A robot advertisement publishing effect evaluation and feedback learning method

Abstract

The application relates to the field of mobile advertisement publishing and intelligent control, and discloses a robot advertisement publishing effect evaluation and feedback learning method; the method divides a continuous advertisement publishing process into multiple publishing segments, collects scene state parameters and publishing action parameters corresponding to the multiple publishing segments, screens effective exposure targets, and processes repeated passing behaviors of the same target; the net attraction effect value, the net interaction effect value and the net guidance effect value are extracted by combining a local baseline window and a matching pair of photos, and feedback samples are generated; the advertisement publishing strategy is updated in a preset action constraint set, and the current strategy is maintained or switched to a basic publishing strategy when the confidence is insufficient or abnormal interference exists. Through the above mode, the accuracy of robot advertisement publishing effect evaluation and the pertinence of subsequent strategy adjustment are improved.

Description

Technical Field

[0001] This invention relates to the field of mobile advertising delivery and intelligent control, specifically to a method for evaluating and learning the effectiveness of robot advertising delivery. Background Technology

[0002] In shopping mall aisles, exhibition hall corridors, waiting areas, and brand display areas, advertising robots are used to move along preset routes and display advertisements and provide interactive guidance at multiple stops using components such as displays and speakers. These devices typically integrate a moving mechanism, an advertising output unit, a sensing unit, and a control unit to achieve simultaneous mobile advertising or fixed-point advertising.

[0003] In existing technologies, advertising robots mostly play advertisements according to preset times, preset locations, or preset material sequences. They also collect data such as the number of viewers, the number of scans, the duration of dwell time, or the number of touches by combining cameras, QR code scanning components, or interactive interfaces to judge the effectiveness of the advertising. Some solutions will also adjust the subsequent playback content or playback rhythm based on the statistical results.

[0004] However, when existing robots move and deploy at multiple locations, the display unit and the sensing unit change synchronously with the position and orientation of the vehicle. During use, the same human target is likely to repeatedly enter the sensing area at adjacent locations or at adjacent time periods. In addition, the density of people, the direction of passage, the state of obstruction, and environmental interference on site will also fluctuate continuously. This causes the collected exposure data and interaction data to be mixed with natural traffic changes and repeated passage information, thereby disrupting the correspondence between the deployment action and the feedback result. Ultimately, this leads to a distortion in the evaluation of advertising effectiveness and makes it difficult to provide an accurate basis for subsequent deployment strategy adjustments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for evaluating and learning the effectiveness of robot advertising, thereby resolving the technical problems present in the prior art.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A method for evaluating and learning the effectiveness of robot advertising delivery, applied to an advertising delivery robot, the advertising delivery robot including a mobile mechanism, an advertising output unit, an environmental perception unit, an interaction detection unit, and a control unit, the method comprising the following steps: S1: The control unit divides the continuous advertising process into multiple delivery segments based on the preset time length, robot pose change amount, or advertising release action switching event, and records the scene state parameters and release action parameters corresponding to each delivery segment. S2: The environmental perception unit collects information on the relative position, head orientation, dwell trajectory, occlusion status, and distance between the human target and the advertising output unit within each delivery segment. The control unit filters valid exposure targets based on the following conditions: the human target's head area or torso center point is located within the advertising visible coverage area; the head orientation is pointing towards the advertising output unit; the continuous dwell time reaches a preset threshold; the distance is within a preset range; and the occlusion ratio is not greater than a preset threshold. S3: The control unit performs anonymous trajectory association based on the continuity of the effective exposure target's position, direction of motion, and timing of appearance in consecutive frames. It identifies the repeated passing behavior of the same target in the same delivery segment or adjacent delivery segments, and includes subsequent exposures in the segment-level effective exposure data according to a preset attenuation coefficient. S4: The interaction detection unit detects the interaction events generated by the effective exposure target in the current delivery segment. The control unit establishes a local baseline window based on the current delivery segment and selects matching comparison segments from historical delivery segments or adjacent time period segments that meet the preset similar conditions in terms of site grid location, time period category and pedestrian flow status, and whose release action parameters are different or have not executed the advertising release action. S5: The control unit compares the effective exposure data and interaction event data of the current delivery segment with the local baseline window and the matching control segment, extracts the net attraction effect value, net interaction effect value and net guidance effect value of the current delivery segment relative to natural pedestrian flow fluctuations and environmental changes, and generates feedback samples based on the net attraction effect value, net interaction effect value, net guidance effect value, scene state parameters, release action parameters, repeated processing results and abnormal interference markers that characterize abnormal fluctuations in pedestrian flow density, occlusion degree or environmental noise. S6: The control unit updates the advertising publishing strategy based on feedback samples within the preset action constraint set. When the confidence level of the feedback sample is lower than the preset threshold or the abnormal interference flag is triggered, the current advertising publishing strategy remains unchanged or switches to the preset basic publishing strategy. S7: The control unit controls the advertising robot to adjust the publishing action parameters in subsequent delivery segments based on the updated advertising publishing strategy, so as to form a closed-loop update process for advertising publishing effect evaluation and feedback learning.

[0007] Preferably, the scene state parameters include site grid location, time period, pedestrian density, pedestrian traffic direction, ambient light intensity, ambient noise intensity, and degree of occlusion; The parameters for the release action include the robot's stopping point, robot orientation, advertising output unit orientation, advertising material identifier, broadcast volume, broadcast duration, and interaction trigger threshold.

[0008] Preferably, when the control unit filters valid exposure targets, it simultaneously meets the following conditions: the head region or torso center point of the human target is located within the visible coverage area of ​​the advertisement; The angle between the head orientation of the human target and the normal of the display surface of the advertising output unit is less than or equal to the first angle threshold. The continuous dwell time of a human target within the current delivery segment is greater than or equal to the first time threshold; The distance between the human target and the advertising output unit is between a first distance threshold and a second distance threshold; The occlusion ratio corresponding to the human target is less than or equal to the first occlusion threshold.

