Railway advertisement delivery method, railway advertisement delivery device, and electronic device

CN121235768BActive Publication Date: 2026-09-29ZHEJIANG SUPCON INFORMATION TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511312948.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-09-15
Publication Date
2026-09-29
Estimated Expiration
2045-09-15

AI Technical Summary

Technical Problem

[0003]鉴于现有技术的上述缺点、不足,本发明提供一种轨道广告投放方法、轨道广告投放装置及电子设备,其解决了现有技术中广告推送无法响应乘客即时需求与偏好的技术问题

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121235768B_ABST
    Figure CN121235768B_ABST
Patent Text Reader

Abstract

The present application relates to a kind of track advertisement delivery method, device and electronic equipment.Track advertisement delivery method belongs to the technical field of advertisement push, including the way of being driven to passenger image feature processing by multi-scale fusion and feature confidence gate, obtains passenger image vector;Obtain the advertisement content to be matched, and the way of being fused to multi-modal label regulation by graphic-textual dual-tower embedding, contrast learning pre-training, advertisement semantic analysis is carried out to the advertisement content to be matched, obtains advertisement semantic vector;Vector matching is carried out to passenger image vector and advertisement semantic vector by factor perception attention matching model, obtains the matching degree score of each advertisement content to be matched respectively to target passenger image;Based on the matching degree score of each advertisement content to be matched respectively to target passenger image, determine the delivery advertisement to passenger in target area, it is favorable to timely push the advertisement meeting the demand and preference of passenger to passenger.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of advertising delivery technology, and in particular to a method, device and electronic equipment for placing railway advertisements. Background Technology

[0002] Subway advertising is a form of offline advertising media that utilizes the space within subway stations and train carriages. Due to its unique enclosed nature, high passenger flow, and mandatory viewing, it has become a crucial component of modern urban marketing. While subway advertising possesses significant dissemination value due to its high passenger density, existing systems generally ignore individual passenger characteristics, pushing advertising content solely based on geographical location and time, failing to respond to passengers' immediate needs and preferences. Summary of the Invention

[0003] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method, device and electronic device for placing rail advertisements, which solves the technical problem that the advertising push in the prior art cannot respond to the real-time needs and preferences of passengers.

[0004] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0005] The first aspect of this invention provides a method for placing advertisements on railway tracks.

[0006] The method for placing railway advertisements proposed in this embodiment of the invention includes:

[0007] A passenger profile of the target area inside the railcar is obtained, and the passenger profile is processed by multi-scale fusion and feature confidence gating to obtain a passenger profile vector.

[0008] The ad content to be matched is obtained, and the ad semantics of the ad content to be matched is analyzed by means of image and text dual-tower embedding, contrastive learning pre-training and multimodal label regulation fusion to obtain the ad semantic vector;

[0009] The passenger profile vector and the advertisement semantic vector are matched using a factor-aware attention matching model to obtain the matching degree score of each advertisement content to be matched to the target passenger profile.

[0010] Based on the matching score of each advertisement content to be matched against the target passenger profile, the advertisements to be delivered to passengers in the target area are determined.

[0011] In some instances, the process of performing portrait feature processing on the passenger profile using a multi-scale fusion and feature confidence gating-driven approach to output a passenger profile vector includes:

[0012] Keyframes containing significant motion or rich facial expressions are extracted from the video stream of the passenger profile based on a time-series filtering method.

[0013] The keyframes are subjected to a multi-task cascaded convolutional neural network algorithm for face detection and key point alignment to obtain an aligned face image.

[0014] The aligned face image is input into a multi-scale, multi-task feature extraction network, which outputs a multi-scale feature map.

[0015] The multi-scale feature map is input into the coarse-grained branch and the fine-grained branch to extract high-level attribute features and detailed attribute features, respectively;

[0016] The high-level attribute features and the detailed attribute features are fused through the attention pooling mechanism to output a unified character profile vector representation.

[0017] The image quality confidence score of the keyframe is determined, and the user profile vector representation and the environmental context feature compensation vector are weighted and fused based on the image quality confidence score to generate the final user profile vector representation; wherein, the image quality confidence score is used to characterize the usability of the image quality of the keyframe.

[0018] In some instances, the method of performing advertising semantic analysis on the content to be matched by fusing image-text dual-tower embedding, contrastive learning pre-training, and multimodal label modulation to obtain an advertising semantic vector includes:

[0019] By comparing the learning loss function, the text encoder and image encoder are jointly pre-trained based on paired and unpaired advertising image and text data pairs to narrow the embedding distance of paired image and text samples and widen the embedding distance of unpaired image and text samples.

[0020] The deep semantic representation and advertising intent vector of the advertising text are extracted using a text encoder.

[0021] Extracting multi-scale visual semantic representations of advertising images using an image encoder;

[0022] Using the advertising intent vector as a control signal, the deep semantic representation and the visual semantic representation are adaptively weighted and fused through a tag-guided attention mechanism to generate the advertising semantic vector.

[0023] In some instances, the process of performing vector matching between the passenger profile vector and the advertising semantic vector using a factor-aware attention matching model to obtain a matching score for each advertisement content to the target passenger profile includes:

[0024] The passenger profile vector and the advertising semantic vector are projected onto multiple factor spaces to obtain the first factor representation of the passenger profile vector on each factor and the second factor representation of the advertising semantic vector on each factor.

[0025] Based on the first factor representation of the passenger profile vector on each factor and the second factor representation of the advertising semantic vector on each factor, the matching score of the passenger profile vector and the advertising semantic vector for each factor is determined.

[0026] Based on the passenger profile vector and the advertising semantic vector, the attention weight corresponding to each factor is determined through an attention mechanism.

[0027] Based on the matching score of each factor and the attention weight corresponding to each factor, a weighted fusion is performed to obtain the basic matching score.

[0028] The passenger profile vector and the advertisement semantic vector are concatenated, and high-order feature interaction is performed through a nonlinear cross network to generate a cross-interaction representation;

[0029] Based on the basic matching score and the cross-interaction representation, a click probability estimate representing the matching quality is obtained;

[0030] The estimated click probability is weighted and combined with the diversity reward item of the ad content to be matched to obtain the matching degree score of the ad content to be matched to the target passenger profile; wherein, the diversity reward item of the ad content to be matched is determined based on the proportion of ads belonging to the ad content category in the current recommendation list and the similarity penalty item within the ad set.

