A first-person operation data collection and motivation monetization system

Abstract

This invention relates to the field of human-computer interaction data processing technology, specifically a first-person operation data acquisition and incentive monetization system. The system includes: a data acquisition module that acquires a target user's first-person perspective original operation video stream and original touch event sequence; a feature processing module that performs keyframe extraction and eye-tracking region annotation on the video stream to generate a time-series operation video segment with a gaze heatmap; and a touch event sequence that reconstructs gesture trajectories and quantifies pressure features to generate a touch intent encoding sequence; and a recording generation module that aligns the two types of data bidirectionally on the timeline to construct a multimodal operation record package. Based on a pre-trained monetization value prediction model, a regression calculation is performed to derive an incentive score, virtual assets are distributed according to the score, and the record is written to a blockchain incentive ledger. This solution enriches the feature dimensions of operation data, achieves accurate time-series fusion of multimodal data, and completes the entire process of operation data value assessment and incentive distribution with reliable evidence storage.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction data processing technology, and in particular to a first-person operation data acquisition and incentive realization system. Background Technology

[0002] Conventional first-person perspective operation data acquisition modes typically involve capturing only single operation video frames or independently recording touch operation events. Data processing merely employs basic video extraction methods or simply replicates touch motion trajectories, failing to deeply analyze the user's visual behavior and deeper touch characteristics during operation, and lacking a mechanism for integrating and correlating various types of operation data. Under existing data acquisition and processing models, operation videos only retain continuous image information, unable to accurately pinpoint the user's eye-tracking focus area during operation; touch events can only record basic swipe and click trajectories, making it difficult to capture implicit operation features such as pressure levels. The generated operation data is limited in scope and cannot fully reconstruct the logic of actual operation behavior.

[0003] Heterogeneous video image data and touch-coded data inherently suffer from temporal misalignment. Conventional stitching methods only perform simple time node correspondence, failing to achieve bidirectional temporal calibration and adaptation, resulting in weak temporal correlation in the generated operation data packets. Furthermore, the lack of standardized value quantification methods and reliable incentive distribution and documentation mechanisms prevents the quantification and conversion of operation data into tangible value. The incentive distribution process also lacks traceable ledger records, making it difficult to complete a closed loop for market-based incentive monetization based on user-generated first-person operation data. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a first-person operation data acquisition and incentive realization system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a first-person operational data acquisition and incentive realization system, comprising: The data acquisition module is used to acquire the original operation video stream and original touch event sequence of the target user from a first-person perspective; The feature processing module is used to perform keyframe extraction and eye-tracking region annotation on the original operation video stream to generate a time-series operation video segment with a gaze heatmap, and to perform gesture trajectory reconstruction and pressure feature quantization on the original touch event sequence to generate a touch intent encoding sequence. The recording generation module is used to perform bidirectional alignment and mounting of the time sequence operation image fragment and the touch intent encoding sequence on the time axis to generate a multimodal operation recording package; The value assessment module is used to call a pre-trained monetization value prediction model to perform incentive amount regression calculation on the multimodal operation record package and generate the operation data incentive score corresponding to the target user. The incentive execution module is used to perform virtual asset distribution operations based on the operation data incentive score and write the distribution record into the blockchain incentive ledger.

[0006] As a further aspect of the present invention, keyframe extraction and eye-tracking region annotation are performed on the original operational video stream to generate a temporal operational image segment with a gaze heatmap, including: The image pixel matrix is ​​extracted frame by frame from the original operation video stream, and the structural similarity is calculated for the image pixel matrix of two adjacent frames to obtain the inter-frame change metric. The pixel matrix of the image whose inter-frame change metric exceeds the dynamic threshold is marked as a candidate keyframe, and the candidate keyframes are sampled at equal intervals in chronological order to obtain the keyframe sequence after frame extraction. For each frame in the keyframe sequence, a pre-trained eye-tracking gaze detection model is invoked to locate the pupil center and generate the gaze point pixel coordinates. Centered on the pixel coordinates of the gaze point, a Gaussian decay heatmap layer is superimposed on the corresponding frame in the keyframe sequence to generate the time-series operational image segment with the gaze heatmap.

[0007] As a further aspect of the present invention, gesture trajectory reconstruction and pressure feature quantization are performed on the original touch event sequence to generate a touch intent encoding sequence, including: Extract the timestamp, horizontal and vertical coordinates of the touch point, and touch pressure value for each touch event from the original touch event sequence; According to the ascending order of the timestamps, the horizontal and vertical coordinates of the touch points in adjacent touch events are connected to form a touch movement trajectory polyline, and Bézier curve fitting is performed on the touch movement trajectory polyline to obtain a smoothed gesture trajectory. The smoothed gesture trajectory is divided into multiple trajectory segments, and the maximum value extraction and average value calculation are performed on the touch pressure value in each trajectory segment to obtain the segment pressure peak value and segment pressure average value. The peak pressure, average pressure, and trajectory direction angle of each trajectory segment are encoded into a touch intent code element. All touch intent codes corresponding to the trajectory segments are then concatenated in chronological order to generate the touch intent encoding sequence.

[0008] As a further aspect of the present invention, the temporal operation image segment and the touch intent encoding sequence are bidirectionally aligned and mounted on the time axis to generate a multimodal operation record package, including: Each frame of the time-series image segment with a gaze heatmap is combined with its corresponding timestamp to form an image time key-value pair, thus obtaining an image timeline queue. Each touch intent code element in the touch intent encoding sequence is combined with its corresponding timestamp to form a touch time key-value pair, thus obtaining a touch time axis queue; Using the image timeline queue as the reference timeline, traverse the touch intent code elements in the touch timeline queue, find the image time key-value pair that satisfies the preset temporal adjacency condition with the timestamp of the touch intent code element, and establish a time alignment association pair; The image with gaze heatmap in the time-aligned association pair is encapsulated with touch intent code in binary concatenation to generate the multimodal operation record package.

[0009] As a further aspect of the present invention, a pre-trained monetization value prediction model is invoked to perform incentive limit regression calculation on the multimodal operation record package, generating an operation data incentive score corresponding to the target user, including: The image sequence with gaze heatmap and the touch intent code sequence are parsed from the multimodal operation record packet, and channel separation and normalization are performed on the image sequence with gaze heatmap to obtain the gaze heatmap feature tensor. Perform one-hot discretization encoding on the touch intent code sequence and project the encoding result onto a continuous vector space to obtain the touch behavior embedding vector; The gaze thermal feature tensor and the touch behavior embedding vector are deeply concatenated along the channel dimension to form a fused operation feature matrix; The fused operational feature matrix is ​​input into the pre-trained monetization value prediction model, and then passed through the convolutional feature extraction layer and the fully connected regression layer in the model for forward calculation, outputting the operational data incentive score.

