An ar interactive garbage classification game system and method

Abstract

The application discloses an AR interactive garbage classification game system and method, the method comprises the following steps: calling a camera to capture an environment image, using a plane detection algorithm to identify a target plane area and rendering a virtual garbage can to generate an AR scene; dynamically generating a challenge task list; using a space hash partitioning technology to divide a three-dimensional grid, constructing a sphere bounding box for an object and an axis-aligned bounding box for a garbage can, and only testing the intersection of objects in the same or adjacent grid during collision detection; responding to a touch operation, when dragging, the object position moves following the touch point, when releasing, gravity, air resistance and collision are applied and the motion state is updated using explicit Euler integration; judging the dropping result in real time according to a garbage classification rule base and triggering a multi-modal feedback; calculating the accuracy rate and generating settlement data. The application improves the stability of the AR scene and the physical simulation performance of the mobile terminal, realizes personalized task generation, and effectively enhances the interest and effect of garbage classification teaching.

Description

Technical Field

[0001] This invention relates to the fields of augmented reality and computer data processing technology, specifically to an AR interactive garbage sorting challenge game system and method. Background Technology

[0002] With the full implementation of waste sorting policies, the demand for public knowledge dissemination and education is becoming increasingly urgent. Traditional waste sorting teaching and assessment methods mainly rely on static graphic and textual promotions, video tutorials, or physical models, which have the following limitations: First, the teaching methods are monotonous, resulting in low user participation. Traditional text-based and video-based teaching lack interactivity, leaving learners in a passive state and failing to stimulate their interest, especially among teenagers, where the learning effect is limited. While physical model classification offers some hands-on experience, the limited number and variety of teaching aids make it difficult to cover the diverse scenarios of waste sorting.

[0003] Second, the hardware costs are high, and the devices are not portable, making large-scale promotion difficult. Some simulated teaching aids that use physical cards or dedicated readers (such as RFID) suffer from high hardware costs, cumbersome preparation, and inconvenience in carrying them, which limits users' ability to learn anytime and anywhere and makes it difficult to achieve large-scale popularization.

[0004] Third, feedback is delayed, limiting teaching effectiveness. Traditional assessment methods mostly involve post-assessment grading or manual scoring, which cannot provide immediate and vivid feedback on correct and incorrect answers during the teaching process. This leads to missed opportunities for error correction and reinforcement, reducing learning efficiency.

[0005] Fourth, the teaching scenarios are detached from reality, making knowledge transfer difficult. The learning process is disconnected from real-life scenarios of waste generation and disposal, making it difficult for learners to apply the knowledge to real-life situations.

[0006] To address the aforementioned issues, some augmented reality (AR)-based solutions for assisting with waste sorting have emerged in the existing technology. However, these existing technologies primarily focus on information prompts and virtual character interactions, making it difficult to effectively solve the unique technical challenges faced in teaching waste sorting in mobile AR scenarios: firstly, how to stably and accurately detect planes suitable for placing virtual objects in environments with sparse textures and varying lighting; secondly, how to dynamically adjust learning content based on the user's cognitive state to achieve personalized and adaptive learning paths; and thirdly, how to balance the realism of physical interaction with real-time smoothness through lightweight design under the constraints of computing resources and power consumption on mobile terminals.

[0007] Therefore, there is an urgent need for a waste sorting popularization solution that is highly engaging, interactive, low-cost, can be implemented anytime and anywhere, and can provide immersive scenario-based teaching, in order to solve the technical challenges in the aforementioned mobile AR scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide an AR interactive garbage sorting challenge game system and method to solve the problems of boring, costly, slow feedback and disconnected scenarios in traditional garbage sorting teaching methods mentioned in the background art. In particular, it overcomes the technical difficulties faced by existing AR solutions when applied to mobile devices, such as poor stability of planar detection, lack of personalization in task generation, and difficulty in balancing the realism and real-time smoothness of physical interaction under the limited computing resources of mobile terminals.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an AR interactive garbage sorting challenge game method, applied to a mobile terminal, the method comprising: In response to the scene activation command, the camera is invoked to capture real-world image data, a plane detection algorithm is used to identify the target plane region, and a virtual sorting trash can model is superimposed and rendered on the spatial anchor point of the target plane region to generate an AR interactive scene. In response to the user's start command, a list of challenge tasks containing multiple virtual waste items is dynamically generated from the waste sorting item question bank; In the AR interactive scene, spatial hash partitioning technology is used to divide the interactive area into a three-dimensional grid. A spherical bounding box is constructed for each virtual trash item, and an axis-aligned bounding box is constructed for each virtual trash can. When performing collision detection, only the bounding box intersection test is performed on virtual trash items and virtual trash cans located in the same grid or adjacent grids. In response to the user's touch operation on the virtual trash item, when the user touches and drags, the position of the virtual trash item is controlled to follow the projection of the touch point on the target plane area; when the user releases the virtual trash item, gravity, air resistance and elastic force when colliding with the virtual trash can are applied to the virtual trash item, and the speed and position of the item are updated using explicit Euler integral with a fixed step size. The system performs real-time logical judgment on the results of the waste sorting interaction based on the waste sorting rule base, and triggers multimodal feedback including visual, auditory and tactile channels based on the judgment results; Record and display game progress, and after the task list is completed, calculate the classification accuracy based on the number of correctly classified tasks and the total number of tasks, and generate settlement data.

[0010] Preferably, the step of identifying the target planar region using a planar detection algorithm specifically includes: For each frame of the real environment image data, an image pyramid is constructed. Multi-scale texture features and edge features are extracted based on the image pyramid and fused through a feature pyramid network (FPN) to generate a planar candidate region heatmap. The camera pose is estimated in real time using a visual inertial odometry system, and the consistency of candidate planar regions in a series of consecutive frames is checked: the detection window length is set to a preset number of consecutive frames, and when the same candidate planar region is detected a preset threshold number of times within the detection window, the region is identified as the target planar region and its spatial anchor point is locked.

[0011] More preferably, the generation of AR interactive scene further includes: after detecting a partial planar region, using a semantic segmentation model to infer the semantic category of the region, and combining the homography matrix of the detected plane to extrapolate the plane extension range outward, so as to generate a virtual interactive region larger than the actual detected area.

