Intelligent agent companion playing recognition method based on multi-feature combination and related equipment

Abstract

The application relates to the field of artificial intelligence and discloses an intelligent agent accompanying play recognition method based on multi-feature combination and related equipment, which comprises the following steps: collecting inertial sensing data of a hardware carrier bound with a target intelligent agent at a first sampling frequency, acquiring first inertial sensing data for feature analysis, switching to a second sampling frequency for collection when a preset triggering condition is met, and acquiring second inertial sensing data; performing multi-dimensional feature extraction and fusion on the second inertial sensing data to obtain fusion motion features to be recognized; inputting the fusion motion features into a pre-trained motion pattern recognition model to recognize a current motion pattern; triggering a corresponding target accompanying play event based on the current motion pattern, and updating a state parameter of the target intelligent agent. The application triggers a corresponding target accompanying play event after motion pattern recognition through hierarchical sampling and multi-feature fusion, realizes accurate linkage between real motion and accompanying play events, and takes into account sampling power consumption.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and related equipment for identifying intelligent agents for companionship based on multi-feature combination. Background Technology

[0002] With the development of generative artificial intelligence, virtual character interaction, and smart hardware technology, AI-powered intelligent agents with growth attributes, emotional feedback capabilities, and long-term companionship abilities are gradually becoming an important direction for the integration of software and hardware. Some existing products drive the state changes of virtual intelligent agent pets or AI characters through chat dialogues, click interactions, and check-in tasks. Other systems further leverage hardware carriers bound to the intelligent agent to trigger adventure events related to the virtual world in the real world.

[0003] Because real-world behaviors such as carrying, deliberately shaking, walking, running, and riding in a vehicle have significant differences in their companionship semantics and events, most existing technologies can only roughly determine whether the hardware carrier is moving by continuously detecting whether the device is moving or shaking. This not only has high power consumption but also makes it difficult to accurately distinguish the interaction semantics of different movement behaviors, resulting in an inability to accurately distinguish different companionship behaviors of users and reducing the accuracy of identifying real-world play events with intelligent pets. Summary of the Invention

[0004] The main objective of this invention is to provide a method and related equipment for identifying intelligent agents who provide companionship based on multiple features, aiming to solve the problem in the prior art of accurately identifying real-world companionship events of intelligent agents with low power consumption.

[0005] The technical solution of the present invention is as follows: The first aspect of this invention provides a method for identifying intelligent agents for companionship based on multi-feature joint analysis, comprising: Inertial sensing data is collected from the hardware carrier bound to the target intelligent agent at a first sampling frequency to obtain the first inertial sensing data. The first inertial sensing data is subjected to feature analysis to confirm whether the preset triggering condition is met. If it is met, the sampling frequency is switched to the second sampling frequency to collect inertial sensing data from the hardware carrier and obtain the second inertial sensing data. The second sampling frequency is greater than the first sampling frequency. Multi-dimensional feature extraction and fusion are performed on the second inertial sensing data to obtain the fused motion features to be identified; The fused motion features are input into a pre-trained motion pattern recognition model to identify the current motion pattern. Based on the current movement pattern, a corresponding target play event is triggered, and the state parameters of the target agent are updated.

[0006] A second aspect of the present invention provides an intelligent agent companion recognition device based on multi-feature joint recognition, comprising: The data sampling module is used to collect inertial sensing data from the hardware carrier bound to the target intelligent agent at a first sampling frequency, and to obtain the first inertial sensing data. The sampling switching module is used to perform feature analysis on the first inertial sensing data, confirm whether the preset triggering conditions are met, and if so, switch to the second sampling frequency to collect inertial sensing data from the hardware carrier to obtain the second inertial sensing data. The second sampling frequency is greater than the first sampling frequency. The feature extraction module is used to extract and fuse multi-dimensional features from the second inertial sensing data to obtain the fused motion features to be identified. The motion recognition module is used to input the fused motion features into a pre-trained motion pattern recognition model to identify and obtain the current motion pattern; The companion play event mapping module is used to trigger corresponding target companion play events based on the current movement mode and update the state parameters of the target intelligent agent.