[0010] As a further aspect of the present invention, the virtual asset issuance operation is performed based on the operational data incentive score, and the issuance record is written into the blockchain incentive ledger, including: Based on the operational data incentive score, the corresponding virtual asset type and virtual asset quantity are retrieved from the preset asset exchange ratio table to generate a list of assets to be issued; The digital wallet address associated with the target user is invoked, and an on-chain transfer operation is performed according to the type and quantity of virtual assets in the list of assets to be distributed, thus completing the virtual asset distribution operation; After the on-chain transfer operation is successfully executed, the transaction hash value of this transfer is obtained from the blockchain network; The operation data incentive score, the list of assets to be issued, the transaction hash value, and the user identifier of the target user are assembled into an incentive record entry, and the incentive record entry is submitted to the smart contract of the blockchain incentive ledger for persistent writing.

[0011] As a further aspect of the present invention, after acquiring the original operation video stream and the original touch event sequence from the first-person perspective of the target user, the method further includes performing integrity verification processing on the original operation video stream and the original touch event sequence to generate a verification pass flag, including: The original operation video stream is subjected to duration statistics and frame rate detection to obtain the total number of video frames and the average video frame rate; The original touch event sequence is statistically analyzed for the number of events and the time span to obtain the total number of touch events and the touch time span. Calculate the ratio of the total number of video frames to the average frame rate of the video to obtain the actual duration of the video, and calculate the product of the total number of touch events and the preset baseline event density to obtain the expected duration of touch. Determine whether the absolute difference between the actual video duration and the expected touch duration is less than the duration tolerance threshold; if so, generate the verification pass flag.

[0012] As a further aspect of the present invention, after generating the verification pass marker, the method further includes performing privacy desensitization filtering on the original operation video stream and the original touch event sequence that have passed the integrity verification, including: For each frame of the original operation video stream that has passed the integrity check, face region detection is performed to locate a rectangular detection box containing facial biometric features; Gaussian blur filtering is applied to the pixel region within the rectangular detection box to generate an anonymized video stream after facial desensitization; The touch pressure values ​​in the original touch event sequence are clipped to a range, and touch pressure values ​​that exceed a preset pressure threshold are replaced with the preset pressure threshold to generate an anonymized touch event sequence with pressure limiting. The anonymized video stream and the anonymized touch event sequence are used as input data for subsequent steps.

[0013] As a further aspect of the present invention, after generating the operation data incentive score corresponding to the target user, the method further includes performing a decay check on the operation data incentive score to prevent duplicate incentives, including: Query the list of historical incentive records associated with the user identifier of the target user from the blockchain incentive ledger, and extract the historical incentive score sequence from the list of historical incentive records; Perform cumulative summation calculation on the historical incentive score sequence to obtain the total historical incentive score of the target user; The operation data incentive score is added to the historical incentive total score to obtain the incentive total score to be verified. Determine whether the total incentive score to be verified exceeds the preset single-user incentive limit. If it does, set the operation data incentive score to zero.

[0014] As a further aspect of the present invention, after writing the distribution records into the blockchain incentive ledger, the method further includes performing periodic summary statistics on the incentive records in the blockchain incentive ledger, including: According to a preset statistical period, all incentive record entries written within a specified period are read from the smart contract of the blockchain incentive ledger. The incentive score for each incentive record entry is summed to obtain the total periodic incentive amount. Group and aggregate the virtual asset types in each incentive record entry to obtain the total number of each virtual asset type issued within the period; The total incentive amount for the period and the total number of virtual assets issued within the period are assembled into statistical report data, and the statistical report data is stored in an off-chain data warehouse.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Keyframe extraction combined with eye-tracking region annotation is performed on the original operation video stream from a first-person perspective to generate time-series operation video segments with gaze heatmaps. This processing method can filter and retain key frames with behavioral representation significance during the operation, simultaneously mark the area of ​​user visual focus, transform abstract visual attention behavior into visualized heatmap information, expand the amount of behavioral feature information carried by the operation video, and fully restore the linkage between visual behavior and operation actions during the operation.

[0016] Gesture trajectory reconstruction combined with pressure feature quantization is performed on the original touch event sequence to generate a standardized touch intent encoding sequence. The temporal operation image fragments and the touch intent encoding sequence are bidirectionally aligned and mounted on the timeline, forming a structured multimodal operation record package. Heterogeneous operation data can achieve mutual calibration and precise matching in the temporal dimension, eliminating time offset deviations between different types of data, and standardizing fragmented operation behavior data into a unified data package format. A pre-trained monetization value prediction model can perform incentive amount regression calculations based on the standardized and complete multimodal data, producing incentive scores that closely match the actual value of the data. Virtual asset issuance behavior can be synchronously recorded in the blockchain incentive ledger, and the circulation records possess tamper-proof storage characteristics. Attached Figure Description

[0017] Figure 1 This is a timing diagram of a first-person operation data acquisition and incentive realization system according to the present invention; Figure 2 A flowchart for generating gaze heatmap time-series image segments by extracting keyframes and annotating eye movements; Figure 3 A flowchart for generating a multimodal recording package by bidirectionally aligning the timeline of temporal images with the timeline of touch intent. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 A first-person operation data acquisition and incentive monetization system includes a data acquisition module, a feature processing module, a record generation module, a value assessment module, and an incentive execution module. The data acquisition module acquires the original operation video stream and original touch event sequence of the target user from a first-person perspective. The feature processing module performs keyframe extraction and eye-tracking region annotation on the original operation video stream to generate a temporal operation image segment with a gaze heatmap, and performs gesture trajectory reconstruction and pressure feature quantization on the original touch event sequence to generate a touch intent encoding sequence. The record generation module performs bidirectional alignment and mounting of the temporal operation image segment and the touch intent encoding sequence on the timeline to generate a multimodal operation record package. The value assessment module calls a pre-trained monetization value prediction model to perform incentive amount regression calculation on the multimodal operation record package to generate an operation data incentive score corresponding to the target user. The incentive execution module executes a virtual asset distribution operation based on the operation data incentive score and writes the distribution record into a blockchain incentive ledger.

[0021] In one embodiment of the present invention, the process of performing keyframe extraction and eye-tracking region annotation on the original operational video stream to generate a temporal operational image segment with a gaze heatmap is as follows: (See attached document) Figure 2The process involves extracting the pixel matrix frame by frame from the original operational video stream and performing structural similarity calculations on the pixel matrices of adjacent frames to obtain inter-frame variation metrics. Pixel matrices with inter-frame variation metrics exceeding a dynamic threshold are marked as candidate keyframes, and these candidate keyframes are sampled at equal intervals in chronological order to obtain a sequence of extracted keyframes. For each frame in the keyframe sequence, a pre-trained eye-tracking gaze detection model is used to locate the pupil center and generate gaze point pixel coordinates. A Gaussian decay heatmap layer is overlaid on the corresponding frame in the keyframe sequence, centered on the gaze point pixel coordinates, to generate a temporal operational video segment with a gaze heatmap.