[0012] Preferably, the dynamic generation of the challenge task list containing multiple virtual junk items specifically includes: The built-in Bayesian knowledge tracing model establishes and maintains a knowledge mastery vector for each user. The knowledge mastery vector contains multiple subcategories, and each subcategory corresponds to a value representing the probability that the user has mastered that subcategory. Based on the correctness of each classification operation, the mastery probability and forgetting parameters of the corresponding sub-category in the knowledge mastery vector are updated in real time. Based on the knowledge mastery vector, the Thompson sampling strategy is used to dynamically select the next virtual trash item to be disposed of from the question bank. The selection process comprehensively considers the following three factors: First factor: Virtual junk items are randomly selected from subcategories where the user's current mastery probability is below a first threshold, with an exploration probability ε. The exploration probability ε gradually decreases as the preset learning process parameter increases. The second factor is to select virtual junk items from subcategories whose probability of being currently known by the user is lower than the second threshold, based on Thompson's sampling results, using a probability of 1-ε, where the second threshold is greater than the first threshold. The third factor: For subcategories that the user has mastered with a probability higher than the second threshold, monitoring is conducted according to the review interval set by the Ebbinghaus forgetting curve. After the interval expires, virtual junk items for that subcategory are automatically inserted as review tasks.

[0013] Furthermore, the method of dynamically selecting the next virtual trash item to be disposed of from the question bank using the Thompson sampling strategy specifically includes: Maintain a Beta distribution model for each subcategory in the waste sorting item question bank. The Beta distribution model includes a success parameter α and a failure parameter β. When a user correctly categorizes a virtual junk item belonging to a specific subcategory, the success parameter α corresponding to that subcategory is incremented by 1. When a user misclassifies a virtual junk item belonging to a specific subcategory, the failure parameter β corresponding to that subcategory is incremented by 1. When selecting the next virtual trash item to be disposed of, a random probability value is sampled from the Beta distribution model corresponding to each subcategory, and the item in the subcategory with the highest sampled probability value is selected as the next virtual trash item to be disposed of.

[0014] Furthermore, the method also includes sequence optimization of the dynamic generation process: The process of generating task sequences is modeled as a Markov decision process, where the state is the user's mastery vector, the action is selecting the next item category, and the reward function is the expected improvement in knowledge retention rate. The optimal item placement order is calculated and output using either offline reinforcement learning or an online Monte Carlo tree search algorithm. If the online Monte Carlo tree search algorithm is used, the following steps are included: constructing a search tree with the current mastery degree vector as the root node, simulating node state transitions using a Bayesian knowledge tracing model, updating node values ​​through backpropagation based on the accumulated rewards obtained from the simulation, and finally selecting the item category corresponding to the child node with the highest value as the next virtual trash item to be deployed.

[0015] Preferably, the step of using spatial hash partitioning technology to divide the interactive area into a three-dimensional grid, and performing bounding box intersection tests only on virtual trash items and virtual trash cans located in the same or adjacent grids, specifically includes: According to the preset grid size, the interactive area in the AR interactive scene is divided into a uniform three-dimensional grid, and a unique hash index is assigned to each grid. In each physical simulation frame, based on the current spatial location of each virtual trash item and each virtual trash can, the hash index of the grid in which it is located is calculated, and the identifiers of the virtual trash items and the virtual trash cans are registered to the object list of the corresponding grid. For each virtual junk item, obtain the hash index of the grid it belongs to, and determine the hash index of the grids adjacent to it; Intersection tests of spherical bounding boxes and axis-aligned bounding boxes are performed only on virtual trash cans located within the grid containing the virtual trash item and its adjacent grids. Specific steps include: Obtain the minimum and maximum values ​​of the axis-aligned bounding box on the X, Y, and Z coordinate axes; The coordinates of the center of the sphere bounding box are restricted to between the minimum and maximum values ​​on the X, Y, and Z axes, respectively, to obtain the point on the axis-aligned bounding box closest to the center of the sphere; Calculate the Euclidean distance between the center of the sphere and the nearest point to obtain the nearest distance; If the nearest distance is less than the radius of the sphere's bounding box, then they are considered to intersect. If the nearest distance is greater than or equal to the radius of the sphere's bounding box, then they are determined to be non-intersecting; For virtual trash cans that are not located in the grid of the virtual trash item or in its adjacent grids, skip the intersection test directly.

[0016] Preferably, the response to the user's touch operation on the virtual trash item specifically includes switching between kinematic control mode and dynamic simulation mode: When a user touches and drags a virtual trash item, the virtual trash item is placed in kinematic control mode. Physical force calculations are ignored, and its spatial position is directly updated to the projection coordinates of the touch point on the target plane area. The orientation of the virtual trash item is controlled so that the angle between its normal and the direction of the camera optical axis is kept at a preset value. When a user releases a virtual trash item, the virtual trash item is immediately switched from kinematic control mode to dynamic simulation mode. The physics engine is activated to calculate gravity, air resistance, and collision response force, and the step of updating the item's velocity and position using explicit Euler integrals with a fixed step size is executed.

[0017] More preferably, the method of updating the velocity and position of an item using an explicit Euler integral with a fixed step size specifically includes: In response to the user's release operation, the position coordinates and velocity vector of the virtual junk item at the moment of release are obtained as the initial state for the dynamics simulation; In each physical simulation frame, perform the following iterative steps: The net force acting on the virtual trash item is calculated. The net force includes: gravity, calculated based on a preset gravitational acceleration constant and mass parameters; air resistance, calculated based on the current velocity vector and a preset air resistance coefficient, with the direction opposite to the velocity direction; and collision response, when it is determined that the virtual trash item collides with the virtual trash can or the target planar area, the collision impulse is calculated based on the collision normal vector and the coefficient of restitution to correct the velocity vector. The acceleration of the current frame is calculated based on the resultant force, and the acceleration is equal to the ratio of the resultant force to the mass. The motion state is updated using the explicit Euler integral formula: the velocity vector of the current frame is added to the product of the acceleration and the fixed time step to obtain the velocity vector of the next frame; the position coordinates of the current frame are added to the product of the velocity vector of the current frame and the fixed time step to obtain the position coordinates of the next frame. The updated position coordinates and velocity vectors are used as the input state for the next physical simulation frame. The above steps are repeated until the virtual trash item stops moving or is dragged back by the user.