[0009] Preferably, when the control unit performs anonymous trajectory association, it jointly determines the positional deviation, motion direction deviation, and occurrence time interval between the human target in the current frame and the human target in the previous frame; When the positional deviation is less than or equal to the first displacement threshold, the angle of the movement direction is less than or equal to the second angle threshold, and the time interval is less than or equal to the second time threshold, the corresponding human target will be identified as the same anonymous trajectory.

[0010] Preferably, when the control unit identifies repeated passing behavior, if the same anonymous trajectory enters the visible coverage area of ​​the advertisement multiple times in the same delivery segment or adjacent delivery segments, and the time interval between two adjacent entries is less than or equal to a third time threshold, it is determined to be repeated passing behavior. The first exposure is counted as a valid exposure in the segment-level valid exposure data, and each subsequent exposure is counted as a segment-level valid exposure data according to a preset attenuation coefficient, which is greater than 0 and less than 1.

[0011] Preferably, the interactive events include QR code scanning events, voice response events, touch events, proximity dwell events, and preset guidance area arrival events; The scanning event is an event formed when a human target reads the QR code or electronic tag displayed by the advertising output unit. The voice response event is an event formed when the interaction detection unit collects the human target's response voice and identifies the preset response word. The touch event is an event formed when a human target touches the interactive interface of the advertising output unit. The near-distance dwell event is an event formed when the distance between the human target and the advertising output unit is less than or equal to a third distance threshold and the duration is greater than or equal to a fourth time threshold. The preset guidance area arrival event is an event formed when the human target moves from the visible coverage area of ​​the advertisement to the preset guidance area.

[0012] Preferably, when the control unit filters the matching comparison segment, it uses the current delivery segment as a benchmark and selects segments from historical delivery segments or segments from adjacent time periods that simultaneously meet the following conditions as matching comparison segments: the site grid positions are the same; The time periods belong to the same time period category; The difference in pedestrian density is less than or equal to the first density threshold; The deviation of the pedestrian flow direction is less than or equal to the third angle threshold; The difference in ambient light intensity is less than or equal to the first light threshold. The difference in environmental noise intensity is less than or equal to the first noise threshold; Furthermore, at least one of the parameters for the published action is different from that of the currently delivered segment, or the corresponding segment has not executed the ad publishing action.

[0013] Preferably, when the control unit extracts the net attraction effect value, net interaction effect value, and net guidance effect value, it uses the effective exposure number, interaction event number, and preset guidance area arrival number of the current delivery segment as the current value, and the weighted average value calculated by the local baseline window and the corresponding data of the matching control segment according to the preset weighting rule as the reference value. The net attraction effect value, net interaction effect value, and net guidance effect value are obtained by subtracting the reference value from the current value. The preset weighting rule determines the weight of each matched comparison segment based on the segment time difference and the scene state parameter difference.

[0014] Preferably, the feedback sample includes scene state parameters, release action parameters, net attraction effect value, net interaction effect value, net guidance effect value, repeated processing results, abnormal interference markers, and confidence level; The abnormal interference flag is triggered by at least one of the following conditions: the pedestrian density exceeds the second density threshold within a preset time, the degree of obstruction exceeds the second obstruction threshold, the environmental noise intensity exceeds the second noise threshold, and the duration of obstruction within the visible coverage area of ​​the advertisement is greater than or equal to the fifth time threshold. The confidence level is determined based on the number of matching comparison segments, the difference in scene state parameters between the currently delivered segment and the matching comparison segments, and the completeness of the interaction event data.

[0015] Preferably, when the control unit updates the advertising publishing strategy within a preset set of action constraints, it only adjusts the robot's stopping point, robot orientation, advertising output unit orientation, advertising material identifier, broadcast volume, broadcast duration, and interaction trigger threshold. The adjustment range of the robot's stopping point is limited to the current field grid or field grids adjacent to the current field grid. The adjustment range of the robot orientation is limited to a preset angle range on both sides of the current orientation. The adjustment range of the advertising output unit orientation is limited to a preset angle range on both sides of the current orientation. The broadcast volume is limited to a preset volume range. The broadcast duration is limited to a preset duration range. The interaction trigger threshold is limited to a preset threshold range. When the confidence level of the feedback sample is lower than a preset threshold or an abnormal interference flag is triggered, the control unit keeps the current publishing action parameters unchanged or adjusts the publishing action parameters to a preset basic parameter combination.

[0016] In summary, the present invention has the following main beneficial effects: This application divides the continuous advertising process of a robot into multiple segments and records scene state parameters and publishing action parameters for each segment. This achieves a one-to-one correspondence between the advertising effect and the robot's actual execution state under specific stopping points, orientations, materials, and broadcast configurations. Existing solutions often rely solely on the entire playback log, total number of viewers, or total number of interactions to statistically analyze advertising effectiveness. This is easily affected by natural fluctuations in pedestrian traffic, changes in robot position, and short-term environmental disturbances, leading to unclear correlations between evaluation results and actual publishing actions. This application, by collecting data and evaluating effects at the segment level, ensures that each advertising publishing action has clear temporal and spatial boundaries. This refines the effect attribution path, improves the accuracy of advertising effect evaluation, and provides a stable data foundation for subsequent strategy adjustments.

[0017] By effectively screening human targets for exposure, associating anonymous trajectories, and processing repeated passages, this application aims to improve the authenticity of exposure data and reduce inflated statistical results. In existing technologies, the common practice is to directly count the number of people passing near the advertising device, entering the screen, or briefly stopping in the exposure count. This method struggles to distinguish between human targets who actually receive the advertising content and those merely passing near the device, and it also fails to eliminate data overlap caused by the same human target repeatedly passing by within a short period. This application filters human targets within the advertising's visible coverage area that meet conditions regarding head orientation, dwell time, distance, and occlusion as effective exposure targets. Furthermore, it applies attenuation weighting to repeated passages along the same anonymous trajectory, ensuring that the exposure count is not simply an additive result, but rather fragment-level effective exposure data that more closely reflects actual advertising contact. This improves the credibility of performance evaluation, enhances the quality of feedback samples, and reduces subsequent learning bias.

