is responsible for semantic feature mapping, while the Redux model focuses on extracting visual features from the image. This model has special parameter constraints—the input image must be set to a fixed aspect ratio to avoid the loss of global features (such as the overall proportions and spatial layout of the elevator car) due to aspect ratio stretching or cropping. Ultimately, this achieves precise alignment between the features of the reference image and the dimensions of the denoising network, laying a stable feature foundation for the subsequent generation process.

[0054] Then, the strength of the transformed semantic features is set. This feature can transform the reference image into multiple tokens. Here, the weight is set to 1.0. This parameter configuration ensures that the semantic features of the reference image play a maximum role in the denoising model, providing key parameter support for preserving the core semantics of the reference image in the subsequent generation process.

[0055] The text input process requires obtaining an accurate description first: input the blank elevator reference image into a large language model (such as ChatGPT) to generate a detailed and reasonable natural language description. Then, through a dual CLIP model (such as the t5-xxl model and the clip_l model), the natural language description is converted into a semantic token. This token serves as the core denoising guidance signal and is finally input into the denoising network to provide linguistic semantic constraints for image generation.

[0056] In the line art image input stage, the line art image is processed by the ControlNet network, and the VAE encoder is called simultaneously to encode its features. The core is to achieve accurate alignment between the spatial structural features of the line art (such as elevator outline and layout proportion) and the aforementioned semantic features, so as to ensure that structural constraints and semantic information are coordinated and unified. Finally, it is used as the overall guiding condition input to the denoising network to enhance the structural accuracy of the generated image.

[0057] The basic parameters of the sampler are set as follows: the number of sampling steps is set to 30 steps to ensure the completeness and fineness of the denoising process through sufficient iterations and avoid loss of details; the CFG intensity is configured to 1.0 to balance the guiding condition constraints and generation flexibility; the sampler function is selected as euler to improve generation efficiency by utilizing its efficient iteration characteristics; the caller function is specified as beta; the denoising intensity is 1.0 to ensure that the noise scheduling matches the denoising network optimization logic.

[0058] Within the K-sampler, based on multimodal guidance conditions such as line drawings and semantic tokens, the feature distribution is gradually optimized through 30 iterations of denoising, effectively eliminating noise interference and ultimately obtaining accurate feature representations in the latent space dimension. Subsequently, the VAE decoder is invoked to initiate a dimensionality transformation process from latent space to normal image space, mapping abstract latent space features into concrete pixel information, and finally outputting a new blank elevator image that conforms to the scene setting and possesses both structural integrity and visual realism.

[0059] Using a pre-generated blank elevator image as the base background, the generation areas for the person and pet dog are first precisely defined through masking control to prevent target elements from overflowing the elevator space. Simultaneously, the ControlNet network is enabled to constrain structural details such as the person's pose and the pet's body proportions to ensure adaptation to the elevator scene dimension. Subsequently, the Redux model is superimposed to strengthen semantic association and deepen the integration of the person, pet, and elevator environment. The CLIP model is used to implement text embedding and provide accurate semantic guidance. Through multi-stage successive generation and optimization, the final output is a complete and realistic image of a person holding a pet dog standing inside the elevator.

[0060] In one possible implementation, using a pre-generated blank elevator image as a fixed background, the generation area is first precisely controlled by manually hand-drawing an image mask: only a reasonable space inside the elevator suitable for a person to stand is defined as the mask area, thus avoiding problems such as positional deviation or exceeding the elevator boundary when the target person is generated, and providing basic constraints for subsequent person positioning.

[0061] In the ControlNet layout detail control stage, a skeleton-driven character model is first built in the Blender virtual scene (ensuring that the character's joint posture and body proportions conform to the real human body structure). After adjusting the virtual camera perspective to match the spatial dimension of the elevator background, an image containing only the character is exported. Then, the Canny edge extraction algorithm is used to process the exported image, stripping away color and texture information, retaining the core structural features such as the character's outline and limb posture, and generating the structured line drawing required by ControlNet, providing precise layout guidance for the generation of character details.

[0062] Meanwhile, the character images exported from Blender are used as reference images for the Redux model. After inputting them into the model, the visual semantic features of the character (such as body shape and posture style) are enhanced to improve the semantic consistency between the generated character and the reference model. Concise text descriptions (such as "a man wearing a dark gray hooded sweatshirt, black slim-fit jeans and white chunky sneakers") are added to supplement the clothing detail constraints through text embedding.

[0063] Finally, the masked area, ControlNet line art, Redux semantic features and text information are collaboratively input into the denoising network, guiding it to generate a foreground character whose pose and structure highly match the Blender virtual character within the designated masked space of the blank elevator, thus achieving a natural blending of the character and the elevator background.

[0064] Using the generated "person + blank elevator" image as a background, the same logic is followed to generate a pet dog: First, a custom mask for the pet dog is hand-drawn to define a reasonable space next to the person in the elevator, ensuring that the pet's position matches the person and elevator scene and avoiding spatial conflicts; then, a skeleton-driven pet dog model is built in the Blender virtual scene under ControlNet control, and after adjusting the posture, a single pet image is exported, and a structured line drawing is obtained through an edge extraction algorithm; this Blender pet image is input into the Redux model to enhance semantics, and paired with text descriptions such as "brown short-haired pet dog, standing docilely," to guide the denoising network to generate a pet dog within the specified mask area that highly matches the Blender model and coordinates with the person.