[0018] Preferably, triggering multimodal feedback including visual, auditory, and tactile channels based on the discrimination result includes: The visual channel is implemented in the following way: when a correct or incorrect feedback is triggered, GPU instantiation rendering technology is used to generate particle effects, and the initial position, initial velocity, and life cycle parameters of the particles are passed to the vertex shader. The GPU then performs particle position updates and rendering operations in parallel for each frame. The tactile channel is implemented by triggering the device's linear motor to emit a short vibration when the judgment result is incorrect. The auditory channel is implemented by playing a positive chord when the judgment result is correct and a low-frequency warning sound when the judgment result is incorrect.

[0019] More preferably, the method further includes: Real-time monitoring of current frame rate and device temperature; When the frame rate is detected to be lower than the first preset threshold or the device temperature is higher than the second preset threshold, performance optimization measures are automatically executed. These performance optimization measures include reducing the accuracy of the physical simulation step size, reducing the number of particle emissions, or reducing the rendering resolution ratio.

[0020] Preferably, the step of performing real-time logical judgment on the disposal interaction results based on the waste classification rule base specifically includes: A pre-set local waste sorting rule base is provided. The waste sorting rule base uses a key-value pair data structure to store the mapping relationship between item identifiers and classification categories, and a Bloom filter is configured as a pre-index for the waste sorting rule base. During the identification process, the existence of virtual junk items is first quickly verified using the Bloom filter: If the verification fails, it is determined that the item identifier does not exist in the local rule base, and the subsequent query process is terminated; If the verification passes, the category result corresponding to the item identifier is obtained through key-value pair hash lookup.

[0021] Preferably, the generation of settlement data further includes: displaying changes in the user's mastery of major categories of waste using a radar chart, and recommending targeted tasks based on changes in mastery.

[0022] Preferably, the method further includes: uploading all interaction data to a cloud server after compression, for optimizing the cognitive state model parameters, and supporting display in the friend leaderboard.

[0023] Secondly, the present invention provides an AR interactive garbage sorting challenge game system, including a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program, is configured to implement the method described in any one of the first aspects through the following modules: The scene construction module is used to: call the camera to capture real environment image data, use a plane detection algorithm to identify the target plane area, and overlay and render a virtual sorting trash can model on the spatial anchor point of the target plane area to generate an AR interactive scene. The task generation module is used to: dynamically generate a list of challenge tasks containing multiple virtual garbage items from the garbage sorting item question bank in response to the user's start command; The collision detection module is used to: divide the interactive area into a three-dimensional grid using spatial hash partitioning technology in the AR interactive scene, construct a spherical bounding box for each virtual trash item, construct an axis-aligned bounding box for each virtual trash can, and only perform bounding box intersection tests on virtual trash items and virtual trash cans located in the same grid or adjacent grids during collision detection. The interactive control module is used to: respond to the user's touch operation on the virtual trash item; when the user touches and drags, control the position of the virtual trash item to follow the projection of the touch point on the target plane area; when the user releases the virtual trash item, apply gravity, air resistance and elastic force when colliding with the virtual trash can to the virtual trash item, and update the speed and position of the item using explicit Euler integral with a fixed step size. The discrimination feedback module is used to: perform real-time logical discrimination on the disposal interaction results based on the garbage classification rule base, and trigger multimodal feedback including visual, auditory and tactile channels based on the discrimination results; The data management module is used to record and display game progress, and after the task list is completed, calculate the classification accuracy based on the number of correctly classified tasks and the total number of tasks, and generate settlement data.

[0024] Thirdly, the present invention provides a mobile terminal, characterized in that it includes a processor, a memory, a camera, and a display screen, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the AR interactive garbage sorting challenge game method as described in any of the first aspects.

[0025] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the AR interactive garbage sorting challenge game method as described in any of the first aspects.

[0026] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1) Improved stability and environmental adaptability of AR scene construction: This invention effectively suppresses the plane detection jitter problem caused by changes in lighting, sparse textures or dynamic interference in the real environment through a plane detection mechanism of multi-scale feature fusion and temporal consistency verification, ensuring the stable anchoring of virtual objects in the real scene and improving usability and immersion in different home environments.

[0027] 2) Personalized and adaptive learning task generation is achieved: Unlike the fixed question bank or simple random question generation of existing technologies, this invention dynamically adjusts the content and difficulty of tasks based on the user's cognitive state. Through a balanced exploration and utilization mechanism, it accurately identifies the user's weak knowledge points and pushes tasks in a targeted manner, which effectively improves the retention rate of garbage classification knowledge and teaching efficiency.

[0028] 3) Optimized physical interaction performance and smoothness on mobile devices: This invention uses lightweight physical simulation strategies such as spatial partitioning collision detection, kinematic / dynamic decoupling control, GPU instantiation rendering, and adaptive frame rate scheduling to significantly reduce the computational overhead of mobile devices while fully preserving realistic physical behaviors such as gravity, air resistance, and collision elasticity. It can maintain a high frame rate real-time interactive experience on low-end and mid-range terminals, solving the technical problem that traditional physical simulation is difficult to balance realism and real-time performance on mobile devices.

[0029] 4) Provides a low-latency, multimodal, and real-time feedback experience: This invention ensures the real-time nature of interactive judgments and the richness of feedback through a fast query mechanism of the local rule base and multi-channel feedback collaboration, enhancing the user's immersion and the immediacy of error correction effects, and avoiding interruptions in interactive response caused by query delays. Attached Figure Description

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating a preferred embodiment of the AR interactive garbage sorting challenge game method of the present invention; Figure 2 This is a structural block diagram of an AR interactive garbage sorting challenge game system according to a preferred embodiment of the present invention. Detailed Implementation

[0032] To make the above and other features and advantages of the present invention clearer, the invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for the purpose of explanation to those skilled in the art and are exemplary only, not restrictive.

[0033] Example 1: like Figure 1 As shown, this embodiment provides an AR interactive garbage sorting challenge game method. This method is applied to a mobile terminal, which includes, but is not limited to, smartphones, tablets, AR glasses, and other devices equipped with cameras, processors, graphics processors, memory, and displays. The following describes each step of the method in detail, using an example of running on a tablet.

[0034] I. Scene Launch and AR Interactive Scene Generation Step S1: In response to the scene activation command, the camera captures real-world image data, a planar detection algorithm is used to identify the target planar region for placing virtual objects, and a virtual sorting trash can model is overlaid and rendered on the spatial anchor points of the target planar region to generate an AR interactive scene. Specifically, this includes the following sub-steps: (1) Multi-scale feature extraction An image pyramid is constructed for each captured RGB image frame. For example, images at three scales—original image, 1 / 2 resolution image, and 1 / 4 resolution image—are extracted. Multi-scale texture features and edge features are extracted based on the image pyramid. In a preferred embodiment, the image at each scale is input into a lightweight convolutional neural network (such as MobileNetV3) to extract texture features, while the Sobel operator is combined to extract edge features. The multi-scale texture feature maps and edge feature maps are input into a feature pyramid network (FPN) for cross-scale fusion to generate a planar candidate region heatmap that is robust to changes in illumination and sparse texture environments.