[0031] In some instances, the temporal filtering method involves filtering keyframes in the video stream of the passenger profile based on the magnitude of optical flow variation or the facial action recognition score, to obtain keyframes whose optical flow variation magnitude or facial action recognition score is higher than a preset threshold.

[0032] In some instances, determining the image quality confidence score for the keyframe includes:

[0033] Determine the blur level, occlusion level, and relative size of the keyframe;

[0034] Based on the blur level, occlusion level, and relative size of the keyframe, the image quality confidence score of the keyframe is determined.

[0035] In some instances, before generating the final user profile vector representation by weighted fusion of the user profile vector representation and the environmental context feature compensation vector based on the image quality confidence score, the method includes:

[0036] Obtain the timestamp feature vector, the station number feature vector of the passenger, the passenger flow density feature vector, and the carriage environment status feature vector corresponding to the monitoring screen;

[0037] The feature vector is obtained by concatenating the timestamp feature vector, the station number feature vector of the passenger, the passenger flow density feature vector, and the carriage environment status feature vector corresponding to the monitoring screen.

[0038] The feature concatenation vector is normalized to obtain the environmental context feature compensation vector.

[0039] In some instances, before extracting the deep semantic representation of the advertising text and the advertising intent vector through a text encoder, the method includes:

[0040] The text encoder and image encoder are trained using an adversarial discriminator to make the distribution of the embedded representations output by the text encoder and image encoder in the semantic space more consistent.

[0041] A second aspect of the present invention provides a track advertising placement device, comprising:

[0042] The passenger profile acquisition module is used to acquire passenger profiles of the target area inside the railcar, and to perform profile feature processing on the passenger profiles through multi-scale fusion and feature confidence gating to obtain passenger profile vectors.

[0043] The advertising semantic analysis module is used to obtain the advertising content to be matched, and to perform advertising semantic analysis on the advertising content to be matched through image and text dual-tower embedding, contrastive learning pre-training and multimodal label regulation fusion to obtain advertising semantic vector;

[0044] The vector matching module is used to perform vector matching between the passenger profile vector and the advertising semantic vector through a factor-aware attention matching model, so as to obtain the matching degree score of each advertising content to be matched to the target passenger profile.

[0045] The advertising delivery module is used to determine which advertisements to deliver to passengers in the target area based on the matching score of each advertisement content to be matched against the target passenger profile.

[0046] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the steps of the method described in the first aspect above.

[0047] This invention discloses a method for placing railway advertisements, comprising: acquiring passenger profiles of a target area within a railway car, and performing profile feature processing on the passenger profiles using multi-scale fusion and feature confidence gating to obtain passenger profile vectors; acquiring advertisement content to be matched, and performing advertisement semantic analysis on the advertisement content to be matched using image-text dual-tower embedding, contrastive learning pre-training, and multimodal label control fusion to obtain advertisement semantic vectors; performing vector matching between the passenger profile vectors and the advertisement semantic vectors using a factor-aware attention matching model to obtain matching degree scores for each advertisement content to be matched against the target passenger profiles; and determining the advertisements to be placed on passengers within the target area based on the matching degree scores of each advertisement content to be matched against the target passenger profiles. This method, by performing vector matching between the passenger profile vectors and the advertisement semantic vectors to obtain matching degree scores for each advertisement content to be matched against the target passenger profiles, and by determining the advertisements to be placed on passengers within the target area based on the matching degree scores of each advertisement content to be matched against the target passenger profiles, facilitates the determination of passenger needs and preferences based on individual passenger characteristics, and timely pushes advertisements that meet passenger needs and preferences to passengers. Attached Figure Description

[0048] Figure 1 A flowchart of a method for placing advertisements on a railway track is provided in an embodiment of the present invention;

[0049] Figure 2 An advertising placement flowchart provided as an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the structure of the track advertising placement device provided in an embodiment of the present invention. Detailed Implementation

[0051] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] The rail advertising delivery method proposed in this invention addresses the technical problem that existing advertising pushes cannot respond to passengers' immediate needs and preferences. It involves vector matching between passenger profile vectors and advertising semantic vectors to obtain a matching score for each advertisement content to the target passenger profile. Based on these matching scores, the method determines which advertisements to deliver to passengers within the target area. This facilitates understanding passenger needs and preferences based on their individual characteristics and allows for timely delivery of advertisements that meet those needs and preferences.

[0053] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0054] Figure 1 This is a flowchart illustrating a method for placing advertisements on a railway track, as provided in an embodiment of the present invention. Figure 1 As shown, the rail advertising placement method proposed in this embodiment of the invention includes:

[0055] Step 100: Obtain passenger portraits of the target area inside the railcar, and perform portrait feature processing on the passenger portraits through multi-scale fusion and feature confidence gating to obtain passenger portrait vectors;

[0056] Step 110: Obtain the advertising content to be matched, and perform advertising semantic analysis on the advertising content to be matched through image and text dual-tower embedding, contrastive learning pre-training and multimodal label regulation fusion to obtain advertising semantic vector;

[0057] Step 120: Perform vector matching between the passenger profile vector and the advertising semantic vector using the factor-aware attention matching model to obtain the matching degree score of each advertising content to be matched to the target passenger profile.

[0058] Step 130: Based on the matching score of each advertisement content to be matched against the target passenger profile, determine the advertisements to be delivered to passengers in the target area.

[0059] In this exemplary embodiment, a dual-tower structure and contrastive learning loss are adopted to clearly distinguish positive and negative image-text sample pairs during the pre-training stage. Through structural learning, a more boundary-conscious image-text semantic embedding space is formed, which significantly improves the accuracy and discriminativeness of image-text alignment.

[0060] By using advertising intent tags (such as promotions, brands, festivals, etc.) as control vectors to guide the fusion process, the model is endowed with the ability to perceive contextual preferences, enabling image and text to achieve feature aggregation under the same intent semantics, thereby enhancing the expressive directionality and task adaptability of multimodal fusion representation.