[0022] The process of generating a touch intent encoding sequence by performing gesture trajectory reconstruction and pressure feature quantization on the original touch event sequence is as follows: Extract the timestamp, horizontal and vertical coordinates of the touch point, and touch pressure value for each touch event from the original touch event sequence; connect the horizontal and vertical coordinates of the touch points in adjacent touch events in ascending order of the timestamps to form a touch movement trajectory polyline, and perform Bezier curve fitting on the touch movement trajectory polyline to obtain a smoothed gesture trajectory; divide the smoothed gesture trajectory into multiple trajectory segments, and perform maximum value extraction and mean value calculation on the touch pressure value within each trajectory segment to obtain the segment pressure peak value and segment pressure mean value; encode the segment pressure peak value, segment pressure mean value, and trajectory direction angle corresponding to each trajectory segment into a touch intent code element, and concatenate all the touch intent code elements corresponding to all trajectory segments in chronological order to generate a touch intent encoding sequence.

[0023] In practice, the target user uses a mobile terminal equipped with a front-facing camera and a touch-sensitive screen to perform interactive operations. The data acquisition module collects the target user's original operation video stream from a first-person perspective in real time. The original operation video stream includes continuous frames of the user's face orientation and hand operation area. At the same time, it collects the original touch event sequence. Each record in the original touch event sequence includes the timestamp of the touch event, the horizontal and vertical coordinates of the touch point, and the touch pressure value. The feature processing module performs keyframe extraction on the original operational video stream as follows: It extracts the pixel matrix frame by frame from the original operational video stream. The pixel matrix stores the color channel value of each pixel position in the frame in the form of a two-dimensional array. It performs structural similarity calculation on the pixel matrices of two temporally adjacent frames and uses the calculated structural similarity value as the inter-frame change metric. A dynamic threshold is set, the specific value of which is obtained by adding the sliding average of the previously processed inter-frame change metrics to a preset constant offset. When the inter-frame change metric value corresponding to a frame exceeds the dynamic threshold, the pixel matrix corresponding to that frame is marked as a candidate keyframe. After obtaining the candidate keyframe set, the candidate keyframes are sampled at equal intervals according to the original timestamp order. The number of sampling intervals is controlled by a preset extraction density parameter. The sampled candidate keyframes are retained and form a keyframe sequence. The feature processing module performs eye-tracking region annotation on the keyframe sequence as follows: For each frame in the keyframe sequence, a pre-trained eye-tracking gaze detection model is called to process the eye region image in the frame. The pre-trained eye-tracking gaze detection model consists of a multi-layer convolutional neural network and is trained on an eye-tracking dataset with pupil center coordinate annotations. The pre-trained eye-tracking gaze detection model locates the pupil center in the input eye image of the frame and outputs the gaze point pixel coordinates. The gaze point pixel coordinates are two-dimensional coordinate values, representing the pixel position on the screen where the user's gaze point falls when viewing the frame. In the frame sequence corresponding to the pixel coordinates of the gaze point, a Gaussian attenuation heatmap layer centered on the gaze point pixel coordinates is superimposed on the frame. The Gaussian attenuation heatmap layer is generated by setting the maximum heat value at the gaze point pixel coordinates as the center value, and using a Gaussian function to calculate the heat attenuation weights of other pixel positions in the layer, so that the heat value gradually decreases from the center outwards. The Gaussian attenuation heatmap layer is then semi-transparently blended and superimposed with the original frame to form an image frame with a gaze heatmap. All image frames with gaze heatmaps in the keyframe sequence are arranged in chronological order to generate a time-series operation video segment with a gaze heatmap.

[0024] The feature processing module performs gesture trajectory reconstruction and pressure feature quantization on the original touch event sequence to generate a touch intent encoding sequence as follows: The timestamp, horizontal and vertical coordinates of the touch point, and touch pressure value are extracted from each touch event in the original touch event sequence. All touch events are traversed in ascending order of timestamps. The horizontal and vertical coordinates of the touch point of the preceding and following touch events are connected by straight line segments, forming a polyline touch movement trajectory composed of multiple connected line segments. A Bézier curve fitting algorithm is used to process the polyline touch movement trajectory. The inflection points on the polyline are used as control points of the Bézier curve. The control point positions of the Bézier curve are solved using least squares approximation, generating a smooth, continuous gesture trajectory without sharp angles. The smoothed gesture trajectory is divided into multiple trajectory segments based on the curvature abrupt changes of the path points or according to a preset fixed path length interval. Each trajectory segment corresponds to a segment on the smoothed gesture trajectory. For each continuous arc segment, pressure feature quantization is performed to extract the touch pressure values ​​corresponding to all touch events contained within that segment. The maximum value is retrieved from the set of touch pressure values ​​as the segment pressure peak, and the arithmetic mean of all values ​​in the set is calculated as the segment pressure mean. The direction angle of the line connecting the start and end points of the segment is calculated, with the horizontal to the right as zero degrees and increasing by degrees counterclockwise. The calculated direction angle value is used as the trajectory direction angle. The segment pressure peak, segment pressure mean, and trajectory direction angle are combined into a triplet. After normalization, the triplet is mapped to a preset discrete codebook space to obtain a unique integer code value as the touch intent code element. A touch intent code element represents the user's touch operation behavior and intent strength characteristics within the segment. All touch intent code elements corresponding to the segment are concatenated sequentially according to the time order of the segment on the smoothed gesture trajectory to form a touch intent encoding sequence.

[0025] In one embodiment of the present invention, the method of bidirectionally aligning and mounting the temporal operation image fragment and the touch intent encoding sequence on the time axis to generate a multimodal operation record package is as follows: See [reference] Figure 3The process involves: forming image time key-value pairs by combining each frame of the time-series operation image segment with its gaze heatmap and its corresponding timestamp, thus creating an image timeline queue; forming touch time key-value pairs by combining each touch intent code element in the touch intent encoding sequence with its corresponding timestamp, thus creating a touch timeline queue; using the image timeline queue as the base timeline, traversing the touch intent code elements in the touch timeline queue, searching for image time key-value pairs whose timestamps satisfy the preset temporal adjacency condition, and establishing time alignment association pairs; and then performing binary concatenation and encapsulation of the image with the gaze heatmap in the time alignment association pairs with the touch intent code elements to generate a multimodal operation record package.

[0026] The process of using a pre-trained monetization value prediction model to perform incentive regression calculation on the multimodal operation record package to generate operation data incentive scores is as follows: Image sequences with gaze heatmaps and touch intent code sequences are parsed from the multimodal operation record package. Channel separation and normalization are performed on the image sequences with gaze heatmaps to obtain gaze heatmap feature tensors. One-hot discretization encoding is performed on the touch intent code sequences, and the encoding results are projected onto a continuous vector space to obtain touch behavior embedding vectors. The gaze heatmap feature tensors and touch behavior embedding vectors are deeply concatenated along the channel dimension to form a fused operation feature matrix. The fused operation feature matrix is ​​input into the pre-trained monetization value prediction model, and forward calculation is performed sequentially through the convolutional feature extraction layer and the fully connected regression layer in the model to output operation data incentive scores.