[0035] (2) Timing consistency check To suppress planar detection jitter caused by factors such as instantaneous reflections and interference from dynamic objects, this embodiment introduces a temporal consistency verification mechanism. Specifically, visual inertial odometry (VIO) is used to estimate the camera pose (position and orientation) in each frame in real time, and candidate planar regions in different frames are mapped to a unified world coordinate system based on the camera pose. A lifetime counter is established for each candidate planar region to record the number of times it is detected in consecutive frames. The detection window length is set to N frames (N is 5 in this embodiment). When the same candidate planar region is detected in at least 3 out of 5 consecutive frames, that is, when the lifetime counter of the region reaches the threshold of 3, the region is confirmed as the target planar region, and its spatial anchor point is locked. This mechanism effectively suppresses false detections and jitter caused by instantaneous interference, ensuring the stable anchoring of virtual objects in the real scene.

[0036] (3) Adaptive expansion of planar boundaries In practical applications, a camera may only capture a portion of a planar area (e.g., only a corner of the floor). To obtain a larger virtual interactive area, this embodiment employs a planar boundary adaptive expansion technique. After detecting a partial planar area, a lightweight semantic segmentation model (such as the mobile version of DeepLabV3+) is used to infer the semantic category of that area, determining whether it belongs to "ground," "desktop," or another category. Combining the homography matrix of the detected plane, the planar extension range is extrapolated outwards, generating a virtual interactive area larger than the actual detected area. This technique allows the system to stably place multiple virtual trash cans even in scenes where only a small portion of the ground is exposed, improving the usability of the AR scene.

[0037] (4) Rendering of virtual trash can model After identifying and locking onto the target planar area, the system overlays and renders four virtual recycling bin models on the spatial anchor points of that plane, corresponding to kitchen waste, recyclables, hazardous waste, and other waste, respectively. The recycling bin models use physically based rendering (PBR) materials, which can reflect light in real time according to ambient lighting, enhancing the integration of virtual objects with the real scene. At this point, the AR interactive scene is complete.

[0038] II. Dynamic Generation of Challenge Task List Step S2: In response to the user's start command, dynamically generate a list of challenge tasks containing multiple virtual waste items from the waste sorting item question bank. This includes the following sub-steps: (1) Cognitive state modeling This embodiment uses a Bayesian Knowledge Tracking (BKT) model to model the user's cognitive state. Specifically, a knowledge mastery vector is established and maintained for each user. The dimensions of this vector correspond to the subcategories in the waste classification database (such as plastic bottles, batteries, leftover food, soiled paper towels, etc.), and the value of each dimension represents the user's mastery probability of that subcategory, with a value range of [0,1]. During model initialization, the mastery probability of all subcategories can be set to a preset value (e.g., 0.3).

[0039] Each time a user completes a classification operation, the model updates the mastery probability of the corresponding sub-category using Bayes' theorem, based on the correctness of the operation. Simultaneously, the model maintains a forgetting parameter for each sub-category to simulate potential knowledge loss over time. The update of the forgetting parameter is modeled based on the Ebbinghaus forgetting curve.

[0040] (2) Single-step decision-making: comprehensively consider exploration, utilization and review When generating the challenge task list, this embodiment uses the Thompson sampling strategy to dynamically select the next virtual trash item to be disposed of from the question bank, and comprehensively considers the following three factors: Exploration Factor: Virtual junk items are randomly selected from subcategories whose current probability of being mastered by the user is lower than a first threshold (set to 0.4 in this embodiment) using an exploration probability ε. The exploration probability ε gradually decreases as the number of task rounds completed by the user increases. For example, the initial value is 0.2, which decreases to 0.9 times the original value after every 100 tasks are completed, and eventually stabilizes at a small value (e.g., 0.05).

[0041] Utilization Factors: Virtual trash items are selected from subcategories whose current probability of being used is below a second threshold (set to 0.6 in this embodiment) based on Thompson sampling results, using a utilization probability of 1-ε. The specific implementation of Thompson sampling is as follows: A Beta distribution model is maintained for each subcategory, containing a success parameter α and a failure parameter β, initially α=1 and β=1. When a user correctly classifies a virtual trash item belonging to a certain subcategory, the success parameter α for that subcategory is incremented by 1; when the classification is incorrect, the failure parameter β is incremented by 1. When selecting the next item, a random probability value is sampled from the Beta distribution corresponding to each subcategory, and the item from the subcategory with the highest sampled probability value is selected as the next virtual trash item to be disposed of.

[0042] Review Factors: For subcategories where the user's current mastery probability is above the second threshold (i.e., already mastered), the system monitors the user's progress according to the review intervals set by the Ebbinghaus forgetting curve. The forgetting curve can be preset to a series of time intervals, such as 1 day, 3 days, 7 days, 15 days, etc. When the time since the last practice for a mastered category reaches the set review interval, the system automatically inserts a virtual junk item from that category as a review task to consolidate memory.

[0043] (3) Sequence optimization: Task planning based on reinforcement learning To further improve the learning effect, this embodiment also provides an optional sequence optimization mode, which models the task sequence generation process as a Markov decision process (MDP) and calculates the optimal task delivery order through reinforcement learning algorithms.

[0044] In the MDP model, the state is the user's mastery vector, the action is selecting the next item category, and the reward function is defined as the expected increase in knowledge retention rate after the item is correctly classified. The state transition function is defined by the Bayesian knowledge tracing model: after presenting a user with an item of a certain category, the model predicts the mastery vector for the next state based on the user's possible classification results (correct or incorrect).

[0045] This embodiment supports two solution methods: Offline reinforcement learning: A large amount of historical user interaction data is pre-collected to construct an experience replay pool. Each experience contains a quadruple of (current state, action, reward, next state). A deep Q-network is constructed as the policy network, taking the current mastery vector as input and outputting the expected Q-value for each possible action. Mini-batch experiences are randomly sampled from the experience replay pool, and the network parameters are updated using gradient descent to minimize the mean squared error between the current Q-value and the target Q-value. After training, the current mastery vector is input into the deep Q-network, and the action with the highest output Q-value is selected as the next item category.