[0061] By performing vector matching between the passenger profile vector and the advertising semantic vector, the matching degree score of each advertisement content to be matched to the target passenger profile is obtained; based on the matching degree score of each advertisement content to be matched to the target passenger profile, the advertisements to be delivered to passengers in the target area are determined. This is beneficial for determining the needs and preferences of passengers based on their individual characteristics, and timely pushing advertisements that meet the needs and preferences of passengers.

[0062] In some instances, the process of performing portrait feature processing on the passenger profile using a multi-scale fusion and feature confidence gating-driven approach to output a passenger profile vector includes:

[0063] Keyframes containing significant motion or rich facial expressions are extracted from the video stream of the passenger profile based on a time-series filtering method.

[0064] The keyframes are subjected to a multi-task cascaded convolutional neural network algorithm for face detection and key point alignment to obtain an aligned face image.

[0065] The aligned face image is input into a multi-scale, multi-task feature extraction network, which outputs a multi-scale feature map.

[0066] The multi-scale feature map is input into the coarse-grained branch and the fine-grained branch to extract high-level attribute features and detailed attribute features, respectively;

[0067] The high-level attribute features and the detailed attribute features are fused through the attention pooling mechanism to output a unified human profile vector representation; wherein, the human profile vector representation includes structured labels representing age, gender, behavior, facial expression and the object being carried.

[0068] The image quality confidence score of the keyframe is determined, and the user profile vector representation and the environmental context feature compensation vector are weighted and fused based on the image quality confidence score to generate the final user profile vector representation; wherein, the image quality confidence score is used to characterize the usability of the image quality of the keyframe.

[0069] In this exemplary embodiment, image frames can be acquired in real time by accessing the station monitoring system, and key frames with obvious movement and rich expressions can be extracted using a time-series filtering method (e.g., based on facial action recognition scoring or optical flow changes).

[0070] In this exemplary embodiment, the MTCNN (Multi-task Cascaded Convolutional Neural Networks) algorithm is used to locate and align facial key points. For facial images with a detection confidence score below a threshold, the "portrait confidence gating mechanism" is used for processing.

[0071] In this exemplary embodiment, the system introduces a multi-scale image coding structure that integrates the recognition capabilities of coarse-grained (gender, age group) and fine-grained (emotion, behavior) tags;

[0072] In this exemplary embodiment, a visual language model backbone is used to generate image-text co-embedding, and feature maps of different resolutions are generated through different levels of the Transformer.

[0073] In this exemplary embodiment, there are coarse-grained branches and fine-grained branches. The main task of the coarse-grained branch is to identify high-level attributes (such as gender and age group), while the main task of the fine-grained branch is to identify detailed attributes (such as emotions, micro-expressions, and specific behaviors).

[0074] The branch results are fused into a unified vector representation using an attention pooling mechanism:

[0075] hpersona=AttnPool([fcoarse,ffine]);

[0076] Here, hpersona is a unified vector representation of a person's profile, AttnPool represents a channel fusion operation based on learnable attention weights, fcoarse is the coarse-grained feature vector, and fffine is the fine-grained feature vector.

[0077] In this exemplary embodiment, the model outputs structured labels such as age, gender, behavior, facial expression, and carried objects, and also outputs a unified vectorized embedding, which is used as a profile feature stream to enter the recommendation matching module.

[0078] In this exemplary embodiment, the temporal filtering method involves filtering key frames in the video stream of the passenger profile based on the amplitude of optical flow change or the facial action recognition score, to obtain key frames whose amplitude of optical flow change or facial action recognition score is higher than a preset threshold.

[0079] In this exemplary embodiment, determining the image quality confidence score of the keyframe includes:

[0080] Determine the blur level, occlusion level, and relative size of the keyframe;

[0081] Based on the blur level, occlusion level, and relative size of the keyframe, the image quality confidence score of the keyframe is determined.

[0082] In this exemplary embodiment, considering the significant fluctuations in the quality of actual subway surveillance images, the system introduces an image quality perception module for scoring the usability of the images:

[0083] λ=σ(w1·blur+w2·occlusion+w3·facearea);

[0084] Where σ refers to the Sigmoid function, which maps a real number to the interval 0–1; λ∈[0,1] represents the image profile confidence; blur∈[0,1] represents the degree of blur; occlusion∈[0,1] represents the degree of occlusion; facearea∈[0,1] represents the relative size; and w1, w2, and w3 are weight coefficients. The final image embedding will participate in the recommendation system in the following manner:

[0085] u=λ·hpersona+(1-λ)·hcontextual;

[0086] Here, `hcontextual` is the environmental context feature compensation vector (such as time, passenger flow density, etc.), used to compensate for the image when the image quality is low, ensuring that "the image vector does not have gaps." Surveillance videos may often have unreliable facial features due to blurring or occlusion. If only `hpersona` (visual image) is relied upon, there will be gaps where "no image is available." Therefore, the context vector `hcontextual` is introduced to provide compensation when the quality gating λ is low. contextual By embedding discrete features (time, station) and quantifying continuous features (passenger density, environmental state), the MLP is mapped to the same dimension. hpersona is a unified person profile vector representation, and u is the final user profile output, i.e., the passenger profile vector output in step 100.

[0087] This application constructs a multi-scale visual language model fusion structure and introduces an image quality perception and confidence gating mechanism, which effectively solves problems such as image blurring, occlusion, and atypical faces in subway scenes. Even in low-quality or complex scenes, it can still output structured, multi-dimensional, and highly reliable user profiles, and achieve accurate modeling of passenger status.

[0088] In some instances, before generating the final user profile vector representation by weighted fusion of the user profile vector representation and the environmental context feature compensation vector based on the image quality confidence score, the method includes:

[0089] Obtain the timestamp feature vector, the station number feature vector of the passenger, the passenger flow density feature vector, and the carriage environment status feature vector corresponding to the monitoring screen;

[0090] The feature vector is obtained by concatenating the timestamp feature vector, the station number feature vector of the passenger, the passenger flow density feature vector, and the carriage environment status feature vector corresponding to the monitoring screen.