[0018] By establishing a local baseline window and matching comparison segments, the net attraction effect value, net interaction effect value, and net guidance effect value are extracted from the current delivery segment. Within a preset set of action constraints, the advertising delivery strategy is updated through feedback learning, achieving a direct closed-loop association between the advertising effect optimization process and the physical delivery actions that the robot can actually execute. Existing solutions often adjust parameters such as materials and time periods based on static advertising evaluation results or platform-level delivery data, which is difficult to apply to dynamic delivery scenarios of mobile robots in complex environments. This application organically combines reference value construction, net effect extraction, abnormal interference labeling, and confidence level determination. Furthermore, it updates parameters such as dwell point, robot orientation, advertising output unit orientation, broadcast volume, broadcast duration, and interaction trigger threshold under constraints. This allows the robot to continuously correct subsequent delivery strategies without deviating from the execution boundaries of the environment, thereby achieving a balance between evaluation accuracy, update stability, and on-site adaptability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 This is a schematic diagram of the site grid, candidate stopping points, advertising visible coverage area, and preset guidance area of ​​the advertising robot of the present invention in the target site.

[0021] Figure 3 This is a schematic diagram of the segmentation and data processing flow of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 refer to Figure 1-3 A method for evaluating and learning the effectiveness of robot advertising delivery, applied to an advertising delivery robot, the advertising delivery robot including a mobile mechanism, an advertising output unit, an environmental perception unit, an interaction detection unit, and a control unit, the method comprising the following steps: S1: The control unit divides the continuous advertising process into multiple delivery segments based on the preset time length, robot pose change amount, or advertising release action switching event, and records the scene state parameters and release action parameters corresponding to each delivery segment. S2: The environmental perception unit collects information on the relative position, head orientation, dwell trajectory, occlusion status, and distance between the human target and the advertising output unit within each delivery segment. The control unit filters valid exposure targets based on the following conditions: the human target's head area or torso center point is located within the advertising visible coverage area; the head orientation is pointing towards the advertising output unit; the continuous dwell time reaches a preset threshold; the distance is within a preset range; and the occlusion ratio is not greater than a preset threshold. S3: The control unit performs anonymous trajectory association based on the continuity of the effective exposure target's position, direction of motion, and timing of appearance in consecutive frames. It identifies the repeated passing behavior of the same target in the same delivery segment or adjacent delivery segments, and includes subsequent exposures in the segment-level effective exposure data according to a preset attenuation coefficient. S4: The interaction detection unit detects the interaction events generated by the effective exposure target in the current delivery segment. The control unit establishes a local baseline window based on the current delivery segment and selects matching comparison segments from historical delivery segments or adjacent time period segments that meet the preset similar conditions in terms of site grid location, time period category and pedestrian flow status, and whose release action parameters are different or have not executed the advertising release action. S5: The control unit compares the effective exposure data and interaction event data of the current delivery segment with the local baseline window and the matching control segment, extracts the net attraction effect value, net interaction effect value and net guidance effect value of the current delivery segment relative to natural pedestrian flow fluctuations and environmental changes, and generates feedback samples based on the net attraction effect value, net interaction effect value, net guidance effect value, scene state parameters, release action parameters, repeated processing results and abnormal interference markers that characterize abnormal fluctuations in pedestrian flow density, occlusion degree or environmental noise. S6: The control unit updates the advertising publishing strategy based on feedback samples within the preset action constraint set. When the confidence level of the feedback sample is lower than the preset threshold or the abnormal interference flag is triggered, the current advertising publishing strategy remains unchanged or switches to the preset basic publishing strategy. S7: The control unit controls the advertising robot to adjust the publishing action parameters in subsequent delivery segments based on the updated advertising publishing strategy, so as to form a closed-loop update process for advertising publishing effect evaluation and feedback learning.

[0024] The mobile mechanism drives the advertising robot to move along a preset cruising route within the target area and stop at preset candidate stopping points; the advertising output unit outputs image ads, video ads, and voice ads, and switches between ad playback mode and interactive mode under the control of the control unit; the environmental perception unit collects spatial position, head orientation, movement trajectory, distance information, occlusion information, and environmental status information of human targets around the robot; the interaction detection unit collects scanning events, touch events, voice response events, close-range dwelling events, and preset guidance area arrival events between human targets and the advertising output unit; the control unit completes the division of delivery segments, effective exposure target selection, anonymous trajectory association, repeated passage identification, local baseline window establishment, matching comparison segment selection, net effect value extraction, feedback sample generation, and advertising delivery strategy update.

[0025] This embodiment is applicable to scenarios with relatively clear boundaries and stable pedestrian flow paths, such as indoor commercial passageways, exhibition hall corridors, waiting areas, brand display areas, and information guidance areas. In these scenarios, the advertising robot typically needs to move between multiple stopping points and adjust its publishing action parameters based on the pedestrian flow, observation behavior, and interaction behavior at each stopping point. If a fixed advertising screen is still used, and the advertising effectiveness is evaluated directly based on the number of viewers, playback count, total number of scans, or facial expression classification results, it is easily affected by natural fluctuations in pedestrian flow, repeated passages, temporary obstructions, and changes in the robot's own position and orientation, making it difficult to truly distinguish the actual effect of the current publishing action parameters. Therefore, this embodiment divides the continuous advertising publishing process of the robot into multiple delivery segments, and completes effective exposure screening, repeated passage processing, reference value construction, net effect extraction, and constrained feedback learning at the delivery segment level, thereby establishing a one-to-one correspondence between the evaluation results and the specific publishing actions performed by the robot on-site.