[0046] Online Monte Carlo Tree Search (MCTS): Constructs a search tree using the current mastery vector as the root node. At each decision step, the following iterative steps are performed until a preset computational budget is reached: Starting from the root node, recursively select child nodes according to a tree strategy (such as the UCT algorithm) until a node that is not fully expanded is reached; add a new child node to this node, corresponding to an unselected item category; starting from the newly expanded node, select item categories using a random strategy and simulate user classification results, using a Bayesian knowledge tracking model to predict state transitions during the simulation, until a preset sequence length is reached; propagate the accumulated reward obtained from the simulation upwards along the path, updating the visit count and average reward of each node on the path. After iteration, select the item category corresponding to the child node with the highest average reward under the root node as the next virtual junk item to be deployed.

[0047] III. Collision Detection and Physics Interaction Simulation Step S3: In the AR interactive scene, spatial hashing partitioning technology is used to divide the interactive area into a three-dimensional grid. A spherical bounding box is constructed for each virtual trash item, and an axis-aligned bounding box is constructed for each virtual trash can. During collision detection, bounding box intersection tests are only performed on virtual trash items and virtual trash cans located in the same or adjacent grids. Specifically, this includes the following sub-steps: (1) Bounding box construction For each virtual trash item, construct a spherical bounding box, with the center of the sphere being the geometric center of the item and the radius being the radius of the item's circumscribed sphere. For each virtual trash can, construct an axis-aligned bounding box (AABB), with its six faces parallel to the coordinate plane of the world coordinate system, and its boundaries determined by the minimum and maximum coordinate values ​​of the trash can model on the X, Y, and Z axes.

[0048] (2) Spatial hash partitioning To reduce unnecessary collision detection calculations, this embodiment employs spatial hash partitioning technology. First, based on a preset grid size (which must be larger than the maximum diameter of the sphere's bounding box to ensure that an object spans at most a few adjacent grids), the interactive area in the AR interactive scene is divided into a uniform three-dimensional grid, and a unique hash index is assigned to each grid.

[0049] In each physical simulation frame, based on the current spatial location of each virtual trash item and each virtual trash can, the hash index of the grid in which it is located is calculated, and the identifiers of the virtual trash items and virtual trash cans are registered in the object list of the corresponding grid.

[0050] During collision detection, for each virtual trash item, the hash index of its surrounding grid is obtained, and the hash indices of the 26 adjacent grids (including those with coplanar, edge-sharing, and vertex-sharing connections) are determined. Bounding box intersection tests are performed only on virtual trash cans located within the grid containing the item and its adjacent grids; for virtual trash cans not located within these grids, the intersection test is skipped.

[0051] (3) Sphere-AABB intersection test For virtual trash items and virtual trash cans that require intersection testing, perform the following steps: Obtain the minimum and maximum values ​​of the axis-aligned bounding box on the X, Y, and Z coordinate axes; The coordinates of the center of the sphere bounding box are restricted to between the minimum and maximum values ​​on the X, Y, and Z axes, respectively, to obtain the point on the axis-aligned bounding box closest to the center of the sphere; Calculate the Euclidean distance between the center of the sphere and the nearest point to obtain the nearest distance; If the nearest distance is less than the radius of the sphere's bounding box, it is determined to be an intersection, and a collision event is generated; if the nearest distance is greater than or equal to the radius of the sphere's bounding box, it is determined to be a non-intersection.

[0052] IV. Touch operation response and motion status update Step S4: Responding to the user's touch operation on the virtual trash item, when the user touches and drags, control the position of the virtual trash item to follow the projection of the touch point on the target plane area; when the user releases the virtual trash item, apply gravity, air resistance, and the elastic force when colliding with the virtual trash can to the virtual trash item, and update the item's velocity and position using an explicit Euler integral with a fixed step size. Specifically, this includes the following sub-steps: (1) Kinematic control mode When a user drags a virtual trash item using touch, the system places the item in kinematic control mode. In this mode, the item's spatial position directly follows the projection of the touch point onto the target plane area, without the physics engine intervening in the calculations. Simultaneously, the system maintains a preset angle between the item's orientation and the camera's optical axis (e.g., always facing the camera) so that the user can clearly identify the item's category.

[0053] (2) Dynamic simulation mode When a user releases a virtual junk item, the system switches the virtual junk item to a dynamic simulation mode, activates the physics engine to apply gravity, air resistance, and collision response to it, and updates its motion parameters using explicit Euler integrals with a fixed step size. The specific steps are as follows: In response to the user's release operation, the position coordinates and velocity vector of the virtual junk item at the moment of release are obtained as the initial state for the dynamics simulation; In each physical simulation frame, perform the following iterative steps: Calculate the net force currently acting on the virtual trash item. The net force consists of three parts: gravity, calculated based on a preset gravitational acceleration constant and mass parameters; air resistance, calculated based on the current velocity vector and a preset air resistance coefficient, with the direction opposite to the velocity direction; and collision response, calculated based on the collision normal vector and the restitution coefficient when the virtual trash item collides with a virtual trash can or a target planar area as determined by the collision detection in step S3, directly correcting the velocity vector. The acceleration of the current frame is calculated based on the resultant force, and the acceleration is equal to the ratio of the resultant force to the mass. The motion state is updated using the explicit Euler integral formula: the velocity vector of the current frame is added to the product of the acceleration and the fixed time step to obtain the velocity vector of the next frame; the position coordinates of the current frame are added to the product of the velocity vector of the current frame and the fixed time step to obtain the position coordinates of the next frame. Use the updated position coordinates and velocity vector as the input state for the next physics simulation frame, and repeat the above steps until the object stops moving or is dragged back by the user.

[0054] V. Real-time Logic Judgment and Multimodal Feedback Step S5: Perform real-time logical judgment on the waste sorting interaction results based on the waste sorting rule base, and trigger multimodal feedback including visual, auditory, and tactile channels based on the judgment results. This specifically includes the following sub-steps: (1) Local rule base accelerates query To ensure real-time feedback even in offline environments, this system uses a pre-built local garbage classification rule base and configures a Bloom filter as a front-end index to accelerate queries.

[0055] The rule base uses a key-value pair data structure to store the mapping relationship between item identifiers and category categories. A Bloom filter is configured for the rule base, which can quickly determine whether an item identifier exists in the rule base in constant time complexity.