[0091] The feature concatenation vector is normalized to obtain the environmental context feature compensation vector.

[0092] In this exemplary embodiment, hcontextual is a feature compensation vector for environmental context (such as time, passenger flow density, etc.); where,

[0093] h contextual =LayerNorm([e time ⊕e station ⊕e crowd ⊕e env W c );

[0094] Where ⊕ represents concatenation, that is, joining multiple vectors end-to-end to form a longer vector, W c LayerNorm is used for linear projection to the same dimension as hpersona, ensuring scale stability.

[0095] e time The timestamp corresponding to the monitored image is discretized and then input into the model as a timestamp feature vector, e. station The station where the passenger is located is assigned a feature vector, e crowd e is the passenger flow density feature vector. env This is a feature vector representing the environmental state of the train carriage, indicating contextual information related to the station hall / carriage environment, which helps compensate for low visual quality in the image. Values ​​range from light intensity (bright / dim / shadow) to camera resolution (HD / SD / blurry).

[0096] In some instances, the method of performing advertising semantic analysis on the content to be matched by fusing image-text dual-tower embedding, contrastive learning pre-training, and multimodal label modulation to obtain an advertising semantic vector includes:

[0097] By comparing the learning loss function, the text encoder and image encoder are jointly pre-trained based on paired and unpaired advertising image and text data pairs to narrow the embedding distance of paired image and text samples and widen the embedding distance of unpaired image and text samples.

[0098] The deep semantic representation and advertising intent vector of the advertising text are extracted using a text encoder.

[0099] Extracting multi-scale visual semantic representations of advertising images using an image encoder;

[0100] Using the advertising intent vector as a control signal, the deep semantic representation and the visual semantic representation are adaptively weighted and fused through a tag-guided attention mechanism to generate the advertising semantic vector.

[0101] In this exemplary embodiment, deep semantic features of advertising text and images are extracted uniformly. By introducing a cross-modal contrastive learning mechanism and a label control structure, a more discriminative and matching multimodal fusion representation is generated as input to the recommendation system. To improve the semantic alignment of advertising text and image content, this module proposes a joint modeling structure of "text-image dual-tower embedding + contrastive learning pre-training + multimodal label control fusion," breaking through the traditional shallow interaction method and constructing an advertising semantic embedding representation with stronger cross-modal expression consistency.

[0102] In this exemplary embodiment, BERT-based...

[0103] (Bidirectional Encoder Representations from Transformers, based on

[0104] The Transformer's bidirectional encoder representation extracts keywords, user intent, and contextual semantics from the ad description; simultaneously, it parses ad intent tags (such as "promotion," "display," and "interaction") as control signals to guide subsequent modal fusion. Among these, h... text =RoBERTa(x text ),e intent =p T E intent ;

[0105] Where, x text For the original text of the advertisement, h text For deep semantic representation of advertising text, p is the probability distribution of intent category, and E is the deep semantic representation of advertising text. intent For the intended embedding matrix, e intent This is the ad intent vector. RoBERT is the ad text encoder.

[0106] In this exemplary embodiment, a Vision Transformer is used to extract multi-head semantic features of the advertising image, including product category, scene elements, brand anchors, etc. Multi-scale attention pooling is employed for aggregation.

[0107] h image =AttnPool([t1,t2,...,t) n ]);

[0108] Among them, t i For image token representation, h image This is for the semantic representation of images.

[0109] In this exemplary embodiment, to improve the semantic alignment capability of text and images, paired advertising text-image pairs can be used as positive samples, and unpaired pairs as negative samples. Through a self-learning contrastive loss function, the model learns that: text and images from the same advertisement should have a short embedding distance, while text and images from different advertisements should have a long embedding distance, establishing "cross-modal matching semantics," allowing the model to know that "this image and this text are indeed from the same advertisement." This is equivalent to semantic-level positive and negative contrastive learning. Introducing a self-learning contrastive loss function aggregates true text-image pairs and widens the gaps between unpaired text and images, forming clear embedding boundaries.

[0110]

[0111] Where τ is the temperature coefficient, τ>0; h text h is the vector representation of the advertising text. image A vector representation of the advertisement image paired with the text; For comparison, the representation of the j-th advertisement image in the batch; cos(h text ,h image The cosine similarity is [-1, 1], with a larger value indicating greater semantic similarity. This represents the sum of similarity between the text and all images within the batch; Essentially, it's cross-entropy loss, aiming to maximize the similarity between the text and the correctly paired image.

[0112] This application achieves deep semantic alignment between advertising images and text content through a dual-tower embedding structure of images and text, cross-modal contrastive learning, and tag-guided fusion mechanism, generating a more unified and distinguishable multimodal advertising representation, which helps improve the expressive consistency and task adaptability of subsequent matching.

[0113] In some instances, the process of performing vector matching between the passenger profile vector and the advertising semantic vector using a factor-aware attention matching model to obtain a matching score for each advertisement content to the target passenger profile includes:

[0114] The passenger profile vector and the advertising semantic vector are projected onto multiple factor spaces to obtain the first factor representation of the passenger profile vector on each factor and the second factor representation of the advertising semantic vector on each factor.

[0115] Based on the first factor representation of the passenger profile vector on each factor and the second factor representation of the advertising semantic vector on each factor, the matching score of the passenger profile vector and the advertising semantic vector for each factor is determined.

[0116] Based on the passenger profile vector and the advertising semantic vector, the attention weight corresponding to each factor is determined through an attention mechanism.

[0117] Based on the matching score of each factor and the attention weight corresponding to each factor, a weighted fusion is performed to obtain the basic matching score.

[0118] The passenger profile vector and the advertisement semantic vector are concatenated, and high-order feature interaction is performed through a nonlinear cross network to generate a cross-interaction representation;

[0119] Based on the basic matching score and the cross-interaction representation, a click probability estimate representing the matching quality is obtained;

[0120] The estimated click probability is weighted and combined with the diversity reward item of the ad content to be matched to obtain the matching degree score of the ad content to be matched to the target passenger profile; wherein, the diversity reward item of the ad content to be matched is determined based on the proportion of ads belonging to the ad content category in the current recommendation list and the similarity penalty item within the ad set.