[0027] In specific implementation, the recording generation module performs bidirectional alignment and mounting of the time-series operation image segments with gaze heatmaps and the touch intent encoding sequence on the time axis to generate a multimodal operation recording package as follows: The recording generation module parses the time-series operation image segments with gaze heatmaps, extracts each frame of the image with gaze heatmap and its corresponding timestamp, and combines each frame of the image with gaze heatmap and its corresponding timestamp into an image time key-value pair. All image time key-value pairs are arranged in ascending order of timestamps to form an image time axis queue. At the same time, the recording generation module parses the touch intent encoding sequence, extracts each touch intent code element and its corresponding timestamp. The timestamp corresponding to the touch intent code element is the start time of the trajectory segment that generated this touch intent code element. Each touch intent code element and its corresponding timestamp are combined into a touch time key-value pair. All touch time key-value pairs are arranged in ascending order of timestamps to form a touch time axis queue. The recording generation module uses the image timeline queue as the base timeline. It iterates through each touch intent code in the touch timeline queue. For the currently iterated touch intent code, it searches the image timeline queue for the image time key-value pair whose timestamp is closest to the touch intent code's timestamp. It calculates the absolute value of the difference between the touch intent code's timestamp and the image time key-value pair's timestamp. If the absolute value is less than or equal to a preset time adjacency threshold δ, the touch intent code and this image time key-value pair are determined to meet a preset temporal adjacency condition, and they are established as a time-aligned association pair. If the absolute value is greater than the preset time adjacency threshold δ, the current touch intent code is discarded, and no association is established. The relationship for establishing time-aligned association pairs satisfies the following constraints:

[0028] in: This represents the timestamp in the image time key-value pair. Indicates the timestamp corresponding to the touch intent code. The preset time adjacency threshold is used. The value is set to an integer multiple of the maximum interval between the touch sampling interval of the acquisition device and the video frame interval. In some embodiments, a preset time adjacency threshold is used. The time interval is set to 100 milliseconds, and the timestamps of both the image timeline queue and the touch timeline queue are aligned and calibrated based on the same network time protocol server. The recording generation module collects all established time-aligned association pairs. For each time-aligned association pair, the image with the gaze heatmap in the time-aligned association pair is converted into a binary data block using a lossless image encoding format. The touch intent code elements in the time-aligned association pair are converted into a 32-bit unsigned integer binary representation. A four-byte separator is written to the end of the binary data block of the image with the gaze heatmap, and then the binary representation of the touch intent code elements is written sequentially to complete the binary splicing and encapsulation, generating a multimodal operation recording segment. The multimodal operation recording segments generated from all time-aligned association pairs are spliced ​​together according to the order of the image timestamps, and a recording header information is appended before the spliced ​​data. The recording header information includes the target user identifier, the start timestamp of the image timeline queue, and the end timestamp of the image timeline queue, finally generating a multimodal operation recording package. Optionally, if there are multiple image time key-value pairs that meet the preset temporal adjacency conditions for the same touch intent code, the record generation module selects the image time key-value pair with the smallest absolute value of the timestamp difference to establish a time alignment association pair.

[0029] In specific implementation, the process by which the value assessment module calls the pre-trained monetization value prediction model to perform incentive amount regression calculation on the multimodal operation record package to generate the operation data incentive score corresponding to the target user is as follows: The value assessment module loads the multimodal operation record package, parses the record header information in the multimodal operation record package to obtain the time range, sequentially extracts all images with gaze heatmaps from the multimodal operation record package to obtain the image sequence with gaze heatmaps, and sequentially extracts all touch intent code elements from the multimodal operation record package to obtain the touch intent code element sequence. The steps for generating a gaze heatmap feature tensor by performing channel separation and normalization on an image sequence with gaze heatmaps are as follows: Each frame of the image sequence with gaze heatmaps is separated into three color channels (red, green, and blue) and an alpha channel representing the transparency of the gaze heatmap overlay. The mean and standard deviation of pixel values ​​are calculated for each of the three color channels, and a normalization operation (subtracting the mean and dividing by the standard deviation) is performed on each pixel value in the three color channels. The maximum value of each pixel value in the alpha channel is normalized, and the alpha channel pixel value is divided by 255 to obtain a normalized heatmap value in the range [0,1]. The normalized red, green, and blue channels are stacked along the channel dimension with the normalized alpha channel to form a single-frame gaze heatmap feature tensor. The shape of the single-frame gaze heatmap feature tensor is height H multiplied by width W multiplied by the number of channels, which is 4. The gaze heatmap feature tensors of all frames in the image sequence with gaze heatmaps are arranged in chronological order to form a gaze heatmap feature tensor sequence. The process of performing one-hot discretization encoding on the touch intent code sequence and projecting it onto a continuous vector space to generate touch behavior embedding vectors is as follows: First, count the total number K possible values ​​of all touch intent codes. Construct a one-hot vector of length K for each touch intent code. The element at the index corresponding to the touch intent code value in the one-hot vector is 1, and the elements at other positions are 0. Multiply the one-hot vector by an embedding matrix of shape K rows and D columns. The elements in the embedding matrix are randomly generated during system initialization and updated through backpropagation during model training. After multiplication, a touch behavior embedding vector of dimension D is obtained, with D set to 64. Each touch intent code in the touch intent code sequence is transformed in the same way to obtain a sequence of touch behavior embedding vectors.

[0030] To fuse the gaze thermal feature tensor sequence and the touch behavior embedding vector sequence in a unified feature space, the value assessment module generates a fusion operation feature matrix for each time step: For the touch behavior embedding vector at time step t, a fully connected mapping layer is used to map it into a feature map with height H, width W, and number of channels D. The fully connected mapping layer contains a learnable weight matrix. The mapping method involves linearly transforming the touch behavior embedding vector and reshaping it into a shape of H multiplied by W multiplied by D. Then, the value of each spatial position in the mapped feature map is deeply concatenated with the gaze thermal feature tensor of the corresponding time step in the gaze thermal feature tensor sequence along the channel dimension. The shape of the fusion operation feature matrix of a single frame after concatenation is height H multiplied by width W multiplied by number of channels (4+D). The fusion operation feature matrices of all time steps are stacked to form a fusion operation feature matrix sequence, which serves as the input to the pre-trained monetization value prediction model. The pre-trained monetization value prediction model consists of a convolutional feature extraction layer and a fully connected regression layer. The convolutional feature extraction layer contains multiple 3D convolutional kernels that perform sliding convolutions on the fused operation feature matrix sequence across the time, height, and width dimensions to extract the spatiotemporal interaction patterns and output a high-level feature vector. The fully connected regression layer receives this high-level feature vector and performs regression mapping through two fully connected network layers. The first fully connected network layer contains 128 hidden units and uses a linear rectified activation function, while the second fully connected network layer outputs a scalar value without an activation function. The parameters of the pre-trained monetization value prediction model are obtained offline using massive user operation record samples and manually labeled monetization value tags. The loss function used during training is the mean squared error between the predicted and labeled incentive values. The final scalar value output by the fully connected regression layer is the incentive score for the target user's operation data. In some embodiments, the fully connected regression layer may additionally output a confidence interval, which is used to measure the estimation uncertainty of the incentive score of the operational data. In this case, the number of output units of the fully connected regression layer is set to 2, and the mean and variance are output respectively.