[0026] To ensure a unified spatial reference for each segment, the control unit first accesses the robot's navigation map, dividing the target area into multiple site grids. Each site grid corresponds one-to-one with its position coordinates on the navigation map, and a grid index table is established in the control unit. Candidate stopping points are set on at least one site grid, and the robot performs the advertisement placement action after reaching the corresponding stopping point during operation. Through this method, subsequent aspects such as site grid location, stopping point switching, local baseline window establishment, matching segment selection, and strategy updates are all based on a unified map coordinate system, without using fuzzy region labels to replace actual locations.

[0027] The environmental perception unit includes a wide-angle camera module, a depth ranging module, a lighting acquisition module, a noise acquisition module, and a pose acquisition module. The wide-angle camera module outputs image frames of human targets within the currently displayed segment; the depth ranging module outputs distance data between the human target and the advertising output unit; the lighting acquisition module outputs the ambient light intensity within the current site grid; the noise acquisition module outputs the ambient noise intensity within the current site grid; and the pose acquisition module outputs the position coordinates, body orientation, and advertising output unit orientation of the advertising robot.

[0028] The interaction detection unit includes a QR code callback interface, a touch capture interface, a voice capture interface, and a guided area entry detection interface. The QR code callback interface records scanning events generated when a human target reads the QR code or electronic tag displayed by the advertising output unit; the touch capture interface records touch operations by a human target on the interactive interface of the advertising output unit; the voice capture interface records the human target's response voice to a preset wake-up word or preset response word; and the guided area entry detection interface records whether the same anonymous trajectory enters a preset guided area from the visible coverage area of ​​the advertisement.

[0029] It should be noted that in this embodiment, the environmental perception unit does not establish individual identity profiles, store facial templates, or use age tags, gender tags, or expression classification results as the basis for updating advertising strategies. The environmental perception unit only extracts anonymous information related to on-site behavior, including the head region location, torso center point location, head orientation, movement trajectory, dwell time, distance information, and occlusion information of the human target. Therefore, the feedback learning in this embodiment is based on the anonymous behavioral data and interaction data that the robot can continuously and stably acquire, rather than on individual identity information or emotion classification information.

[0030] In this embodiment, the control unit records scene state parameters and release action parameters within each delivery segment. Scene state parameters characterize the environmental state of the current delivery segment, while release action parameters characterize the release action actually performed by the robot in the current delivery segment.

[0031] The scene state parameters include the location of the site grid, time period, pedestrian density, pedestrian flow direction, ambient light intensity, ambient noise intensity, and occlusion level. The site grid location is determined by the site grid where the robot is currently stationary. The time period is divided into fixed time periods according to the system's pre-set operating schedule. Pedestrian density is determined by the number of anonymous trajectories crossing the current site grid boundary within a sliding time window and the current site grid area. The sliding time window uses the starting point of the currently deployed segment as a reference, and its window length is consistent with or a fixed proportion of the preset time length. Using sliding time window statistics instead of single-frame population statistics reduces the impact of instantaneous fluctuations on subsequent matching and comparison segment selection. The pedestrian flow direction is obtained by averaging the displacement direction vectors of each anonymous trajectory within the current sliding time window. Ambient light intensity is collected by the light acquisition module. Ambient noise intensity is collected by the noise acquisition module in a non-voice response state. The degree of occlusion is determined by the proportion of the head or torso area of ​​the human target that is occluded by other targets or obstacles. When there are multiple effective exposure targets in the current delivery segment, the average occlusion ratio or the maximum occlusion ratio can be used as the degree of occlusion of the current delivery segment.

[0032] The release action parameters include robot stopping point, robot orientation, advertising output unit orientation, advertising material identifier, broadcast volume, broadcast duration, and interaction trigger threshold. The robot stopping point is a pre-written dockable location in the navigation map. The robot orientation is the angle of the robot chassis relative to the navigation map reference direction. The advertising output unit orientation is the angle of the display surface normal or the speaker's main radiation direction relative to the navigation map reference direction. The advertising material identifier is the material number of the currently released segment. The broadcast volume is the audio output level of the advertising output unit. The broadcast duration is the continuous output duration of the current advertising material. The interaction trigger threshold is the quantity threshold used by the control unit when switching the advertising output unit from advertising playback mode to interactive mode.

[0033] When the current delivery segment is generated, the control unit writes the above-mentioned scene state parameters and release action parameters into the segment record table. When the current delivery segment ends, the effective exposure data, interaction event data, abnormal interference markers, net effect value and confidence level corresponding to the delivery segment are encapsulated into a segment data unit for subsequent reference value calculation and strategy update.

[0034] In this embodiment, the control unit divides the continuous advertising process into multiple segments based on a preset time length, robot pose change, or advertising release action switching event. The control unit terminates the current segment and generates the next segment when any of the following occurs: The duration of the currently displayed segment has reached the preset time length; The robot's pose change exceeds the preset pose change threshold; The advertising creative identifier has changed; The broadcast volume has changed; The broadcast duration setting has been changed; The advertising output unit switches between advertising playback mode and interactive mode.

[0035] The change in robot pose is determined by the positional offset between the current position and the starting position of the current segment, and the angular offset between the current robot orientation and the starting orientation of the current segment. The control unit records the starting position and the starting robot orientation at the beginning of the current segment delivery. In each subsequent sampling period, the current position is compared with the starting position, and the current robot orientation is compared with the starting robot orientation. When the positional offset reaches a preset displacement threshold, or the angular offset reaches a preset angle threshold, the control unit determines that the current pose has significantly deviated from the segment's starting state and terminates the current segment delivery.

[0036] Through the above method, the scene state parameters and release action parameters within each release segment remain relatively stable in a short period of time, so that the subsequent evaluation object is no longer the daily patrol task or the entire playback log, but an actual release action performed by the robot at a specific stop point, a specific orientation, a specific material, a specific volume, and a specific mode.

[0037] Within each delivery segment, the environmental perception unit continuously outputs a set of human targets, from which the control unit filters for valid exposure targets. To avoid misclassifying human targets that merely pass by the robot without actually touching the advertising content as exposure targets, in this embodiment, valid exposure targets must simultaneously meet the following conditions: The head area or center point of the torso of the human target is located within the visible coverage area of ​​the advertisement.