[0121] In this exemplary embodiment, based on multimodal user profiles and advertising semantic vectors, adjustable personalized advertising recommendations are achieved through factor-aware dynamic matching, ranking learning, and joint optimization of diversity. The factor-aware attention fusion network dynamically learns the matching weights of users and advertisements on various factors (such as age and emotion). A nonlinear cross-interaction structure captures high-order correlations between user features and advertising attributes. End-to-end ranking learning and joint optimization of diversity model match quality and diversity constraints in a unified manner, avoiding phased conflicts.

[0122] In this exemplary embodiment, the user and advertisement representations are projected onto the k-th factor space using a factor-aware attention matching model:

[0123] f u,k =P u,k (u),f a,k =P a,k (a);

[0124] Where: P u,k and P a,k Let f be the factor projection matrix. u,k It is the representation of the user profile on the k-th factor, f a,k This is the representation of the advertisement on the k-th factor. For the k-th factor, calculate the matching score:

[0125] s k =g k (f u,k ,f a,k ) = MLP k ([f u,k ⊕f a,k ]);

[0126] Where k represents the k-th factor, the model calculates a matching score s for each factor dimension separately. k ⊕ indicates vector concatenation, MLP k (·) represents a feedforward neural network (multilayer perceptron) for the k-th factor.

[0127] Factor attention weights are calculated based on overall user and ad information, with the weight α representing the factor attention weight. k :

[0128]

[0129] Here, W is a learnable linear transformation matrix that projects the concatenated vectors into an "attention space". tanh(·) is the activation function, and w k α is the attention parameter vector for the k-th factor, where exp(·) is used for exponentialization to ensure positive scores and amplify differences. k The attention weights for factor k. The base matching score h. match Through attention weight α k The weighted fusion yields:

[0130] h match =∑ k α k ·s k ; where s k For each factor, a matching score is given.

[0131] Non-linear cross-interaction structure: To further capture the higher-order cross-relationships between passenger features and advertising attributes, a CrossNet-style cross layer is introduced:

[0132] Input concatenation: h (0) =[u⊕a];

[0133] For the l-th layer: h (l+1) =h (0) ·(h (l) ) T W (l) +b (l) +h (l) ;

[0134] After L=3 layers of intersection: h cross =h (L) ;

[0135] Among them, h (0) h is the initial input to CrossNet (layer 0). (l) W is the latent vector representation of the l-th layer. (l) b is the weight matrix of the l-th layer. (l) h is the bias vector of the l-th layer.cross This is the final cross-interaction representation.

[0136] In some instances, before extracting the deep semantic representation of the advertising text and the advertising intent vector through a text encoder, the method includes:

[0137] The text encoder and image encoder are trained using an adversarial discriminator to make the distribution of the embedded representations output by the text encoder and image encoder in the semantic space more consistent.

[0138] In this exemplary embodiment, to improve the semantic consistency between images and text, an adversarial discriminator D is introduced to attempt to distinguish whether the input embedding is "text" or "image". This makes the outputs of the image encoder and text encoder tend to be modally indistinguishable in unsupervised adversarial scenarios, making the distribution forms of the two modalities more consistent (e.g., mean / variance, latent space geometry), and reducing "modal discrepancies". This is equivalent to alignment at the distribution level.

[0139]

[0140] Where x text For the sample advertising text, x image E is an advertising image sample. T E I This is a text and image encoder that maps text and images to semantic embedding spaces, respectively. D is the discriminator, representing the confidence that the vector comes from the text, logD(E T (x text )) indicates that the discriminator expects the output of the real text embedding to be close to 1, log(1-D(E) I (x image ))) indicates that the discriminator expects the image embedding output to be close to 0.

[0141] The system incorporates advertising intent tags (such as promotion / brand / festival) as control vectors in multimodal feature fusion to adjust the attention weights of different modalities:

[0142] a = Attn([h text ,h image ]|e intent ); where a is the final semantic vector of the advertisement, which includes the semantics of both text and image, and is also controlled by the "intent tag", and is directly used as the input of the recommendation matching module.

[0143] In this exemplary embodiment, the fused semantic vector will serve as the representation input for the advertisement in the recommendation matching module. Wherein, e intentThe ad intent vector (e) acts as a "control signal" during fusion, adjusting the attention weights of text and images. For example, in promotional scenarios, more attention is paid to the discount information in the text, while in branding scenarios, more attention is paid to the logo and visual anchors in the image. Attn guides attention through tags; during the fusion phase, the ad intent vector e... intent As a control signal, to adjust the text representation h text and image representation h image The weight distribution.

[0144] In this exemplary embodiment, after obtaining the matching score of each ad content to the target passenger profile, the ad content with the highest matching score can be determined as the ad to be delivered to the target passenger profile, and the ad can be delivered. When there are multiple target passengers in the target area, the overall matching score of a single ad content for all passengers can be sorted, and the ad content with the highest matching score can be determined as the ad to be delivered to the target passenger profile in the target area, and the ad can be delivered. For example, if there are 3 target passengers in the target area, then 3 matching scores are determined for each ad content for the 3 passengers, and then the 3 matching scores are added together as the overall matching score of the ad content for all target passengers.

[0145] In this exemplary embodiment, in order to obtain the final ad ranking score, both match quality and diversity may be considered.

[0146] Matching quality term: Q = σ(MLP) q ([h match ⊕h cross ]));

[0147] Where σ(·) is the Sigmoid activation, and the output is the estimated click probability.

[0148] Diversity item:

[0149] Where, p c This represents the percentage of ads belonging to category c in the current recommendation list. For ad collection The internal similarity penalty term (to avoid excessive duplicate ads) is γ, which is the coefficient for balancing entropy and penalty.

[0150] The estimated click probability is weighted and combined with the diversity bonus of the ad content to be matched to obtain the matching score S of the ad content to the target passenger profile, including:

[0151] S = Q + βD; where β is a learnable parameter that can be dynamically adjusted; Q is the estimated click probability of the matching quality item; and D is the diversity bonus item for the ad content to be matched.

[0152] The proposed method for subway advertising headers aims to break through the traditional "static content + coarse-grained matching" model of subway advertising and establish an intelligent advertising recommendation architecture based on dynamic user profiles of passengers and integrating multimodal semantic understanding and environmental adaptation mechanisms.