[0007] A third aspect of the present invention provides a computer device including at least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described intelligent agent companion recognition method based on multi-feature combination.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by one or more processors, cause the one or more processors to perform the above-described intelligent agent companion recognition method based on multi-feature joint processing.

[0009] Beneficial Effects: This invention discloses a method and related equipment for identifying intelligent agents for companion gaming based on multi-feature fusion. Compared with existing technologies, this invention's embodiments acquire first inertial sensing data by collecting inertial sensing data from a hardware carrier bound to a target intelligent agent at a first sampling frequency; perform feature analysis on the first inertial sensing data to confirm whether a preset triggering condition is met; if so, switch to a second sampling frequency to collect inertial sensing data from the hardware carrier, acquiring second inertial sensing data, where the second sampling frequency is greater than the first sampling frequency; perform multi-dimensional feature extraction and fusion on the second inertial sensing data to obtain fused motion features to be identified; input the fused motion features into a pre-trained motion pattern recognition model to identify the current motion pattern; trigger the corresponding target companion gaming event based on the current motion pattern and update the state parameters of the target intelligent agent. This invention achieves precise linkage between real-world motion and companion gaming events by triggering corresponding target companion gaming events after motion pattern recognition through hierarchical sampling and multi-feature fusion, while considering sampling power consumption. Attached Figure Description

[0010] To more clearly illustrate the solutions in this invention, the accompanying drawings used in the description of the embodiments of this invention will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for the intelligent agent companion recognition method based on multi-feature joint analysis provided in an embodiment of the present invention. Figure 2 A flowchart of an intelligent agent companion recognition method based on multi-feature joint analysis provided in an embodiment of the present invention; Figure 3 A schematic diagram of the functional modules of the intelligent agent play-along recognition device based on multi-feature combination provided in an embodiment of the present invention; Figure 4 A schematic diagram of the hardware structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The embodiments of the invention are described below in conjunction with the accompanying drawings.

[0013] The intelligent agent companion recognition method based on multi-feature joint analysis provided in this invention can be applied to, for example... Figure 1The hardware and virtual intelligent agent collaborative application environment shown includes a terminal device 101, a hardware carrier 102 bound to the target intelligent agent, a network 103, and a server 104. The network 103 serves as a medium to provide a communication link between the terminal device 101, the hardware carrier 102, and the server 104. The network 103 can include various connection types, such as wired and / or wireless communication links (e.g., Bluetooth, Wi-Fi, NFC, etc.).

[0014] Users can interact with server 104 via network 103 using terminal device 101 and hardware carrier 102 to receive or send messages, etc. Terminal device 101 may have a client installed that supports virtual scenarios. For example, when the virtual scenario is pet companionship, the client could be a virtual pet application. Users can log in to the client to view, edit, or control the interactions of the bound target intelligent agent in the virtual scenario (this is just an example). Terminal device 101 can be various electronic devices with a display screen and web browsing support, including but not limited to smartphones, tablets, and desktop computers.

[0015] The hardware carrier 102, bound to the target intelligent agent, can serve as a real-world interaction medium between the user and the target intelligent agent. It can sense the user's close-range touch, shaking, movement, and other real-world activities in real time and associate them as corresponding interactive events. The hardware carrier 102 can be a plush toy, smart wearable accessory, desktop smart ornament, smart speaker, etc., that integrates motion sensing and near-field communication recognition functions. The hardware carrier 102 has built-in inertial sensors (accelerometer, gyroscope, magnetometer, etc.) to collect inertial sensing data when the user carries or operates the hardware carrier.

[0016] Server 104 can be a server that provides various services, such as a backend server that supports the content browsed by the user using terminal device 101 and hardware carrier 102 (this is just an example). The backend server can analyze and process data such as received user requests, and feed the processing results back to the user through terminal device 101 and hardware carrier 102. Server 104 can be a cloud server, a server of a distributed system, or a server combined with blockchain.