[0038] The angle between the head orientation of the human target and the normal of the display surface of the advertising output unit is less than or equal to the first angle threshold.

[0039] The duration of continuous dwell time of a human target within the current delivery segment is greater than or equal to the first time threshold.

[0040] The distance between the human target and the advertising output unit is between the first distance threshold and the second distance threshold.

[0041] The occlusion ratio corresponding to the human target is less than or equal to the first occlusion threshold.

[0042] The control unit first determines the visible coverage area of ​​the advertisement within the current site grid and its adjacent areas based on the installation height of the advertisement output unit, the size of the display surface, the normal of the display surface, the display field of view, and a first distance threshold and a second distance threshold. This visible coverage area is represented as a closed polygonal region on the navigation map and rotates synchronously as the orientation of the advertisement output unit changes.

[0043] Subsequently, the control unit performs filtering in a fixed order. First, it determines whether the head region or torso center point of the human target is within the visible coverage area of ​​the advertisement. Next, it determines whether the angle between the head orientation and the normal of the advertisement output unit's display surface meets the requirements. Then, it determines whether the continuous dwell time meets the requirements. Next, it determines whether the distance is within a preset range. Finally, it determines whether the occlusion ratio is less than or equal to a preset threshold. Only when all the above conditions are met will the control unit include the human target in the set of valid exposure targets.

[0044] The reason for using head orientation rather than movement direction as the basis for observation is that in real-world scenarios, a human target may be walking forward while turning their head to observe the advertisement. The reason for using continuous dwell time rather than instantaneous appearance as the contact criterion is to avoid misjudging a human target that momentarily crosses the field of view as having received the advertisement content. The reason for using dual-sided distance thresholds is to simultaneously eliminate ineffective observation situations caused by distances that are too close or too far.

[0045] After obtaining the valid exposure targets within the current delivery segment, the control unit performs anonymous trajectory association on the valid exposure targets across consecutive frames. The anonymous trajectory is only used to represent the continuity of the behavior of the same human target within a short time range and is not used to identify its identity.

[0046] The control unit jointly determines the positional deviation, motion direction deviation, and occurrence time interval between the human target in the current frame and the human target in the previous frame. When the positional deviation is less than or equal to a first displacement threshold, the motion direction angle is less than or equal to a second angle threshold, and the occurrence time interval is less than or equal to a second time threshold, the corresponding human target is associated with the same anonymous trajectory. If multiple candidate human targets simultaneously meet the conditions, the candidate target with the smallest positional deviation and the smallest direction deviation is selected for association. If no subsequent target meeting the conditions is found within a preset time, the anonymous trajectory is considered to have ended.

[0047] After an anonymous trajectory is established, the control unit continues to track the anonymous trajectory within the current delivery segment and its adjacent delivery segments. When the same anonymous trajectory enters the visible coverage area of ​​the advertisement multiple times within the same delivery segment or adjacent delivery segments, and the time interval between two consecutive entries is less than or equal to a third time threshold, the control unit determines it as a repeated traversal behavior.

[0048] For repeated traversal behaviors, this embodiment does not accumulate them as multiple independent exposure samples, but uses an attenuation weighting method. The first exposure is counted as one effective exposure in the segment-level effective exposure data, and each subsequent exposure is counted as segment-level effective exposure data according to a preset attenuation coefficient. The reason for this setting is that when the same human target passes through the advertising area multiple times in a short period of time, although it constitutes multiple entries, its additional contribution to the advertising dissemination effect is usually less than the contribution of the first contact. If it is still accumulated with equal weights, it is easy to exaggerate the attraction effect of the robot at that stop point.

[0049] In this embodiment, the repeated traversal processing result includes at least the following: the number of anonymous tracks within the currently delivered segment that are determined to be repeated traversal behaviors, the number of repeated traversals corresponding to each anonymous track, and the segment-level effective exposure number after attenuation weighting. The control unit writes the above results into the feedback sample for subsequent net effect extraction and release strategy updates.

[0050] The interaction detection unit is used to collect QR code scanning events, voice response events, touch events, near-field dwell events, and preset guidance area arrival events.

[0051] The QR code scanning event is generated by the QR code callback interface when it detects that the terminal device is reading the QR code or electronic tag. The voice response event is generated by the voice acquisition interface after it has collected the preset response words. The touch event is generated by the touch acquisition interface when it detects that a human target is touching the interactive interface. The near-field dwell event is determined by the control unit based on the distance between the human target and the advertising output unit and the duration of dwell. The preset guidance area arrival event is generated based on an anonymous trajectory entering the preset guidance area from the advertising's visible coverage area.

[0052] In this embodiment, the interaction trigger threshold directly participates in the switching from ad playback mode to interactive mode. The control unit continuously counts the number of interactive candidate targets within the current delivery segment. When the number of interactive candidate targets is greater than or equal to the interaction trigger threshold, the ad output unit is controlled to switch from ad playback mode to interactive mode. The interactive candidate target is a human target that simultaneously meets the following conditions: it has been determined as a valid exposure target, the distance between the human target and the ad output unit is less than or equal to a third distance threshold, and the continuous dwell time is greater than or equal to a fourth time threshold. Therefore, the interaction trigger threshold is not merely a recorded parameter, but an executable control parameter directly related to mode switching.

[0053] In interactive mode, the ad output unit can switch the original ad playback page to an interactive page, or overlay consultation prompts, discount entry points, QR code guidance, or voice Q&A prompts on the ad playback page. The specific content of the interactive page can be configured according to project requirements, but the triggering time of the interactive mode is determined by the comparison between the number of interactive candidate targets and the interactive triggering threshold.

[0054] In order to separate the effects of the current release action itself from natural fluctuations in pedestrian traffic, the control unit simultaneously establishes a local baseline window and a matching control segment within each release segment.