[0153] This application addresses the challenges posed by inconsistent video input quality and complex passenger behavior in subway scenarios, proposing the following:

[0154] A multi-scale, multi-task embedding network structure is proposed, which integrates coarse-grained (gender, age) and fine-grained (emotion, behavior) features through a visual language model (VLM) to form a unified passenger profile vector.

[0155] This application calculates image confidence based on indicators such as image blur, occlusion, and face area, and performs weighted gating fusion of image vector and environment vector based on the confidence.

[0156] This application is beneficial for improving the recognition robustness of images with occlusion, blurriness, and atypical emotions. It is an innovative method in the direction of deep image feature fusion and quality gating weighting and has independent protection.

[0157] Traditional advertising content encoding often suffers from semantic fragmentation between images and text. To address this, this application employs a contrastive learning embedding model based on a dual-tower structure of images and text, combined with InfoNCE loss to enhance image-text alignment capabilities; it introduces advertising intent tags (such as promotion / display / festival) as guiding vectors to adjust the image-text fusion attention mechanism, thereby constructing a fusion vector with controllable semantic preferences; and it utilizes a modal adversarial alignment mechanism to spatially unify image-text representations.

[0158] The above scheme improves the consistency of advertising content vector expression and task controllability, and is an innovation in cross-modal semantic enhancement and control fusion structure. It can be protected as an independent claim point for the input structure of the recommender system.

[0159] This application constructs a passenger-advertisement matching model with adjustable factor granularity, the core of which includes:

[0160] Factor-aware attention structure dynamically adjusts the weight of different features (such as age and emotion) and advertising content in the matching process;

[0161] A non-linear cross-interaction structure captures the high-order coupling relationship between passenger profiles and advertising attributes;

[0162] The matching quality score and ad diversity constraints are jointly optimized end-to-end.

[0163] This application overcomes the shortcomings of traditional recommendation systems that decouple matching and diversity in their modeling, and achieves unified modeling of personalized recommendations and advertising exposure control. It has high engineering practicality and theoretical innovation, and can be protected as one of the core rights.

[0164] This application constructs a complete closed-loop process of passenger profiling, advertising content, recommendation matching, environmental control, and feedback optimization, unifies data flow and model interface standards, and significantly improves the system's scalability and engineering implementation capabilities.

[0165] This application addresses the problems of current rail transit advertising systems, such as "static content, lack of user perception, and slow response to environmental changes." It proposes an intelligent advertising system that integrates multimodal perception, semantic modeling, and dynamic control, significantly improving the accuracy of ad matching, the flexibility of advertising strategies, and the overall adaptability of the system. Specifically, this is reflected in the following aspects:

[0166] This application improves the accuracy and robustness of passenger profiles. For example, it constructs a multi-scale visual language model fusion structure and introduces image quality perception and confidence gating mechanisms, which effectively solves problems such as image blurring, occlusion, and atypical faces in subway scenarios. Even in low-quality or complex scenarios, it can still output structured, multi-dimensional, and highly reliable user profiles, achieving accurate modeling of passenger status.

[0167] Enhance the semantic expressiveness of advertising content. For example, this application achieves deep semantic alignment between advertising images and text content through a dual-tower embedding structure of images and text, cross-modal contrastive learning, and a tag-guided fusion mechanism, generating a more unified and distinguishable multimodal advertising representation, which helps improve the expressive consistency and task adaptability of subsequent matching.

[0168] This enables factor-aware and personalized control of ad matching. For example, through factor-aware attention mechanisms and higher-order interaction structures, ad content can be dynamically matched based on different passenger profile factors (such as age, gender, and emotion), breaking through the traditional "static rule recommendation" or "single feature matching" model and achieving a more personalized and scenario-adaptive intelligent matching process.

[0169] This application enhances both the effectiveness and diversity of ad ranking. For example, it employs a joint ranking optimization strategy that integrates ad click probability prediction with diversity control factors in its modeling. This improves recommendation accuracy while avoiding content monotony and homogenization, thereby increasing the attractiveness and acceptance of ad content from the passenger's perspective.

[0170] This paper introduces a modal adversarial mechanism to improve the consistency of the representation space. For example, existing image-text coding models typically learn their own semantic representations, resulting in a significant semantic gap between modalities. This application uses an adversarial discriminator to make the outputs of the image encoder and text encoder nearly modally indistinguishable in the embedding space, fundamentally compressing the semantic distribution gap between images and text and solving the "modal misalignment" problem.

[0171] In addition to a large visual language model, this application can use traditional convolutional neural networks or lightweight image classification models based on transfer learning to extract passenger appearance features (such as age and gender). Its advantages lie in its lightweight nature and fast inference speed, making it suitable for deployment on resource-constrained edge devices. However, in complex tasks such as emotion recognition and behavior perception, its recognition accuracy and stability may be lower than large models, and its adaptability to scene changes is weaker.

[0172] In terms of feature extraction from advertising text, lightweight natural language processing models can be used as alternatives.

[0173] BERT-like pre-trained models have the advantages of fast inference speed and low computational resource consumption, making them suitable for low-latency scenarios. However, their performance is limited in tasks such as understanding long texts and cross-domain semantic modeling.

[0174] Besides deep learning matching models, classic recommendation methods such as collaborative filtering, matrix factorization, and content recommendation can be used to model the association between user profiles and advertisements. Among them, collaborative filtering is suitable for scenarios with sufficient user interaction history, and is simple and efficient but suffers from the cold start problem; matrix factorization models are suitable for stable interest modeling, but their responsiveness to real-time context changes is insufficient.

[0175] In terms of dynamic weighting, traditional methods such as rule-based context scoring and linear regression models can also be used to score and integrate environmental factors. This method is simple to implement and highly interpretable, but it lacks self-learning ability and is difficult to cope with complex or nonlinear environmental changes.

[0176] This application can employ a diversity ranking method based on category deduplication or complex networks, which is suitable for recommendation tasks that require a certain degree of content diversity, but may be slightly inferior in terms of ranking accuracy or user experience.