[0017] It should be understood that the number of terminal devices 101, hardware carriers 102, networks 103, and servers 104 mentioned above is merely illustrative. Depending on the implementation needs, there can be any number of terminal devices 101, hardware carriers 102, networks 103, and servers 104. For example, a single user can correspond to one terminal device 101 and one hardware carrier 102, and multiple users can achieve interactive linkage between target intelligent agents through the network.

[0018] like Figure 2As shown, the intelligent agent companion recognition method based on multi-feature joint analysis provided in this embodiment of the invention specifically includes the following steps: S201. Inertial sensing data is collected from the hardware carrier bound to the target intelligent agent at a first sampling frequency to obtain the first inertial sensing data.

[0019] In this embodiment, the target intelligent agent is an artificial intelligence virtual character (such as an AI pet or virtual partner) with companionship and growth attributes. It is bound to a hardware carrier, and the movement state of the hardware carrier is associated with the virtual world events of the target intelligent agent. Users can trigger companionship events for the target intelligent agent through action interactions with the hardware carrier, such as shaking, running, jumping, etc., to achieve virtual-real linkage between real-world movement and play events.

[0020] Specifically, in order to achieve low-power companion event recognition, the hardware carrier enters a low-power listening mode by default after startup. The built-in inertial sensors, including accelerometers, gyroscopes, magnetometers, etc., continuously collect inertial sensing data at a low first sampling frequency to reduce the power consumption of the hardware carrier. After the collection is completed, the first inertial sensing data can be temporarily stored in the local cache of the hardware carrier without real-time uploading to the server. Subsequent high-frequency sampling and data processing are only started when the triggering conditions are met, further reducing power consumption.

[0021] This embodiment uses a low-power sampling frequency to continuously monitor inertial sensing data, minimizing the energy consumption of the hardware carrier, making it more suitable for the battery life requirements of small smart hardware, avoiding rapid power consumption caused by frequent high-frequency sampling, and ensuring long-term stable operation of the device.

[0022] S202. Perform feature analysis on the first inertial sensing data to confirm whether the preset triggering conditions are met. If they are met, switch to the second sampling frequency to collect inertial sensing data from the hardware carrier and obtain second inertial sensing data. The second sampling frequency is greater than the first sampling frequency.

[0023] In this embodiment, the first inertial sensor data obtained through continuous low-power monitoring is subjected to feature analysis to determine whether a preset trigger condition is met, thereby identifying whether the user is actively engaging in interactive gaming. Specifically, the preset trigger condition is a judgment criterion based on the motion characteristics of the interactive gaming behavior. Either the trigger condition is determined after accumulating a certain number of sets of first inertial sensor data, or after accumulating first inertial sensor data for a preset time period; this embodiment does not limit the determination to either. If the determination result indicates that the preset trigger condition is met, the system immediately switches to a second sampling frequency for enhanced sampling (i.e., the second sampling frequency is greater than the first sampling frequency). If the determination result indicates that the preset trigger condition is not met, the system continues to maintain the first sampling frequency for low-power monitoring, cyclically executing feature analysis and trigger judgment.

[0024] For example, a user puts the hardware carrier bound to the intelligent pet into their pocket. After the device is turned on, it continuously collects inertial sensing data at a first sampling frequency, such as 5Hz. At this time, the device is in a low-power state. If the user takes the hardware carrier out of the pocket and deliberately shakes the device to interact with the intelligent pet, the acceleration modulus and angular velocity modulus in the first inertial sensing data collected at this time will show significant changes. After feature analysis, the preset trigger conditions are met, and the device immediately switches to a second sampling frequency, such as 50Hz, for enhanced sampling in order to more accurately capture the motion data during the shaking process.

[0025] This embodiment accurately filters out potential companion behaviors by analyzing the features of the first inertial sensor data, avoiding energy waste caused by invalid high-frequency sampling, and achieving a balance between low-power monitoring and high-precision recognition. At the same time, the high-frequency collected second inertial sensor data can also more accurately capture motion details, providing high-quality data for subsequent multi-dimensional feature extraction and ensuring the accuracy of recognition.