[0031] In one embodiment of the present invention, the steps of performing virtual asset issuance based on the operation data incentive score and writing the issuance record into the blockchain incentive ledger are as follows: querying the corresponding virtual asset type and quantity from a preset asset exchange ratio table according to the operation data incentive score to generate a list of assets to be issued; calling the digital wallet address associated with the target user and performing an on-chain transfer operation according to the virtual asset type and quantity in the list of assets to be issued to complete the virtual asset issuance operation; after the on-chain transfer operation is successfully executed, obtaining the transaction hash value of this transfer from the blockchain network; assembling the operation data incentive score, the list of assets to be issued, the transaction hash value, and the target user's user identifier into an incentive record entry, and submitting the incentive record entry to the smart contract of the blockchain incentive ledger for persistent writing.

[0032] After the distribution records are written into the blockchain incentive ledger, the process of periodically summarizing and statistically analyzing the incentive records in the blockchain incentive ledger is as follows: according to the preset statistical period, all incentive record entries written within the specified period are read from the smart contract of the blockchain incentive ledger; the incentive score of the operation data in each incentive record entry is accumulated to obtain the total incentive amount for the period; the virtual asset types in each incentive record entry are grouped and aggregated to obtain the total quantity of each virtual asset type distributed within the period; the total incentive amount for the period and the total quantity of each virtual asset type distributed within the period are assembled into statistical report data, and the statistical report data is stored in the off-chain data warehouse.

[0033] In practical implementation, the incentive execution module executes virtual asset distribution operations based on the operation data incentive score and writes the distribution record to the blockchain incentive ledger as follows: After obtaining the operation data incentive score generated by the value assessment module, the incentive execution module retrieves a preset asset exchange ratio table using the operation data incentive score as the query key. The preset asset exchange ratio table is stored in the read-only memory built into the incentive execution module. The preset asset exchange ratio table defines the mapping relationship between the numerical range in which the operation data incentive score falls and the virtual asset type and quantity. A data structure for a preset asset exchange ratio table is shown in Table 1.

[0034] Table 1: Asset Exchange Rate Mapping Table

[0035] The incentive execution module compares the incentive score of the operation data with the lower and upper limits of the incentive score in the asset exchange ratio mapping table row by row. It locates the row within the range of the operation data incentive score, reads the corresponding virtual asset type and quantity, and generates a list of assets to be distributed. This list includes the target user identifier, the operation data incentive score, a list of virtual asset types to be distributed, and a list of virtual asset quantities to be distributed. The incentive execution module calls the digital wallet address associated with the target user. This digital wallet address is generated through the blockchain wallet creation interface when the target user registers in the system. The public key of the digital wallet address is stored in the user profile, and the private key is securely stored by the target user's client. The incentive execution module constructs a blockchain transfer transaction. The target address of the blockchain transfer transaction is the digital wallet address. The call parameters of the blockchain transfer transaction include the virtual asset type and quantity from the list of assets to be distributed. The initiating point of the blockchain transfer transaction is the system incentive distribution contract address. The incentive execution module uses the signing key of the system incentive distribution contract address to sign the blockchain transfer transaction and broadcasts it to the blockchain network. After the consensus nodes in the blockchain network verify the legality of the signature, they execute an on-chain transfer operation in the smart contract virtual machine, transferring the specified virtual asset type and quantity from the system incentive distribution contract address to the target user's associated digital wallet address, thus completing the virtual asset distribution operation.

[0036] After a successful on-chain transfer, the incentive execution module queries the transaction receipt from the full node interface of the blockchain network. It then extracts the transaction hash value from the receipt; the transaction hash value is a fixed-length hexadecimal string calculated using a hash algorithm. The incentive execution module assembles the operation data incentive score, the list of assets to be distributed, the transaction hash value, and the target user's user identifier into an incentive record entry. The assembly method involves converting the operation data incentive score into a text string, sequentially concatenating the virtual asset types and quantities from the list of assets to be distributed into a comma-separated text string, preserving the transaction hash value in its original hexadecimal string format, and preserving the user identifier in a universally unique identifier format. These four data fields are connected by vertical bar separators. The incentive execution module calls the write function exposed in the smart contract of the blockchain incentive ledger. The write function is named recordIncentive. The write function receives the incentive record entry string and the caller's identity signature as input parameters. The incentive execution module constructs a call transaction and sends the call transaction to the contract address corresponding to the blockchain incentive ledger. After receiving the call transaction, the smart contract of the blockchain incentive ledger appends the incentive record entry string to the log storage mapping maintained inside the contract. The key of the log storage mapping is the combined hash value of the user identifier and the timestamp, and the value of the log storage mapping is the incentive record entry string, thus completing the persistent write.

[0037] After the incentive records are written to the blockchain incentive ledger, the system performs a periodic summary and statistical analysis of the incentive records in the blockchain incentive ledger as follows: A scheduled task is set up within the system. The scheduled task triggers the summary and statistical task according to a preset statistical period, which is set to be once every 24 hours. After the summary and statistical task starts, it reads all incentive record entries written within the specified period from the smart contract of the blockchain incentive ledger. The reading method is to call the rangeQuery function exposed in the smart contract. The rangeQuery function receives a start timestamp and an end timestamp as input parameters. The start timestamp is the current time minus the number of seconds corresponding to the preset statistical period, and the end timestamp is the current time. The smart contract iterates through all key-value pairs in the log storage mapping, filters out log storage mapping key-value pairs whose timestamps fall within the specified period, and returns a list of corresponding incentive record entry strings. The summary and statistical task parses the returned list of incentive record entry strings one by one, and performs an accumulation operation on the operation data incentive score in each incentive record entry to obtain the total periodic incentive amount. The total periodic incentive amount is calculated as the sum of the operation data incentive scores of all incentive record entries within the period, expressed as:

[0038] in: This represents the total amount of periodic incentives. This indicates the total number of stimulus record entries within the period. Indicates the first The incentive score is the operational data in each incentive record entry. In some embodiments, the summary statistics task parses and accumulates the data while reading incentive record entries in a streaming manner, without needing to load all incentive record entry strings into memory simultaneously. The summary statistics task groups and aggregates the virtual asset types in each incentive record entry. The grouping and aggregation method is to use the virtual asset type as the grouping key, traversing all the parsed virtual asset types and virtual asset quantities in all incentive record entries. When the same virtual asset type is encountered, the virtual asset quantity is added to the count value of the corresponding grouping key. After the traversal is completed, the total number of each virtual asset type issued within the period is obtained. It can be understood that for incentive record entries containing multiple virtual asset types in the asset list to be issued, the summary statistics task performs grouping and aggregation counting independently for each virtual asset type.