[0177] Considering the resource constraints and accuracy requirements of different deployment environments, the above methods are of practical significance in certain scenarios and can serve as a flexible supplementary solution for engineering implementation.

[0178] MTCNN (Multi-task Cascaded Convolutional Networks) is a classic face detection and keypoint localization algorithm. It employs a three-stage cascaded structure to sequentially generate, refine, and localize five keypoints (eyes, nose tip, and corners of the mouth), offering advantages such as high detection accuracy, high speed, and strong adaptability. In this system, MTCNN is used to detect passenger faces in real-time from surveillance video and perform alignment and normalization processing, providing accurate face region input for subsequent feature recognition and user profile modeling.

[0179] Cross-ModalAttention is a deep learning method for semantic alignment of multimodal data, capable of modeling the interdependencies between features of different modalities (such as text and images). By introducing an attention mechanism, the system can dynamically allocate attention weights between images and text, achieving precise alignment and complementary fusion of semantic information. In this system, this mechanism is used to deeply fuse advertising text with image content, improving the completeness and accuracy of advertising semantic understanding, thereby enhancing the matching effect between passenger profiles and advertising content.

[0180] Multi-factor fusion mechanism refers to a strategy in recommender systems that comprehensively considers multiple dimensions of feature factors to rank and filter candidate options. In this system, this mechanism combines user profile features such as passenger age, gender, emotion, and behavioral status with information such as the target audience, content type, and timeliness of advertisements. It uses weighted modeling and nonlinear fusion methods (such as multilayer perceptrons) for matching and scoring to ensure that the recommendation results are both relevant and consistent with the context, achieving more accurate and personalized advertising.

[0181] The Bandit algorithm is a strategy optimization method used to dynamically balance exploration and exploitation, particularly suitable for online recommendation scenarios. This system employs the Thompson Sampling strategy, updating parameters in real-time based on user feedback (such as click-through rate and dwell time) after ad delivery, enabling the recommendation model to self-correct and evolve. This mechanism, combined with user feedback data, constructs a behavior-driven feedback loop, significantly improving the system's adaptability and long-term benefits.

[0182] Figure 2 This is a flowchart illustrating an advertising placement process provided in an embodiment of the present invention. Figure 2 As shown, the advertising placement process includes:

[0183] Step 200: Perform passenger profile analysis through the passenger user profile analysis module, including video acquisition and frame processing, face detection and alignment, multi-scale feature extraction, and passenger profile output.

[0184] Step 210: Extract text features, extract image features, learn image-text comparison, and output advertising vectors through the advertising multi-dimensional feature extraction module;

[0185] Step 220: By combining the multi-dimensional feature extraction module of the advertisement with factor perception attention, the passenger profile vector and the advertisement semantic vector are matched through the cross-interaction network, and the delivery decision is output through joint ranking optimization.

[0186] This invention provides a device for placing advertisements on a railway track. Figure 3 This is a schematic diagram of the structure of a track advertising placement device provided in an embodiment of the present invention. Figure 3 As shown, it includes:

[0187] The passenger profile acquisition module 30 is used to acquire passenger profiles of the target area inside the railcar, and to perform profile feature processing on the passenger profiles through multi-scale fusion and feature confidence gating to obtain passenger profile vectors.

[0188] The advertising semantic analysis module 31 is used to obtain the advertising content to be matched, and to perform advertising semantic analysis on the advertising content to be matched through the fusion of image and text dual-tower embedding, contrastive learning pre-training and multimodal label regulation to obtain advertising semantic vector;

[0189] The vector matching module 32 is used to perform vector matching between the passenger profile vector and the advertising semantic vector through the factor-aware attention matching model to obtain the matching degree score of each advertising content to be matched to the target passenger profile.

[0190] The advertising delivery module 33 is used to determine the advertising to be delivered to passengers in the target area based on the matching score of each advertising content to be matched against the target passenger profile.

[0191] In this exemplary embodiment, a dual-tower structure and contrastive learning loss are adopted to clearly distinguish positive and negative image-text sample pairs during the pre-training stage. Through structural learning, a more boundary-conscious image-text semantic embedding space is formed, which significantly improves the accuracy and discriminativeness of image-text alignment.

[0192] By using advertising intent tags (such as promotions, brands, festivals, etc.) as control vectors to guide the fusion process, the model is endowed with the ability to perceive contextual preferences, enabling image and text to achieve feature aggregation under the same intent semantics, thereby enhancing the expressive directionality and task adaptability of multimodal fusion representation.

[0193] By performing vector matching between the passenger profile vector and the advertising semantic vector, the matching degree score of each advertisement content to be matched to the target passenger profile is obtained; based on the matching degree score of each advertisement content to be matched to the target passenger profile, the advertisements to be delivered to passengers in the target area are determined. This is beneficial for determining the needs and preferences of passengers based on their individual characteristics, and timely pushing advertisements that meet the needs and preferences of passengers.

[0194] This invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the track advertising placement method described in the above embodiments.

[0195] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0196] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0197] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0198] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0199] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for placing advertisements on railway tracks, characterized in that, include: A passenger portrait of a target area within a railcar is acquired, and its features are processed using a multi-scale fusion and feature confidence gating approach to obtain a passenger portrait vector. The process of processing the passenger portrait features using multi-scale fusion and feature confidence gating to obtain the passenger portrait vector includes: extracting keyframes containing significant motion or rich facial expressions from the video stream of the passenger portrait using a temporal filtering method; performing face detection and keypoint alignment on the keyframes using a multi-task cascaded convolutional neural network algorithm to obtain aligned face images; and inputting the aligned face images into a... A multi-scale, multi-task feature extraction network outputs multi-scale feature maps. These feature maps are then input to coarse-grained and fine-grained branches to extract high-level and detailed attribute features, respectively. An attention pooling mechanism is used to fuse the high-level and detailed attribute features, outputting a unified user profile vector representation. An image quality confidence score is determined for each keyframe, and based on this score, the user profile vector representation and the environmental context feature compensation vector are weighted and fused to generate the final user profile vector representation. The image quality confidence score characterizes the usability of the keyframe's image quality. The ad content to be matched is obtained, and the ad semantics of the ad content to be matched is analyzed by means of image and text dual-tower embedding, contrastive learning pre-training and multimodal label regulation fusion to obtain the ad semantic vector; The passenger profile vector and the advertisement semantic vector are matched using a factor-aware attention matching model to obtain the matching degree score of each advertisement content to be matched to the target passenger profile. Based on the matching score of each advertisement content to be matched against the target passenger profile, the advertisements to be delivered to passengers in the target area are determined.