[0056] During the discrimination process, a Bloom filter is first used to quickly verify the existence of virtual junk item identifiers. If the verification fails, the item identifier is determined not to exist in the local rule base, and the subsequent query process is terminated; if the verification passes, the classification result corresponding to the item identifier is obtained through a key-value hash lookup. This mechanism can reduce the time of a single discrimination to less than 5ms.

[0057] (2) Multimodal feedback Based on the discrimination results, the system triggers multimodal feedback including visual, auditory, and tactile channels: Visual Feedback: When correctly categorized, the target trash can plays an absorption animation and displays a score prompt; when incorrectly categorized, the virtual trash item bounces back to its original position, and a semi-transparent prompt box is generated next to the item, displaying the correct categorization information. Upon triggering correct or incorrect feedback, GPU instantiation rendering technology is used to generate particle effects. The initial position, initial velocity, and lifecycle parameters of the particles are passed to the vertex shader, and the GPU executes particle position updates and rendering operations in parallel for each frame.

[0058] Auditory feedback: Play a positive chord sound when the classification is correct; play a low-frequency warning sound when the classification is incorrect.

[0059] Haptic feedback: When a classification error occurs, the device's linear motor emits a short vibration.

[0060] (3) Adaptive performance optimization The system monitors the current frame rate and device temperature in real time. When the frame rate is detected to be lower than a first preset threshold (e.g., 50fps) or the device temperature is detected to be higher than a second preset threshold (e.g., 45℃), performance optimization measures are automatically implemented. Optimization measures include: reducing the accuracy of the physics simulation step size, reducing the number of particle emissions, or reducing the rendering resolution ratio to ensure smooth interaction without lag.

[0061] VI. Progress Recording and Settlement Data Generation Step S6: Record and display the game progress, and after completing the task list, calculate the classification accuracy based on the number of correctly classified tasks and the total number of tasks, and generate settlement data. This includes the following sub-steps: During the game, the system displays the current progress, current score, and remaining time in real time. After completing the classification of all virtual trash items in this level, the system enters the settlement screen, calculates the classification accuracy based on the number of correctly classified items and the total number of tasks, and awards virtual rewards based on the score and level difficulty.

[0062] The settlement interface uses a radar chart to display the user's changes in mastery over various categories of waste, helping users intuitively understand their strengths and weaknesses. The system recommends targeted tasks based on changes in mastery.

[0063] All interaction data is compressed and uploaded to the cloud server to optimize the parameters of the Bayesian knowledge tracing model, while also supporting a friend leaderboard feature. Application scenario example: To more intuitively understand the implementation process of this invention, the following example illustrates how a user learns to sort garbage using this method in a home environment.

[0064] Scene Startup: User Xiaoming opens the application installed on his tablet and points the tablet's camera at his living room floor. The screen displays a real-time view of the living room. The system quickly identifies the floor area using a planar detection algorithm and stably "places" four virtual trash cans at spatial anchor points in that area, labeled "Kitchen Waste," "Recyclable Waste," "Hazardous Waste," and "Other Waste," respectively. The trash can models change in real-time as Xiaoming moves his viewpoint, appearing as if they actually exist on the living room floor.

[0065] Task Generation: Xiaoming selects the "Kitchen Battle" level to start the game. Based on Xiaoming's past learning data, the system identifies that he has a high error rate in the subcategories of "Soiled Paper Towels" (belonging to Other Waste) and "Plastic Bottles" (belonging to Recyclable Waste). He has mastered "Eggshells" (kitchen waste), but it has been more than 3 days since his last practice. Taking into account exploration, utilization, and review, the system dynamically generates a challenge task list containing 100 virtual waste items. The frequency of soiled paper towels has been increased, eggshells have been added as a review task, and some items that Xiaoming has not yet encountered (such as expired medicine) have been randomly selected as exploration tasks.

[0066] Interactive Experience: A virtual "eggshell" icon falls from the top of the screen. Xiaoming drags it with his finger and puts it into the "kitchen waste" bin. During the dragging process, the eggshell always follows the finger and faces the camera. The moment it is released, the system determines that the classification is correct, the kitchen waste bin plays an absorption animation and displays "+10 points", and at the same time plays a crisp positive chord sound effect.

[0067] Next, a used napkin appeared, and Xiaoming mistakenly dragged it into the "recyclable waste" bin. After releasing it, the item immediately bounced back to its original position, and a semi-transparent prompt box popped up next to it, displaying: "Error! Soiled tissues belong to other waste and cannot be recycled." At the same time, the device emitted a short vibration and a low-frequency warning sound. The item remained semi-transparent for the next 3 seconds to prevent repeated misoperation.

[0068] Progress Check: Xiaoming completed the classification of all 100 items in 2 minutes and 30 seconds, correctly classifying 95 items for a score of 95. The system displays "Congratulations! You have earned the 'Kitchen Cleaning Expert' badge!" and shows Xiaoming's progress in each category of waste using a radar chart. Xiaoming noticed his understanding of the "Hazardous Waste" category was low, so the system automatically recommended "Hazardous Waste Intensive Practice".

[0069] Data synchronization: The data from this interaction has been uploaded to the cloud server, updating Xiaoming's cognitive state model parameters. On Xiaoming's friend leaderboard, his score surpasses that of 80% of his friends.

[0070] Example 2: like Figure 2 As shown, this embodiment provides an AR interactive garbage sorting challenge game system. The system runs on a mobile terminal, which includes a processor, a graphics processor, a memory, a camera, and a display screen. The memory stores a computer program, and the processor, when executing the computer program, is configured to implement the method described in Embodiment 1 through the following modules: The scene construction module, in response to a scene launch command, calls the camera to capture real-world image data, uses a plane detection algorithm to identify target planar regions for placing virtual objects, and overlays and renders virtual sorting trash can models on spatial anchor points within the target planar regions to generate an AR interactive scene. This module specifically includes: a multi-scale feature extraction unit, used to construct an image pyramid for each frame and extract multi-scale features; a temporal consistency verification unit, used to perform temporal voting to confirm valid planes using VIO; and a plane boundary expansion unit, used to deduce the plane extension range using a semantic segmentation model.

[0071] The task generation module, in response to the user's start command, dynamically generates a list of challenge tasks containing multiple virtual waste items from a waste sorting item question bank. This module specifically includes: a cognitive state modeling unit, used to maintain the user's knowledge mastery vector using a Bayesian knowledge tracking model; a task dynamic adjustment unit, used to select items by comprehensively considering exploration, utilization, and review factors using a Thompson sampling strategy; and a sequence optimization unit, used to calculate the optimal task sequence using reinforcement learning or the MCTS algorithm.