[0055] The local baseline window consists of two sub-windows at the stop point corresponding to the current delivery segment. The first sub-window is located before the start of the current delivery segment, when the robot has already reached the same stop point, and both the robot's orientation and the ad output unit's orientation are stable, but the current target ad creative has not yet started outputting. The second sub-window is located after the end of the current delivery segment, when the robot has not yet left the same stop point, but the output of the current target ad creative has stopped. The control unit counts the number of effective impressions, the number of interaction events, and the number of visits to the preset guidance area within these two sub-windows, and merges the two to form the local baseline window data. The local baseline window is used to reflect the natural passage when the robot is not performing the current ad delivery action at the same stop point and in the same orientation.

[0056] In addition to the local baseline window, the control unit also filters matching reference segments from historical delivery segments and adjacent time segment segments. Matching reference segments must simultaneously meet the following conditions: the site grid location is the same, the time period belongs to the same time period category, the difference in pedestrian density is less than or equal to the first density threshold, the deviation in pedestrian traffic direction is less than or equal to the third angle threshold, the difference in ambient light intensity is less than or equal to the first light threshold, the difference in ambient noise intensity is less than or equal to the first noise threshold, and at least one of the publishing action parameters is different from the currently delivered segment, or the corresponding segment has not executed the advertising publishing action.

[0057] The requirement that the matched comparison segment and the currently displayed segment be in the same site grid and the same time period category is to ensure that they are under basically consistent spatial and temporal conditions; the requirement that the differences in pedestrian density, pedestrian traffic direction, ambient light intensity, and ambient noise intensity be within a preset range is to ensure that the scene conditions of the two segments are basically similar; and the requirement that at least one of the publishing action parameters be different or that no advertising publishing action be performed is to ensure that the reference object can reflect the natural baseline or approximate baseline under different behavioral configurations.

[0058] After obtaining the number of effective exposures, interaction events, and guided arrivals for the current delivery segment, the control unit further extracts the net attraction effect value, net interaction effect value, and net guided effect value. The current values ​​are not directly used as the final effect values; instead, they need to be compared with reference values ​​formed by the local baseline window and matched control segments.

[0059] To ensure that the weighting method for the matched control fragments has a clear implementation basis, the control unit first calculates the first... The overall difference between each matched control segment and the currently delivered segment :

[0060] In the formula, For the first The overall difference value between each matched control segment and the currently delivered segment; For the first The time difference between each matched control segment and the currently delivered segment; The current crowd density for the segment being shown; For the first Crowd density in a matched control segment; For the first The deviation of pedestrian flow direction between the matched control segment and the currently displayed segment; and These are the currently displayed segment and the segment number. The ambient light intensity of the matched control fragments; and These are the currently displayed segment and the segment number. The ambient noise intensity of each matched control segment; and These are the currently displayed segment and the segment number. The degree of occlusion of each matched control fragment; to These are the weighting coefficients corresponding to each difference item, and all are preset values ​​greater than 0.

[0061] The control unit determines the weight of each matched control segment in the calculation of the reference value based on the magnitude of the comprehensive difference value. The smaller the comprehensive difference value, the closer the corresponding matched control segment is to the currently delivered segment, and the greater its contribution to the reference value; the larger the comprehensive difference value, the smaller its contribution to the reference value.

[0062] Subsequently, the control unit weights the local baseline window data and the data of each matched control segment according to a unified rule to obtain reference values ​​for the current segment in three dimensions: effective exposure, interactive events, and guided arrival. Then, the exposure reference value is subtracted from the number of segment-level effective exposures of the current segment, the interaction reference value is subtracted from the number of interactive events of the current segment, and the guided arrival value of the current segment is subtracted from the guided arrival reference value to obtain the net attraction effect value, net interaction effect value, and net guided effect value.

[0063] Through the above method, the effect value obtained in this embodiment is not simply the number of viewers, the number of QR codes scanned, or the number of arrivals, but rather the net incremental result of the current published action parameters relative to natural fluctuations in pedestrian flow and environmental changes. This processing method can avoid misrepresenting the natural heat of the dwell point itself, changes in peak and trough pedestrian flow, or changes in the local environment as the effect of the robot's published action itself.

[0064] After extracting the net attraction effect value, net interaction effect value, and net guidance effect value, the control unit generates a feedback sample for the current delivery segment. The feedback sample includes scene state parameters, release action parameters, net attraction effect value, net interaction effect value, net guidance effect value, repeated pass-through results, abnormal interference markers, and confidence level.

[0065] An abnormal interference flag is triggered by at least one of the following conditions: the pedestrian density exceeds a second density threshold within a preset time; the degree of obstruction exceeds a second obstruction threshold; the environmental noise intensity exceeds a second noise threshold; or the duration of obstruction within the visible coverage area of ​​the advertisement is greater than or equal to a fifth time threshold. When an abnormal interference flag is triggered, the control unit may still retain the record of the current segment, but will not use that segment as a high-confidence sample to directly update the publishing strategy.

[0066] Confidence score is used to characterize whether the current feedback sample is suitable for participating in strategy updates. The control unit determines the confidence score of the current feedback sample based on the number of matching comparison segments, the degree of scene difference between the currently delivered segment and the matching comparison segments, and the completeness of interaction event data. If the number of matching comparison segments is too small, the scene difference is too large, or there are issues such as missing image frames, interrupted pose data, or lost interaction callbacks in the currently delivered segment, the confidence score of the current feedback sample will decrease. Confidence score is not an evaluation value for advertising effectiveness, but rather an evaluation value for sample reliability.

[0067] In this embodiment, the control unit does not perform arbitrary searches within the open action space, but updates the advertising publishing strategy within a preset set of action constraints. The preset set of action constraints includes only the robot's stopping point, robot orientation, advertising output unit orientation, advertising material identifier, broadcast volume, broadcast duration, and interaction trigger threshold.

[0068] Specifically, the adjustment range of the robot's stopping point is limited to the current site grid or the site grid adjacent to the current site grid; the adjustment range of the robot's orientation is limited to a preset angle range on both sides of the current orientation; the adjustment range of the advertising output unit's orientation is limited to a preset angle range on both sides of the current orientation; the broadcast volume is limited to a preset volume range; the broadcast duration is limited to a preset duration range; the interaction trigger threshold is limited to a preset threshold range; and the advertising material identifier is only switched in the material candidate library corresponding to the current scene.