[0039] The summary statistics task assembles the total incentive amount for the period and the total issuance quantity of each virtual asset type within the period according to a predefined statistical report format. This report format includes the start and end timestamps of the statistical period, the total incentive amount, an enumeration list of virtual asset types, and a list of issuance quantities corresponding to each virtual asset type. The summary statistics task sends the statistical report data to an off-chain data warehouse via a secure transmission layer encryption protocol. The off-chain data warehouse is a relational database deployed in an offline server cluster. Upon receiving the statistical report data, the off-chain data warehouse parses the fields in the statistical report format, writes the total incentive amount into the period incentive summary table, writes the total issuance quantity of each virtual asset type into the asset issuance details table, and records the corresponding statistical period identifier in the asset issuance details table to support historical data backtracking. Optionally, if the blockchain network becomes congested when the preset statistical period is triggered, causing a timeout in reading incentive record entries, the summary statistics task executes a retry mechanism, retrying up to three times with an exponential backoff retry interval. If it still fails after three attempts, an alarm message is output to the system monitoring platform.

[0040] In one embodiment of the present invention, after acquiring the original operation video stream and the original touch event sequence from the first-person perspective of the target user, the process of performing integrity verification processing on the original operation video stream and the original touch event sequence to generate a verification pass mark is as follows: The original operation video stream is subjected to duration statistics and frame rate detection to obtain the total number of video frames and the average video frame rate; the original touch event sequence is subjected to event quantity statistics and time span calculation to obtain the total number of touch events and the touch time span; the ratio of the total number of video frames to the average video frame rate is calculated to obtain the actual video duration, and the product of the total number of touch events and the preset baseline event density is calculated to obtain the expected touch duration; it is determined whether the absolute difference between the actual video duration and the expected touch duration is less than the duration tolerance threshold, and if so, a verification pass mark is generated.

[0041] After generating the verification pass mark, the privacy desensitization filtering steps for the original operation video stream and original touch event sequence that passed the integrity verification are as follows: perform face region detection on each frame of the original operation video stream that passed the integrity verification, and locate the rectangular detection box containing facial biometric features; perform Gaussian blur filtering on the pixel region within the rectangular detection box to generate an anonymized video stream after face desensitization; perform numerical range clipping on the touch pressure values ​​in the original touch event sequence, and replace touch pressure values ​​exceeding the preset pressure threshold with the preset pressure threshold to generate an anonymized touch event sequence after pressure limiting; use the anonymized video stream and anonymized touch event sequence as input data for subsequent steps.

[0042] In specific implementation, after the data acquisition module obtains the original operation video stream and original touch event sequence from the first-person perspective of the target user, it performs integrity verification processing on the original operation video stream and original touch event sequence to generate verification through the following method: calling the video parsing component to decode the original operation video stream frame by frame and accumulate the frame count to obtain the total number of video frames, and reading the encapsulation layer information of the original operation video stream from the video parsing component, extracting the nominal frame rate value from the encapsulation layer information as the average frame rate of the video; calling the touch event parsing component to count records and extract the first and last timestamps of the original touch event sequence, taking the number of records in the touch event sequence as the total number of touch events, and taking the difference between the timestamp of the last touch event and the timestamp of the first touch event as the touch time span. After obtaining four sets of basic statistics—total video frames, average video frame rate, total number of touch events, and touch duration—the actual video duration is calculated as the ratio of the total video frames to the average video frame rate. The expected touch duration is calculated as the product of the total number of touch events and the preset baseline event density. The preset baseline event density is set in the system according to the touch sampling frequency corresponding to the mobile terminal model used by the target user. One optional configuration parameter value is shown in Table 2.

[0043] Table 2: Preset Baseline Event Density Configuration Table

[0044] It is understandable that the preset baseline event density is the reciprocal of the touch sampling frequency. When the original touch event sequence is complete and without dropped frames, the expected touch duration should be close to the actual video duration. The integrity check criterion is whether the absolute difference between the actual video duration and the expected touch duration is less than the duration tolerance threshold. The duration tolerance threshold is set to the greater of 5% of the actual video duration and a fixed compensation value of 1.5 seconds. The mathematical expression of the judgment rule is as follows:

[0045] in: Indicates the actual duration of the video. Indicates the expected duration of touch input. Indicates taking and The larger of the two values ​​is considered. When the above inequality holds, it is determined that the original operation video stream and the original touch event sequence meet the consistency requirements in the time dimension, a verification pass flag is generated, and the verification pass flag is bound to the original operation video stream and the original touch event sequence and passed to the next processing stage; when the inequality does not hold, it is determined that there is an abnormality in data integrity, the original operation video stream and the original touch event sequence of the current batch are discarded, and a re-acquisition instruction is sent to the data acquisition module. In some embodiments, after receiving the re-acquisition instruction, the data acquisition module will trigger the target user client to pop up a prompt interface for re-authorizing data acquisition.

[0046] After generating a verification pass marker, the system performs privacy desensitization filtering on the original operation video stream and original touch event sequence that passed integrity verification. The privacy desensitization processing component loads the original operation video stream that passed integrity verification, performs face region detection on each frame of the original operation video stream, calls a pre-trained lightweight face detection model to slide candidate boxes on the frame and calculates the similarity between the image features within the candidate boxes and the face feature template. When the similarity exceeds a preset face confidence threshold, the candidate box is marked as a rectangular detection box containing facial biometric features. The lightweight face detection model adopts a single-stage detection architecture and is trained to convergence on a public face dataset. Each frame may detect zero or one or more rectangular detection boxes. Gaussian blur filtering is performed on the pixel region within each rectangular detection box. The Gaussian blur filtering uses a Gaussian kernel of size 15 pixels by 15 pixels, with the distribution parameter σ of the Gaussian kernel set to 3.0. The Gaussian blur filtering process involves sliding the Gaussian kernel pixel-by-pixel within the rectangular detection box and replacing the center pixel value with a weighted sum of the neighboring pixel values. After applying Gaussian blur filtering to all detected rectangular detection boxes, an anonymized video stream with desensitized faces is obtained. In the anonymized video stream, the non-face areas of the frames retain their original pixel values.

[0047] The privacy desensitization component simultaneously performs numerical range pruning on the original touch event sequence to generate an anonymized touch event sequence. This anonymized sequence is generated by iterating through each touch event record in the original sequence, extracting the touch pressure value from each record, and comparing it to a preset pressure threshold. The preset pressure threshold is set to 90% of the maximum range defined in the touch sensor hardware specifications for motion touch scenarios. When the touch pressure value is less than or equal to the preset pressure threshold, the original value is retained. When the value exceeds the threshold, the value is replaced with the threshold. It is understood that touch pressure values ​​exceeding the preset threshold may reflect individual biometric characteristics of the user's finger pressure; truncating the upper limit reduces the risk of identifying the user through pressure patterns. In some embodiments, the horizontal and vertical coordinates of the touch points in the original touch event sequence are also subjected to spatial perturbation processing, with the perturbation amplitude set to no more than two-thousandths of the touch screen pixel density. The privacy desensitization processing component writes the anonymized video stream generated by Gaussian blur filtering and the anonymized touch event sequence generated by numerical range cropping into the input buffer queue of the feature processing module, respectively. The anonymized video stream and the anonymized touch event sequence serve as input data for subsequent feature extraction and encoding stages.