[0072] The interactive simulation module is used in the AR interactive scene to divide the interactive area into a three-dimensional mesh using spatial hashing partitioning technology. It constructs a spherical bounding box for each virtual trash item and an axis-aligned bounding box for each virtual trash can. During collision detection, it only performs bounding box intersection tests on virtual trash items and virtual trash cans located in the same or adjacent meshes. It also responds to user touch operations on virtual trash items, controlling the position of the virtual trash item to follow the projection of the touch point onto the target plane area when the user touches and drags it. When the user releases the virtual trash item, it applies gravity, air resistance, and the elastic force from a collision with a virtual trash can, and updates the item's velocity and position using explicit Euler integrals with a fixed step size. Specifically, this module includes: a collision detection unit for constructing spherical / AABB bounding boxes, dividing the three-dimensional mesh, and performing adjacent mesh detection; a motion control unit for managing the switching between kinematic and dynamic modes; and a rendering optimization unit for implementing GPU-instantiated particle effects and adaptive frame rate scheduling.

[0073] The discrimination feedback module is used to perform real-time logical discrimination of the disposal interaction results based on the waste classification rule base, and trigger multimodal feedback including visual, auditory, and tactile channels based on the discrimination results. This module specifically includes: a local rule base query unit, used to accelerate key-value pair lookup through Bloom filters; and a multimodal feedback output unit, used to coordinate the feedback output of visual, auditory, and tactile channels.

[0074] The data management module records and displays game progress, calculates classification accuracy based on the number of correctly classified tasks and the total number of tasks after the task list is completed, generates settlement data, and synchronizes the interaction data to the cloud server.

[0075] The collaborative work of the above modules realizes the steps described in Example 1, and constructs a complete AR interactive waste sorting teaching closed loop.

[0076] This invention also provides a mobile terminal, including a processor, a memory, a camera, and a display screen. The memory stores a computer program, which, when executed by the processor, implements the AR interactive garbage sorting challenge game method as described in Embodiment 1.

[0077] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the AR interactive garbage sorting challenge game method as described in Embodiment 1.

[0078] In summary, this invention, through a series of algorithm optimizations and hardware resource scheduling strategies such as multi-scale plane detection, cognitive diagnostic task generation, and lightweight physical simulation engine, successfully transforms garbage classification education from a traditional static mode into an AR interactive experience with stable scenes, intelligent content, and realistic and smooth interaction. It effectively overcomes the technical challenges of existing technologies in terms of plane detection stability, personalized task generation, and the difficulty in balancing the realism of mobile physical interaction with real-time performance.

[0079] The specific embodiments of the present invention have been described in detail above, but they are merely examples, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent transformations and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A method for an AR interactive garbage sorting challenge game, characterized in that, Applied to a mobile terminal, the method includes: In response to the scene activation command, the camera is invoked to capture real-world image data, a plane detection algorithm is used to identify the target plane region, and a virtual sorting trash can model is superimposed and rendered on the spatial anchor point of the target plane region to generate an AR interactive scene. In response to the user's start command, a list of challenge tasks containing multiple virtual waste items is dynamically generated from the waste sorting item question bank; In the AR interactive scene, spatial hash partitioning technology is used to divide the interactive area into a three-dimensional grid. A spherical bounding box is constructed for each virtual trash item, and an axis-aligned bounding box is constructed for each virtual trash can. When performing collision detection, only the bounding box intersection test is performed on virtual trash items and virtual trash cans located in the same grid or adjacent grids. In response to the user's touch operation on the virtual trash item, when the user touches and drags, the position of the virtual trash item is controlled to follow the projection of the touch point on the target plane area; when the user releases the virtual trash item, gravity, air resistance and elastic force when colliding with the virtual trash can are applied to the virtual trash item, and the speed and position of the item are updated using explicit Euler integral with a fixed step size. The system performs real-time logical judgment on the results of the waste sorting interaction based on the waste sorting rule base, and triggers multimodal feedback including visual, auditory and tactile channels based on the judgment results; Record and display game progress, and after the task list is completed, calculate the classification accuracy based on the number of correctly classified tasks and the total number of tasks, and generate settlement data.

2. The method according to claim 1, characterized in that, The method of identifying the target planar region using a planar detection algorithm specifically includes: An image pyramid is constructed for each frame of the real environment image data. Multi-scale texture features and edge features are extracted based on the image pyramid and fused through a feature pyramid network to generate a planar candidate region heatmap. The camera pose is estimated in real time using a visual inertial odometry system, and the consistency of candidate planar regions in a series of consecutive frames is checked: the detection window length is set to a preset number of consecutive frames, and when the same candidate planar region is detected a preset threshold number of times within the detection window, the region is identified as the target planar region and its spatial anchor point is locked.

3. The method according to claim 2, characterized in that, The generation of AR interactive scenes also includes: After detecting a partial planar region, the semantic category of the region is inferred using a semantic segmentation model. Combined with the homography matrix of the detected plane, the planar extension range is extrapolated to generate a virtual interactive region larger than the actual detected area.

4. The method according to claim 1, characterized in that, The dynamically generated list of challenge tasks containing multiple virtual junk items specifically includes: The built-in Bayesian knowledge tracing model establishes and maintains a knowledge mastery vector for each user. The knowledge mastery vector contains multiple subcategories, and each subcategory corresponds to a value representing the probability that the user has mastered that subcategory. Based on the correctness of each classification operation, the mastery probability and forgetting parameter of the corresponding sub-category in the knowledge mastery vector are updated in real time. Based on the knowledge mastery vector, the Thompson sampling strategy is used to dynamically select the next virtual trash item to be disposed of from the question bank. The selection process comprehensively considers the following three factors: First factor: Virtual junk items are randomly selected from subcategories where the user's current mastery probability is below a first threshold, with an exploration probability ε. The exploration probability ε gradually decreases as the preset learning process parameter increases. The second factor is to select virtual junk items from subcategories whose probability of being currently known by the user is lower than the second threshold, based on Thompson's sampling results, using a probability of 1-ε, where the second threshold is greater than the first threshold. The third factor: For subcategories that the user has mastered with a probability higher than the second threshold, monitoring is conducted according to the review interval set by the Ebbinghaus forgetting curve. After the interval expires, virtual junk items for that subcategory are automatically inserted as review tasks.