[0069] The control unit maintains the action score by issuing action combinations for each candidate in the action constraint set. For the... The candidate release action combination, in the first The action rating after the latest update Determine using the following formula:

[0070] In the formula, For the first Before the next update Action scores for each candidate release action combination; For the first After the first update Action scores for each candidate release action combination; To update the step size, and ; This represents the confidence level of the feedback sample corresponding to the currently delivered segment. This represents the net attraction effect value of the currently displayed segment; This represents the net engagement value of the currently displayed segment; This represents the net engagement effect of the currently delivered segment; , and These are the weighting coefficients corresponding to the net attraction effect value, net interaction effect value, and net guidance effect value, respectively, and all are preset values ​​greater than 0.

[0071] Based on the updated action scores, the control unit selects candidate action combinations with higher scores from a preset action constraint set that satisfy the current site grid constraints, orientation constraints, and volume constraints as the advertising delivery strategy for subsequent segments. When the confidence level of the current feedback sample is lower than a preset threshold or an abnormal interference flag is triggered, the control unit does not perform the above score update, but maintains the current advertising delivery strategy or switches to a preset basic delivery strategy. The preset basic delivery strategy consists of a combination of preset basic parameters, including default dwell point, default machine orientation, default advertising output unit orientation, default advertising material identifier, default broadcast volume, default broadcast duration, and default interaction trigger threshold.

[0072] It should be noted that the control unit is not required to immediately switch to the highest-scoring action combination after each segment is delivered. Instead, execution rules such as the minimum number of stable running segments, the minimum score difference, and the maximum switching frequency can be set. This avoids the robot frequently adjusting its stopping point, orientation, or volume due to a single, accidental high-value sample, ensuring the entire system maintains stable operation in real-world scenarios.

[0073] In this embodiment, a complete execution process of the advertising delivery robot is as follows: After the advertising robot enters the target area grid according to its cruise route and stops at its designated stop point, the control unit generates the current delivery segment and writes the current scene state parameters and delivery action parameters. The environmental perception unit continuously collects images of human targets, distance data, lighting data, noise data, and pose data. The control unit sequentially filters human targets by location, orientation, dwell time, distance, and occlusion to form a set of effective exposure targets for the current delivery segment. Subsequently, the control unit performs anonymous trajectory association on effective exposure targets between consecutive frames, identifies repeated passing behaviors, and generates segment-level effective exposure data and repeated passing processing results. Simultaneously, the interaction detection unit continuously collects QR code scanning events, voice response events, touch events, near-field dwelling events, and preset guidance area arrival events. The control unit establishes a local baseline window at the location corresponding to the current stop point and filters matching comparison segments from historical segments and adjacent time period segments, forming a reference value based on the time and scene differences between the current segment and the matching comparison segment. Then, the control unit extracts net attraction effect value, net interaction effect value, and net guidance effect value, generates feedback samples, and decides whether to perform action score updates based on confidence level and abnormal interference markers. Finally, the control unit, based on the updated advertising delivery strategy, controls the advertising delivery robot to adjust the stopping point, robot orientation, advertising output unit orientation, advertising material identification, broadcast volume, broadcast duration, and interaction trigger threshold in subsequent delivery segments, and continues to perform the next round of data collection, net effect extraction, and strategy update.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating and learning feedback on the effectiveness of robot advertising, characterized in that, An advertising delivery robot, comprising a mobile mechanism, an advertising output unit, an environmental perception unit, an interaction detection unit, and a control unit, is applied to an advertising delivery robot. The method includes the following steps: S1: The control unit divides the continuous advertising process into multiple delivery segments based on the preset time length, robot pose change amount, or advertising release action switching event, and records the scene state parameters and release action parameters corresponding to each delivery segment. S2: The environmental perception unit collects information on the relative position, head orientation, dwell trajectory, occlusion status, and distance between the human target and the advertising output unit within each delivery segment. The control unit filters valid exposure targets based on the following conditions: the human target's head area or torso center point is located within the advertising visible coverage area; the head orientation is pointing towards the advertising output unit; the continuous dwell time reaches a preset threshold; the distance is within a preset range; and the occlusion ratio is not greater than a preset threshold. S3: The control unit performs anonymous trajectory association based on the positional continuity, motion direction continuity and occurrence sequence continuity of the effective exposure target in consecutive frames, identifies the repeated passing behavior of the same target in the same delivery segment or adjacent delivery segments, and includes subsequent exposures in the segment-level effective exposure data according to a preset attenuation coefficient. S4: The interaction detection unit detects the interaction events generated by the effective exposure target in the current delivery segment. The control unit establishes a local baseline window based on the current delivery segment and selects matching comparison segments from historical delivery segments or adjacent time period segments that meet the preset similar conditions in terms of site grid location, time period category and pedestrian flow status, and whose release action parameters are different or have not executed the advertising release action. S5: The control unit compares the effective exposure data and interaction event data of the current delivery segment with the local baseline window and the matching control segment, extracts the net attraction effect value, net interaction effect value and net guidance effect value of the current delivery segment relative to natural pedestrian flow fluctuations and environmental changes, and generates feedback samples based on the net attraction effect value, net interaction effect value, net guidance effect value, scene state parameters, release action parameters, repeated processing results and abnormal interference markers that characterize abnormal fluctuations in pedestrian flow density, occlusion degree or environmental noise. S6: The control unit updates the advertising publishing strategy based on feedback samples within the preset action constraint set. When the confidence level of the feedback sample is lower than the preset threshold or the abnormal interference flag is triggered, the current advertising publishing strategy remains unchanged or switches to the preset basic publishing strategy. S7: The control unit controls the advertising robot to adjust the publishing action parameters in subsequent delivery segments based on the updated advertising publishing strategy, so as to form a closed-loop update process for advertising publishing effect evaluation and feedback learning.