[0048] In one embodiment of the present invention, after generating the operation data incentive score corresponding to the target user, the process of performing attenuation verification on the operation data incentive score to prevent duplicate incentives is as follows: query the list of historical incentive records associated with the user identifier of the target user from the blockchain incentive ledger, and extract the historical incentive score sequence from the historical incentive record list; perform cumulative summation calculation on the historical incentive score sequence to obtain the historical incentive total score of the target user; add the operation data incentive score to the historical incentive total score to obtain the incentive total score to be verified; determine whether the incentive total score to be verified exceeds the preset single-user incentive limit, and if it exceeds, set the operation data incentive score to zero.

[0049] In practical implementation, after generating the operation data incentive score corresponding to the target user, the system performs attenuation verification on the operation data incentive score to prevent duplicate incentives. The process is as follows: The attenuation verification module queries the smart contract of the blockchain incentive ledger for a list of historical incentive records associated with the user identifier of the target user. The query method is to call the queryByUser function exposed in the smart contract. The queryByUser function receives the user identifier as input parameter, traverses all key-value pairs with the user identifier as a prefix in the internal log storage mapping of the contract, and returns a list of all historical incentive record entries under that user identifier. The attenuation verification module parses each incentive record entry in the historical incentive record list, extracts the incentive score field recorded in the incentive record entry, and arranges all the extracted incentive scores in ascending order according to the write timestamp of the incentive record entry to form a historical incentive score sequence. The attenuation verification module performs a cumulative summation calculation on the historical incentive score sequence to obtain the target user's total historical incentive score. The cumulative summation method is to add each element in the historical incentive score sequence one by one, starting from the first element, until all elements in the historical incentive score sequence have been traversed. The accumulated sum is the total historical incentive score. The calculation process of the total historical incentive score is expressed as follows:

[0050] in: This represents the target user's total historical incentive score. This indicates the total number of historical incentive record entries. Indicates the th in the historical incentive score sequence The historical incentive score corresponding to each record.

[0051] The attenuation verification module adds the current operational data incentive score generated by the value assessment module to the historical total incentive score to obtain the total incentive score to be verified. This total incentive score is then compared with a preset single-user incentive cap. The preset single-user incentive cap is a global constant configured by the system management terminal. The value of the single-user incentive cap is set to a fixed value according to the operational strategy; in one configuration, the preset single-user incentive cap is set to 5000. When the total incentive score to be verified exceeds the preset single-user incentive cap, it is determined that the current operational data incentive score will cause the target user's historical cumulative incentive amount to exceed the maximum allowed by the system. The attenuation verification module sets the operational data incentive score to zero and generates an attenuation verification record written to the attenuation log. The attenuation log includes the target user's user identifier, the original operational data incentive score, the historical total incentive score, the attenuated operational data incentive score, and the attenuation timestamp. In some embodiments, after setting the operational data incentive score to zero, the attenuation verification module sends a notification message to the target user's client through the system message notification interface, indicating that the incentive cap has been reached. When the total incentive score to be verified is less than or equal to the preset single-user incentive limit, the attenuation verification module keeps the operation data incentive score unchanged and passes the original value of the operation data incentive score to the incentive execution module for subsequent virtual asset issuance operations.

[0052] It is understood that the calculation range of the historical incentive total score is defined based on the preset incentive cumulative effective period, which is set as the span of natural months from the current moment backward. When querying the historical incentive record list, the decay verification module appends the start timestamp and end timestamp to the input parameters of the queryByUser function. The start timestamp is the current moment minus the number of seconds corresponding to the incentive cumulative effective period, and the end timestamp is the current moment. The smart contract only returns historical incentive record entries whose write timestamps fall within the interval between the start and end timestamps, and the historical incentive total score is calculated only based on the historical incentive score sequence within this time interval. In some embodiments, the single-user incentive upper limit is not a fixed value, but is set according to the user level of the target user. The mapping relationship between user level and single-user incentive upper limit is stored in the system configuration center. Before performing numerical comparison, the decay verification module first retrieves the single-user incentive upper limit value corresponding to the target user level from the configuration center. Optionally, when the blockchain incentive ledger times out due to network partitioning, the decay verification module enables a local cache copy. The local cache copy is updated synchronously by the decay verification module after each successful writing of an incentive record entry. The local cache copy stores the most recently successfully queried historical incentive total score and query timestamp. If the interval between the query timestamp of the local cache copy and the current time exceeds the preset cache validity period, the local cache copy is rejected and a decay verification failure alarm is triggered, temporarily suspending the issuance of incentives. The decay verification process will be re-executed after the blockchain incentive ledger becomes available again.

[0053] After the decay verification module completes the decay verification process for the operation data incentive score, if the operation data incentive score is not set to zero, it sends the operation data incentive score along with the decay verification to the incentive execution module via a flag. The incentive execution module then executes the virtual asset distribution operation based on the operation data incentive score. If the operation data incentive score is set to zero, the decay verification module submits the zero-value operation data incentive score and the decay verification record to the blockchain incentive ledger for notarization. The submission method is to call the recordIncentive write function of the smart contract, filling the operation data incentive score field with zero, and filling the virtual asset type and virtual asset quantity fields with empty values ​​in the passed incentive record entry. A decay flag is also added to the incentive record entry to indicate that the current record is a zero-value record generated by the decay verification. By notarizing the zero-value record in the blockchain incentive ledger, the system maintains a complete incentive decision audit trail, and the control logic of the single-user incentive cap is transparent and verifiable on the consensus nodes of the entire network.

[0054] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A first-person operational data acquisition and incentive monetization system, characterized in that, include: The data acquisition module is used to acquire the original operation video stream and original touch event sequence of the target user from a first-person perspective; The feature processing module is used to perform keyframe extraction and eye-tracking region annotation on the original operation video stream to generate a time-series operation video segment with a gaze heatmap, and to perform gesture trajectory reconstruction and pressure feature quantization on the original touch event sequence to generate a touch intent encoding sequence. The recording generation module is used to perform bidirectional alignment and mounting of the time sequence operation image fragment and the touch intent encoding sequence on the time axis to generate a multimodal operation recording package; The value assessment module is used to call a pre-trained monetization value prediction model to perform incentive amount regression calculation on the multimodal operation record package and generate the operation data incentive score corresponding to the target user. The incentive execution module is used to perform virtual asset distribution operations based on the operation data incentive score and write the distribution record into the blockchain incentive ledger.

2. The first-person operation data acquisition and incentive realization system according to claim 1, characterized in that, The original operational video stream is subjected to keyframe extraction and eye-tracking region annotation to generate a temporal operational video segment with a gaze heatmap, including: The image pixel matrix is ​​extracted frame by frame from the original operation video stream, and the structural similarity is calculated for the image pixel matrix of two adjacent frames to obtain the inter-frame change metric. The pixel matrix of the image whose inter-frame change metric exceeds the dynamic threshold is marked as a candidate keyframe, and the candidate keyframes are sampled at equal intervals in chronological order to obtain the keyframe sequence after frame extraction. For each frame in the keyframe sequence, a pre-trained eye-tracking gaze detection model is invoked to locate the pupil center and generate the gaze point pixel coordinates. Centered on the pixel coordinates of the gaze point, a Gaussian decay heatmap layer is superimposed on the corresponding frame in the keyframe sequence to generate the time-series operational image segment with the gaze heatmap.