2. The method for placing rail advertisements according to claim 1, characterized in that, The method of fusing image-text dual-tower embedding, contrastive learning pre-training, and multimodal label control to perform advertising semantic analysis on the content to be matched yields an advertising semantic vector, including: By contrastive learning loss function, the text encoder and image encoder are jointly pre-trained based on paired and unpaired advertising image and text data pairs to narrow the embedding distance of paired image and text samples and widen the embedding distance of unpaired image and text samples. The deep semantic representation and advertising intent vector of the advertising text are extracted using a text encoder. Extracting multi-scale visual semantic representations of advertising images using an image encoder; Using the advertising intent vector as a control signal, the deep semantic representation and the visual semantic representation are adaptively weighted and fused through a tag-guided attention mechanism to generate the advertising semantic vector.

3. The method for placing rail advertisements according to claim 1, characterized in that, The step of performing vector matching between the passenger profile vector and the advertising semantic vector using a factor-aware attention matching model to obtain a matching score for each advertisement content to the target passenger profile includes: The passenger profile vector and the advertising semantic vector are projected onto multiple factor spaces to obtain the first factor representation of the passenger profile vector on each factor and the second factor representation of the advertising semantic vector on each factor. Based on the first factor representation of the passenger profile vector on each factor and the second factor representation of the advertising semantic vector on each factor, the matching score of the passenger profile vector and the advertising semantic vector for each factor is determined. Based on the passenger profile vector and the advertising semantic vector, the attention weight corresponding to each factor is determined through an attention mechanism. Based on the matching score of each factor and the attention weight corresponding to each factor, a weighted fusion is performed to obtain the basic matching score. The passenger profile vector and the advertisement semantic vector are concatenated, and high-order feature interaction is performed through a nonlinear cross network to generate a cross-interaction representation; Based on the basic matching score and the cross-interaction representation, a click probability estimate representing the matching quality is obtained; The estimated click probability is weighted and combined with the diversity reward item of the ad content to be matched to obtain the matching degree score of the ad content to be matched to the target passenger profile; wherein, the diversity reward item of the ad content to be matched is determined based on the proportion of ads belonging to the ad content category in the current recommendation list and the similarity penalty item within the ad set.

4. The method for placing rail advertisements according to claim 1, characterized in that, The temporal filtering method involves filtering key frames in the video stream of the passenger profile based on the amplitude of optical flow change or the facial action recognition score, to obtain key frames whose amplitude of optical flow change or facial action recognition score is higher than a preset threshold.

5. The method for placing rail advertisements according to claim 1, characterized in that, Determining the image quality confidence score of the keyframe includes: Determine the blur level, occlusion level, and relative size of the keyframe; Based on the blur level, occlusion level, and relative size of the keyframe, the image quality confidence score of the keyframe is determined.

6. The method for placing rail advertisements according to claim 1, characterized in that, Before generating the final user profile vector representation by weighted fusion of the user profile vector representation and the environmental context feature compensation vector based on the image quality confidence score, the method includes: Obtain the timestamp feature vector, the station number feature vector of the passenger, the passenger flow density feature vector, and the carriage environment status feature vector corresponding to the monitoring screen; The feature vector is obtained by concatenating the timestamp feature vector, the station number feature vector of the passenger, the passenger flow density feature vector, and the carriage environment status feature vector corresponding to the monitoring screen. The feature concatenation vector is normalized to obtain the environmental context feature compensation vector.

7. The method for placing rail advertisements according to claim 2, characterized in that, Before extracting the deep semantic representation of the advertising text and the advertising intent vector through the text encoder, the method includes: The text encoder and image encoder are trained using an adversarial discriminator to make the distribution of the embedded representations output by the text encoder and image encoder in the semantic space more consistent.

8. A track advertising placement device, characterized in that, include: The passenger profile acquisition module is used to acquire passenger profiles of a target area within the railcar, and to perform profile feature processing on the passenger profiles using a multi-scale fusion and feature confidence gating approach to obtain a passenger profile vector. The process of performing profile feature processing on the passenger profiles using a multi-scale fusion and feature confidence gating approach to obtain the passenger profile vector includes: extracting keyframes containing significant motion or rich facial expressions from the video stream of the passenger profiles based on a temporal filtering method; performing face detection and keypoint alignment on the keyframes using a multi-task cascaded convolutional neural network algorithm to obtain aligned face images; and then... The image is input into a multi-scale, multi-task feature extraction network, which outputs a multi-scale feature map. The multi-scale feature map is then input into coarse-grained and fine-grained branches to extract high-level attribute features and detailed attribute features, respectively. These high-level attribute features and detailed attribute features are fused using an attention pooling mechanism to output a unified user profile vector representation. An image quality confidence score is determined for each keyframe, and the user profile vector representation is weighted and fused with the environmental context feature compensation vector based on this image quality confidence score to generate the final user profile vector representation. The image quality confidence score characterizes the usability of the image quality in the keyframe. The advertising semantic analysis module is used to obtain the advertising content to be matched, and to perform advertising semantic analysis on the advertising content to be matched through image and text dual-tower embedding, contrastive learning pre-training and multimodal label regulation fusion to obtain advertising semantic vector; The vector matching module is used to perform vector matching between the passenger profile vector and the advertising semantic vector through a factor-aware attention matching model, so as to obtain the matching degree score of each advertising content to be matched to the target passenger profile. The advertising delivery module is used to determine which advertisements to deliver to passengers in the target area based on the matching score of each advertisement content to be matched against the target passenger profile.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Application method of virtual reality using the 3D background screen and direct advertisement

    KR1020090122897A

  • System And Method For Creating Individualized Mobile and Visual Advertisments Using Facial Recognition

    US20160150260A1