2. The method for evaluating and learning the effectiveness of robot advertising as described in claim 1, characterized in that, The scene state parameters include site grid location, time period, pedestrian density, pedestrian traffic direction, ambient light intensity, ambient noise intensity, and degree of occlusion. The parameters for the release action include the robot's stopping point, robot orientation, advertising output unit orientation, advertising material identifier, broadcast volume, broadcast duration, and interaction trigger threshold.

3. The method for evaluating and learning the effectiveness of robot advertising as described in claim 2, characterized in that, When the control unit filters valid exposure targets, it simultaneously meets the following conditions: the head area or torso center point of the human target is located within the visible coverage area of ​​the advertisement. The angle between the head orientation of the human target and the normal of the display surface of the advertising output unit is less than or equal to the first angle threshold. The continuous dwell time of a human target within the current delivery segment is greater than or equal to the first time threshold; The distance between the human target and the advertising output unit is between a first distance threshold and a second distance threshold; The occlusion ratio corresponding to the human target is less than or equal to the first occlusion threshold.

4. The method for evaluating and learning the effectiveness of robot advertising as described in claim 3, characterized in that, When the control unit performs anonymous trajectory association, it jointly determines the positional deviation, motion direction deviation, and occurrence time interval between the human target in the current frame and the human target in the previous frame. When the positional deviation is less than or equal to the first displacement threshold, the angle of the movement direction is less than or equal to the second angle threshold, and the time interval is less than or equal to the second time threshold, the corresponding human target will be identified as the same anonymous trajectory.

5. The method for evaluating and learning the effectiveness of robot advertising as described in claim 4, characterized in that, When the control unit identifies repeated passing behavior, if the same anonymous trajectory enters the visible coverage area of ​​the advertisement multiple times in the same or adjacent delivery segments, and the time interval between two adjacent entries is less than or equal to the third time threshold, it is determined to be repeated passing behavior. The first exposure is counted as a valid exposure in the segment-level valid exposure data, and each subsequent exposure is counted as a segment-level valid exposure data according to a preset attenuation coefficient, which is greater than 0 and less than 1.

6. The method for evaluating and learning the effectiveness of robot advertising as described in claim 5, characterized in that, The interactive events include QR code scanning events, voice response events, touch events, near-field dwell events, and preset guidance area arrival events; The scanning event is an event formed when a human target reads the QR code or electronic tag displayed by the advertising output unit. The voice response event is an event formed when the interaction detection unit collects the human target's response voice and identifies the preset response word. The touch event is an event formed when a human target touches the interactive interface of the advertising output unit. The near-distance dwell event is an event formed when the distance between the human target and the advertising output unit is less than or equal to a third distance threshold and the duration is greater than or equal to a fourth time threshold. The preset guidance area arrival event is an event formed when the human target moves from the visible coverage area of ​​the advertisement to the preset guidance area.

7. The method for evaluating and learning the effectiveness of robot advertising as described in claim 6, characterized in that, When the control unit filters the matching comparison segment, it uses the current delivery segment as a benchmark and selects segments from historical delivery segments or segments from adjacent time periods that simultaneously meet the following conditions as matching comparison segments: the site grid positions are the same; The time periods belong to the same time period category; The difference in pedestrian density is less than or equal to the first density threshold; The deviation of the pedestrian flow direction is less than or equal to the third angle threshold; The difference in ambient light intensity is less than or equal to the first light threshold. The difference in environmental noise intensity is less than or equal to the first noise threshold; Furthermore, at least one of the parameters for the published action is different from that of the currently delivered segment, or the corresponding segment has not executed the ad publishing action.

8. The method for evaluating and learning the effectiveness of robot advertising as described in claim 7, characterized in that, When the control unit extracts the net attraction effect value, net interaction effect value, and net guidance effect value, it uses the effective exposure number, interaction event number, and preset guidance area arrival number of the current delivery segment as the current value, and the weighted average value calculated by the local baseline window and the corresponding data of the matching control segment according to the preset weighting rule as the reference value. The net attraction effect value, net interaction effect value, and net guidance effect value are obtained by subtracting the reference value from the current value. The preset weighting rule determines the weight of each matched comparison segment based on the segment time difference and the scene state parameter difference.

9. The method for evaluating and learning the effectiveness of robot advertising as described in claim 8, characterized in that, The feedback samples include scene state parameters, release action parameters, net attraction effect value, net interaction effect value, net guidance effect value, repeated processing results, abnormal interference markers, and confidence level. The abnormal interference flag is triggered by at least one of the following conditions: the pedestrian density exceeds the second density threshold within a preset time, the degree of obstruction exceeds the second obstruction threshold, the environmental noise intensity exceeds the second noise threshold, and the duration of obstruction within the visible coverage area of ​​the advertisement is greater than or equal to the fifth time threshold. The confidence level is determined based on the number of matching comparison segments, the difference in scene state parameters between the currently delivered segment and the matching comparison segments, and the completeness of the interaction event data.

10. The method for evaluating and learning the effectiveness of robot advertising as described in claim 9, characterized in that, When the control unit updates the advertising publishing strategy within the preset action constraint set, it only adjusts the robot's stopping point, robot orientation, advertising output unit orientation, advertising material identifier, broadcast volume, broadcast duration, and interaction trigger threshold. The adjustment range of the robot's stopping point is limited to the current field grid or the field grid adjacent to the current field grid. The adjustment range of the robot orientation is limited to the preset angle range on both sides of the current orientation. The adjustment range of the advertising output unit orientation is limited to the preset angle range on both sides of the current orientation. The broadcast volume is limited to a preset volume range. The broadcast duration is limited to a preset duration range. The interaction trigger threshold is limited to a preset threshold range. When the confidence level of the feedback sample is lower than the preset threshold or an abnormal interference flag is triggered, the control unit keeps the current release action parameters unchanged or adjusts the release action parameters to a preset combination of basic parameters.

Images