5. The method according to claim 4, characterized in that, The method further includes sequence optimization of the dynamic generation process: The process of generating task sequences is modeled as a Markov decision process, where the state is the user's mastery vector, the action is selecting the next item category, and the reward function is the expected improvement in knowledge retention rate. The optimal item placement order is calculated and output using either offline reinforcement learning or an online Monte Carlo tree search algorithm. If the online Monte Carlo tree search algorithm is used, the following steps are included: constructing a search tree with the current mastery degree vector as the root node, simulating node state transitions using a Bayesian knowledge tracing model, updating node values ​​through backpropagation based on the accumulated rewards obtained from the simulation, and finally selecting the item category corresponding to the child node with the highest value as the next virtual trash item to be deployed.

6. The method according to claim 1, characterized in that, The spatial hash partitioning technique is used to divide the interactive area into a three-dimensional grid. Bounding box intersection tests are performed only on virtual trash items and virtual trash cans located within the same or adjacent grids. Specifically, this includes: According to the preset grid size, the interactive area in the AR interactive scene is divided into a uniform three-dimensional grid, and a unique hash index is assigned to each grid. In each physical simulation frame, based on the current spatial location of each virtual trash item and each virtual trash can, the hash index of the grid in which it is located is calculated, and the identifiers of the virtual trash items and the virtual trash cans are registered to the object list of the corresponding grid. For each virtual junk item, obtain the hash index of the grid it belongs to, and determine the hash index of the grids adjacent to it; Intersection tests of spherical bounding boxes and axis-aligned bounding boxes are performed only on virtual trash cans located within the grid containing the virtual trash item and its adjacent grids. Specific steps include: Obtain the minimum and maximum values ​​of the axis-aligned bounding box on the X, Y, and Z coordinate axes; The coordinates of the center of the sphere bounding box are restricted to between the minimum and maximum values ​​on the X, Y, and Z axes, respectively, to obtain the point on the axis-aligned bounding box closest to the center of the sphere; Calculate the Euclidean distance between the center of the sphere and the nearest point to obtain the nearest distance; If the nearest distance is less than the radius of the sphere's bounding box, then they are considered to intersect. If the nearest distance is greater than or equal to the radius of the sphere's bounding box, then they are determined to be non-intersecting; For virtual trash cans that are not located in the grid of the virtual trash item or in its adjacent grids, skip the intersection test directly.

7. The method according to claim 1, characterized in that, The response to user touch operations on virtual trash items specifically includes switching between kinematic control mode and dynamic simulation mode: When a user touches and drags a virtual trash item, the virtual trash item is placed in kinematic control mode. Physical force calculations are ignored, and its spatial position is directly updated to the projection coordinates of the touch point on the target plane area. The orientation of the virtual trash item is controlled so that the angle between its normal and the camera optical axis is kept at a preset value. When a user releases a virtual trash item, the virtual trash item is immediately switched from kinematic control mode to dynamic simulation mode. The physics engine is activated to calculate gravity, air resistance, and collision response force, and the step of updating the item's velocity and position using explicit Euler integrals with a fixed step size is executed.

8. The method according to claim 1, characterized in that, The triggering of multimodal feedback, including visual, auditory, and tactile channels, based on the discrimination result includes: The visual channel is implemented in the following way: when a correct or incorrect feedback is triggered, GPU instantiation rendering technology is used to generate particle effects, and the initial position, initial velocity, and life cycle parameters of the particles are passed to the vertex shader. The GPU then performs particle position updates and rendering operations in parallel for each frame. The tactile channel is implemented by triggering the device's linear motor to emit a short vibration when the judgment result is incorrect. The auditory channel is implemented by playing a positive chord when the judgment result is correct and a low-frequency warning sound when the judgment result is incorrect. The method further includes: Real-time monitoring of current frame rate and device temperature; When the frame rate is detected to be lower than the first preset threshold or the device temperature is higher than the second preset threshold, performance optimization measures are automatically executed. These performance optimization measures include reducing the accuracy of the physical simulation step size, reducing the number of particle emissions, or reducing the rendering resolution ratio.

9. The method according to claim 1, characterized in that, The real-time logical judgment of the disposal interaction results based on the waste classification rule base specifically includes: A pre-set local waste sorting rule base is provided. The waste sorting rule base uses a key-value pair data structure to store the mapping relationship between item identifiers and classification categories, and a Bloom filter is configured as a pre-index for the waste sorting rule base. During the identification process, the existence of virtual junk items is first quickly verified using the Bloom filter: If the verification fails, it is determined that the item identifier does not exist in the local rule base, and the subsequent query process is terminated; If the verification passes, the category result corresponding to the item identifier is obtained through key-value pair hash lookup.

10. An AR interactive garbage sorting challenge game system, characterized in that, The system includes a memory and a processor, the memory storing a computer program, and the processor, when executing the computer program, being configured to implement the method as described in any one of claims 1 to 9 through the following modules: The scene construction module is used to: call the camera to capture real environment image data, use a plane detection algorithm to identify the target plane area, and overlay and render a virtual sorting trash can model on the spatial anchor point of the target plane area to generate an AR interactive scene. The task generation module is used to: dynamically generate a list of challenge tasks containing multiple virtual garbage items from the garbage sorting item question bank in response to the user's start command; The collision detection module is used to: divide the interactive area into a three-dimensional grid using spatial hash partitioning technology in the AR interactive scene, construct a spherical bounding box for each virtual trash item, construct an axis-aligned bounding box for each virtual trash can, and only perform bounding box intersection tests on virtual trash items and virtual trash cans located in the same grid or adjacent grids during collision detection. The interactive control module is used to: respond to the user's touch operation on the virtual trash item; when the user touches and drags, control the position of the virtual trash item to follow the projection of the touch point on the target plane area; when the user releases the virtual trash item, apply gravity, air resistance and elastic force when colliding with the virtual trash can to the virtual trash item, and update the speed and position of the item using explicit Euler integral with a fixed step size. The discrimination feedback module is used to: perform real-time logical discrimination on the disposal interaction results based on the garbage classification rule base, and trigger multimodal feedback including visual, auditory and tactile channels based on the discrimination results; The data management module is used to record and display game progress, and after the task list is completed, calculate the classification accuracy based on the number of correctly classified tasks and the total number of tasks, and generate settlement data.

Images