[0065] Using the generated "person-dog-elevator" image as a base, the natural connection path of the leash is first determined based on the spatial logic of the person's hand and the dog's neck. Then, the leash generation area is locked by local masking, and the leash shape is constrained by ControlNet to ensure visual coherence with the person's movements, the dog's posture, and the elevator environment. Finally, a complete multi-target composite image containing the person, the dog, and the leash is output.

[0066] Example 2

[0067] To address the issue of low segmentation accuracy for slender objects, the elevator scene-specific segmentation model accompanying this invention adopts an end-to-end "encoder-decoder" architecture, adapting to the segmentation requirements of real-world elevator scenarios. Figure 3 As can be seen, its specific structure is as follows:

[0068] Encoder: Built on ResNet50, containing 5 feature extraction layers:

[0069] Stage 1 (Conv1+BN+ReLU+MaxPool) downsamples the input 640×640×3 image to 160×160×64 and extracts shallow edge features;

[0070] Stages 2 through 5 (containing 3, 4, 6, and 3 residual blocks respectively) are progressively downsampled to 10×10×2048, simultaneously increasing the number of feature channels and semantic information to capture deep features of people, pet dogs, and leashes.

[0071] Attention module: Embedded at the Stage 5 output of the encoder, including channel attention (SE-Layer) and spatial attention (CBAM spatial branch).

[0072] Channel attention: Channel weights are generated through “global average pooling fully connected dimensionality reduction — ReLU — fully connected dimensionality increase — Sigmoid” to enhance key channels such as traction rope texture and character edges;

[0073] Spatial attention: Spatial weights are generated through "channel average pooling - 3×3 convolution - Sigmoid" to focus on target areas inside the elevator (such as people's hands and pets' necks) and suppress background interference such as elevator walls and buttons;

[0074] Decoder: Employs a "step-by-step upsampling + skip connection" structure to restore pixel-level segmentation accuracy;

[0075] Output layer: The number of channels is adjusted by 1×1 convolution, and a sigmoid activation function is used to output pixel-level probability masks of people, pet dogs, and leashes.

[0076] In addition, to address the limitations of traditional segmentation models in terms of accuracy caused by the low pixel percentage (only 2%~5%) of traction ropes and class imbalance in elevator scenarios, a combined loss function of hard-to-separate sample focusing loss and small target segmentation intersection-union ratio loss is used to set the loss function of the deep learning network.

[0077] The difficult-to-separate sample focusing loss (focusing on difficult-to-separate target samples and reducing the weight of easily separable background samples) is specifically as follows: .

[0078] in The category balance coefficients (humans = 0.3, pet dogs = 0.3, leash = 0.4) are used to compensate for the low proportion of leash samples. This represents the model's predicted probability for the target category. Samples close to 0 are easily distinguishable, and samples close to 1 are difficult to distinguish. Amplify the loss weights for difficult-to-distinguish samples; (A commonly used focusing coefficient is used to balance the loss contribution of difficult-to-separate and easy-to-separate samples).

[0079] The small target segmentation intersection-union ratio loss (optimizing the segmentation accuracy of small targets such as traction ropes, and measuring the segmentation matching degree by the intersection-union ratio) is as follows: .

[0080] in The true label of the i-th pixel (1 = target pixel, 0 = background pixel); is the model prediction value for the i-th pixel; N is the total number of pixels in a single image; This is the minimum value to avoid calculation errors caused by a denominator of 0.

[0081] The combined loss function (weighted fusion, balancing the optimization of hard-to-separate samples with the need for small target segmentation) is as follows: .

[0082] In this embodiment, the weight setting is based on the fact that "small target segmentation of the traction rope" is a core technical challenge in the elevator scenario. The intersection and union loss of small target segmentation is more direct in optimizing its accuracy, so it is given a higher weight of 0.6. The hard-to-separate sample focusing loss helps to solve the class imbalance problem of "high background ratio and low target ratio", and its weight is set to 0.4.

[0083] Example 3

[0084] like Figure 2 As shown, this embodiment proposes a specific implementation example of an elevator monitoring image data synthesis system based on surveillance video. The system includes:

[0085] The data acquisition and modeling module is used to acquire the structural parameters of the target elevator and the real empty elevator images captured by the monitoring camera, and to construct a virtual empty elevator image of the target elevator based on the structural parameters; it is also used to set a virtual camera corresponding to the virtual empty elevator image based on the camera parameters of the monitoring camera.

[0086] The layout constraint generation module is used to extract image edges from the virtual elevator empty image to obtain a line drawing image, and then use the line drawing image as a layout constraint.

[0087] The text constraint generation module is used to extract features from the scene description text of the virtual elevator empty image based on the text encoder, obtain text vectors, and use the text vectors as text constraints.