3. The first-person operation data acquisition and incentive realization system according to claim 1, characterized in that, The original touch event sequence is subjected to gesture trajectory reconstruction and pressure feature quantization to generate a touch intent encoding sequence, including: Extract the timestamp, horizontal and vertical coordinates of the touch point, and touch pressure value for each touch event from the original touch event sequence; According to the ascending order of the timestamps, the horizontal and vertical coordinates of the touch points in adjacent touch events are connected to form a touch movement trajectory polyline, and Bézier curve fitting is performed on the touch movement trajectory polyline to obtain a smoothed gesture trajectory. The smoothed gesture trajectory is divided into multiple trajectory segments, and the maximum value extraction and average value calculation are performed on the touch pressure value in each trajectory segment to obtain the segment pressure peak value and segment pressure average value. The peak pressure, average pressure, and trajectory direction angle of each trajectory segment are encoded into a touch intent code element. The touch intent codes elements corresponding to all trajectory segments are then concatenated in chronological order to generate the touch intent encoding sequence.

4. The first-person operation data acquisition and incentive realization system according to claim 1, characterized in that, The time-series operation image fragments and the touch intent encoding sequence are bidirectionally aligned and mounted on the timeline to generate a multimodal operation record package, including: Each frame of the time-series image segment with a gaze heatmap is combined with its corresponding timestamp to form an image time key-value pair, thus obtaining an image timeline queue. Each touch intent code element in the touch intent encoding sequence is combined with its corresponding timestamp to form a touch time key-value pair, thus obtaining a touch time axis queue; Using the image timeline queue as the reference timeline, traverse the touch intent code elements in the touch timeline queue, find the image time key-value pair that satisfies the preset temporal adjacency condition with the timestamp of the touch intent code element, and establish a time alignment association pair; The image with gaze heatmap in the time-aligned association pair is encapsulated with touch intent code in binary concatenation to generate the multimodal operation record package.

5. The first-person operation data acquisition and incentive realization system according to claim 1, characterized in that, The pre-trained monetization value prediction model is invoked to perform incentive limit regression calculation on the multimodal operation record package, generating an operation data incentive score corresponding to the target user, including: The image sequence with gaze heatmap and the touch intent code sequence are parsed from the multimodal operation record packet, and channel separation and normalization are performed on the image sequence with gaze heatmap to obtain the gaze heatmap feature tensor. Perform one-hot discretization encoding on the touch intent code sequence and project the encoding result onto a continuous vector space to obtain the touch behavior embedding vector; The gaze thermal feature tensor and the touch behavior embedding vector are deeply concatenated along the channel dimension to form a fused operation feature matrix; The fused operational feature matrix is ​​input into the pre-trained monetization value prediction model, and then passed through the convolutional feature extraction layer and the fully connected regression layer in the model for forward calculation, outputting the operational data incentive score.

6. The first-person operation data acquisition and incentive realization system according to claim 1, characterized in that, Based on the operational data incentive score, virtual asset distribution operations are performed, and the distribution records are written into the blockchain incentive ledger, including: Based on the operational data incentive score, the corresponding virtual asset type and virtual asset quantity are retrieved from the preset asset exchange ratio table to generate a list of assets to be issued; The digital wallet address associated with the target user is invoked, and an on-chain transfer operation is performed according to the type and quantity of virtual assets in the list of assets to be distributed, thus completing the virtual asset distribution operation; After the on-chain transfer operation is successfully executed, the transaction hash value of this transfer is obtained from the blockchain network; The operation data incentive score, the list of assets to be issued, the transaction hash value, and the user identifier of the target user are assembled into an incentive record entry, and the incentive record entry is submitted to the smart contract of the blockchain incentive ledger for persistent writing.

7. The first-person operation data acquisition and incentive realization system according to claim 1, characterized in that, After acquiring the original operation video stream and original touch event sequence of the target user from a first-person perspective, the method further includes performing integrity verification processing on the original operation video stream and the original touch event sequence, generating a verification pass flag, including: The original operation video stream is subjected to duration statistics and frame rate detection to obtain the total number of video frames and the average video frame rate; The original touch event sequence is statistically analyzed for the number of events and the time span to obtain the total number of touch events and the touch time span. Calculate the ratio of the total number of video frames to the average frame rate of the video to obtain the actual duration of the video, and calculate the product of the total number of touch events and the preset baseline event density to obtain the expected duration of touch. Determine whether the absolute difference between the actual video duration and the expected touch duration is less than the duration tolerance threshold; if so, generate the verification pass flag.

8. A first-person operation data acquisition and incentive realization system according to claim 7, characterized in that, After generating the verification pass marker, the method further includes performing privacy desensitization filtering on the original operation video stream and the original touch event sequence that passed the integrity verification, including: For each frame of the original operation video stream that has passed the integrity check, face region detection is performed to locate a rectangular detection box containing facial biometric features; Gaussian blur filtering is applied to the pixel region within the rectangular detection box to generate an anonymized video stream after facial desensitization; The touch pressure values ​​in the original touch event sequence are clipped to a range, and touch pressure values ​​that exceed a preset pressure threshold are replaced with the preset pressure threshold to generate an anonymized touch event sequence with pressure limiting. The anonymized video stream and the anonymized touch event sequence are used as input data for subsequent steps.

9. A first-person operational data acquisition and incentive realization system according to claim 1, characterized in that, After generating the operation data incentive score corresponding to the target user, the method further includes performing a decay check on the operation data incentive score to prevent duplicate incentives, including: Query the list of historical incentive records associated with the user identifier of the target user from the blockchain incentive ledger, and extract the historical incentive score sequence from the list of historical incentive records; Perform cumulative summation calculation on the historical incentive score sequence to obtain the total historical incentive score of the target user; The operation data incentive score is added to the historical incentive total score to obtain the incentive total score to be verified. Determine whether the total incentive score to be verified exceeds the preset single-user incentive limit. If it does, set the operation data incentive score to zero.

10. A first-person operational data acquisition and incentive realization system according to claim 1, characterized in that, After writing the distribution records into the blockchain incentive ledger, the process also includes periodically summarizing and statistically analyzing the incentive records in the blockchain incentive ledger, including: According to a preset statistical period, all incentive record entries written within a specified period are read from the smart contract of the blockchain incentive ledger. The incentive score for each incentive record entry is summed to obtain the total periodic incentive amount. Group and aggregate the virtual asset types in each incentive record entry to obtain the total number of each virtual asset type issued within the period; The total incentive amount for the period and the total number of virtual assets issued within the period are assembled into statistical report data, and the statistical report data is stored in an off-chain data warehouse.

Images