[0026] S203. Perform multi-dimensional feature extraction and fusion on the second inertial sensing data to obtain the fused motion features to be identified.

[0027] In this embodiment, the second inertial sensing data is high-frequency collected refined data containing rich motion details. For the second inertial sensing data collected for potential companionship behavior, features that can characterize the motion pattern are extracted from different dimensions, and integrated to obtain fused motion features that can comprehensively characterize the current motion state.

[0028] Specifically, multiple input channels can be used to simultaneously receive second inertial sensor data such as acceleration, angular velocity, and magnetometer data. Each channel undergoes independent denoising and normalization processing to remove environmental interference. For example, moving average filtering can be used to remove noise, and min-max normalization can be used to map the data to the [0,1] interval to avoid the impact of dimensional differences between different dimensions of data on feature fusion. Then, multi-dimensional features are extracted and fused, including time domain, frequency domain, attitude, rhythm, and context dimensions, to form fused motion features. This avoids recognition bias caused by single-dimensional features and improves the accuracy of motion pattern recognition.

[0029] This embodiment comprehensively captures motion details from multiple dimensions such as time domain, frequency domain, and attitude, effectively avoiding the limitations of single-dimensional features. It integrates motion information from different dimensions to reduce recognition bias caused by incomplete features, and provides more comprehensive input features for subsequent motion pattern recognition models.

[0030] S204. Input the fused motion features into the pre-trained motion pattern recognition model to identify and obtain the current motion pattern.

[0031] In this embodiment, the motion pattern recognition model can be a classification model built on a deep learning architecture, including a rule determiner, a machine learning classifier, a deep learning model, or a fusion structure of them. It is pre-trained using a large amount of fused motion feature data labeled with motion pattern tags, so that it can accurately identify different motion patterns, including stationary mode, portable mode, deliberate shaking mode, walking mode, running mode, etc.

[0032] Specifically, a pre-trained motion pattern recognition model is invoked. The fused motion feature vector, extracted and integrated from multiple dimensions, is input into the model for inference calculations. The model can sample convolutional neural networks and long short-term memory network architectures, extracting local time-series features through convolution, and further filtering key motion features through an attention mechanism. It then classifies and outputs the probability distribution of each motion pattern, identifying the motion pattern with the highest probability value as the current motion pattern. For example, when a user shakes a portable sensing device, the extracted fused motion feature vector is input into the motion pattern recognition model. The model outputs the probability distribution of each motion pattern: intentional shaking mode 95%, portable mode 3%, and stationary mode 2%. Therefore, the current motion pattern is identified as "intentional shaking mode".

[0033] This embodiment uses a pre-trained model to quickly and accurately identify motion patterns by extracting and fusing features from multiple dimensions. It can not only distinguish whether the hardware carrier is moving, but also finely distinguish different motion patterns, effectively distinguish various active companion behaviors, provide a reliable basis for subsequent companion events, significantly reduce the probability of motion pattern misjudgment, and ensure the accuracy and efficiency of companion recognition.

[0034] S205. Trigger the corresponding target play event based on the current movement mode, and update the state parameters of the target agent.

[0035] In this embodiment, the target companionship event is a virtual event that corresponds one-to-one with the movement pattern of the hardware carrier in real-world activities. It is used to map the user's real-world companionship behavior into interactive feedback from the virtual intelligent agent, enhancing the user's sense of immersion and companionship. Specifically, a one-to-one correspondence between different movement patterns and companionship events can be pre-set, and the mapping rules can be dynamically adjusted according to the user's actual needs to ensure that each real-world behavior generates reasonable and identifiable event feedback in the virtual world.

[0036] After matching and triggering the target companionship world based on the currently identified movement pattern, the target agent's state parameters are updated according to the event type. These state parameters include emotion value, intimacy level, growth value, and physical strength value, reflecting the target agent's growth changes under the current companionship event. The updated state parameters are then synchronized to the server, mobile terminal, and hardware carrier to ensure consistency across multiple devices.