[0088] The reference vector generation module is used to extract image features from real-world empty elevator images based on an image visual language model, and use them as corresponding reference vectors.

[0089] The image generation module has a built-in denoising network model, which receives layout constraints, text constraints and reference vectors, and generates different foreground-free synthetic elevator empty images in batches.

[0090] The dataset construction module is used to acquire individual reference images corresponding to various subjects in the reference image based on the virtual camera, and synthesize elevator monitoring images and construct elevator monitoring image datasets using a foreground-free synthesized elevator empty image as the base background.

[0091] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

Claims

1. An elevator monitoring image data synthesis method based on monitoring video, characterized by, The method comprises: S01, obtaining the structural parameters of the target elevator and the real elevator empty image captured by the monitoring camera, constructing a virtual elevator empty image of the target elevator according to the structural parameters, and setting a virtual camera corresponding to the virtual elevator empty image according to the camera parameters of the monitoring camera; S02, image edge extraction is performed on the virtual elevator empty image to obtain a line drawing image, and the line drawing image is taken as a layout constraint; S03, feature extraction is performed on the scene description text of the virtual elevator empty image based on a text encoder to obtain a text vector, and the text vector is taken as a text constraint; S04, image features in the real elevator empty image are extracted based on an image visual language model as a corresponding reference vector; S05, the layout constraint, the text constraint and the reference vector are input into a denoising network model to batch generate different foreground-free synthetic elevator empty images; S06, a single reference image corresponding to each type of subject in the reference image is obtained based on the virtual camera, and an elevator monitoring image dataset is generated with the foreground-free synthetic elevator empty image as the background.

2. The method of claim 1, wherein the method is based on monitoring video. The denoising network model is constructed based on the open source image generation models Flux-canny model and Flux-fill model; the Flux-canny model is a background core generation model, and the Flux-fill model is a foreground core generation model.

3. The method of claim 1, wherein the method comprises: In S06, before generating the elevator monitoring image dataset, the following sub-steps are further included: The mask area between each type of subject is determined based on a mask control method; Skeleton-driven subject models of each type are built in a virtual elevator simulation scene, the perspective of the virtual camera is matched with the spatial dimension of the virtual elevator simulation scene, and a single image corresponding to each type of subject is exported; A Canny edge extraction algorithm is used to process the single image to output corresponding contour structure features and generate corresponding structured line drawings; A virtual foreground image corresponding to a subject is input as a Redux model reference, and a preset detail content is encoded in the Redux model to output a detail constraint; The mask area and the structured line drawing are input into a ControlNet network, and the detail constraint is input into a Flux-fill model to output a complete elevator monitoring image.

4. The method of claim 1, wherein the method is based on monitoring video. The virtual elevator empty image is constructed based on Blender software combined with elevator standard setting scene physical rules.

5. The method of claim 1, wherein the method is based on monitoring video. The camera parameters include the field of view angle, distortion parameters and image resolution of the real camera.

6. The method of claim 1, wherein the method is based on monitoring video. The denoising network model encodes the line drawing image through a ControlNet network and a VAE encoder to generate a spatial structure feature constraint.

7. The method of claim 1, wherein the method is based on monitoring video. The subjects include people, non-human beings and associated objects; the generation order of each type of subject in the generation process of the virtual elevator empty image is: first, generate a person model, then generate a non-human being model, and finally generate an associated object model between the person model and the non-human being model.

8. The method of claim 1, wherein the method is based on monitoring video. In S03, the generation step of the text vector is specifically: Input the real elevator empty image into a deployed large language model to generate a scene description text; The scene description text is semantically converted by a dual CLIP model to correspondingly generate the text vector.

9. An elevator monitoring image data synthesis system based on monitoring video, characterized by, The system comprises: A data acquisition and modeling module is configured to acquire structural parameters of a target elevator and a real elevator hall image captured by a monitoring camera, and construct a virtual elevator hall image of the target elevator according to the structural parameters; The data acquisition and modeling module is further configured to set a virtual camera corresponding to the virtual elevator hall image according to camera parameters of the monitoring camera; A layout constraint generation module is configured to perform image edge extraction on the virtual elevator hall image to obtain a line drawing image, and take the line drawing image as a layout constraint; A text constraint generation module is configured to perform feature extraction on scene description text of the virtual elevator hall image based on a text encoder to obtain a text vector, and take the text vector as a text constraint; A reference vector generation module is configured to extract image features in the real elevator hall image based on an image visual language model as a corresponding reference vector; An image generation module is configured to receive the layout constraint, the text constraint and the reference vector, and batch generate different foreground-free synthetic elevator hall images; A dataset construction module is configured to acquire single reference images corresponding to each type of subject in a reference image based on the virtual camera, and synthesize elevator monitoring images and construct an elevator monitoring image dataset with the foreground-free synthetic elevator hall images as a base background. 10.A computer device, comprising a processor and a memory, and characterized in that, The memory stores computer program instructions, and the computer program instructions are executed by the processor to implement the steps of the elevator monitoring image data synthesis method based on monitoring video according to any one of claims 1-8.

Images