[0037] For example, the current movement mode is "deliberate shaking mode", and the matched target play event is "play interaction event". After the event is triggered, the shaking interaction animation of the AI ​​pet agent can be displayed on the mobile terminal. At the same time, the agent's intimacy level (+5) and emotion value (+3) are updated. The updated status parameters are synchronized to the server to ensure that the user can see the latest agent status when logging in next time.

[0038] Based on the accurate differentiation of different motion modes, this embodiment deeply links the user's different real-world companionship behaviors with the virtual intelligent agent's companionship event triggering and state updates. While taking into account sampling power consumption, it achieves accurate linkage between real-world motion and companionship events, enhancing the immersiveness and companionship experience of the user's interaction with the intelligent agent, enriching the interaction scenarios, and improving user stickiness and experience.

[0039] In the above embodiments, this invention discloses a method for identifying intelligent agents for companion gaming based on multi-feature fusion. The method involves acquiring first inertial sensing data by collecting inertial sensing data from a hardware carrier bound to a target intelligent agent at a first sampling frequency; performing feature analysis on the first inertial sensing data to confirm whether a preset triggering condition is met; if so, switching to a second sampling frequency to collect inertial sensing data from the hardware carrier to acquire second inertial sensing data, where the second sampling frequency is greater than the first sampling frequency; extracting and fusing multi-dimensional features from the second inertial sensing data to obtain fused motion features to be identified; inputting the fused motion features into a pre-trained motion pattern recognition model to identify the current motion pattern; triggering a corresponding target companion gaming event based on the current motion pattern and updating the state parameters of the target intelligent agent. This invention achieves precise linkage between real-world motion and companion gaming events by triggering corresponding target companion gaming events after motion pattern recognition through hierarchical sampling and multi-feature fusion, while considering sampling power consumption.

[0040] In this embodiment, step S202 includes: Acceleration and angular velocity data are extracted from multiple consecutive sampling windows from the first inertial sensing data; The acceleration and angular velocity data within the multiple consecutive sampling windows are analyzed for magnitude variation, rhythm variation, reciprocating motion determination, and static state interruption determination. Confirm whether there is at least one of the following: acceleration magnitude change greater than the first threshold, angular velocity magnitude change greater than the second threshold, periodic rhythmic change, reciprocating motion, or interruption of the stationary state. If so, the preset triggering condition is met.

[0041] In this embodiment, when confirming whether the preset triggering condition is met based on the first inertial sensing data obtained from low-power sampling, acceleration data and angular velocity data within multiple consecutive sampling windows are extracted from the first inertial sensing data. By using multiple consecutive sampling windows, the accuracy of feature analysis is improved, and misjudgment caused by the randomness of data from a single window is avoided. The specific sampling window duration can be selected from 0.5 to 2 seconds, or other window durations, etc. Acceleration data and angular velocity data from several sampling windows, such as 3 or other numbers, are extracted continuously to ensure that continuous motion change features can be captured.

[0042] Subsequently, feature analysis was performed on the acceleration and angular velocity data within multiple consecutive sampling windows from multiple analytical perspectives, including modulus variation analysis, rhythm variation analysis, reciprocating motion determination, and static state interruption determination. Among them, the magnitude change analysis determines whether there are significant changes in acceleration and angular velocity to reflect whether the user has taken active actions. It calculates the difference between the maximum and minimum values ​​of acceleration magnitude and angular velocity magnitude within each sampling window. If the difference exceeds the corresponding threshold, it indicates that there is a significant motion change. The rhythm change analysis determines whether the motion is periodic. Active play behaviors (such as shaking and walking) usually have obvious periodicity. The peak frequency of the acceleration or angular velocity signal can be analyzed by Fast Fourier Transform (FFT) to confirm whether the periodic characteristics of the acceleration or angular velocity signal within the continuous sampling window meet the set frequency range and amplitude threshold to determine the rhythm. The reciprocating motion determination determines whether there are short-term high-amplitude repetitive oscillations (such as deliberate shaking). It can be determined by calculating the number of zero crossovers and the amplitude of the acceleration direction change within the continuous sampling window. If the number of zero crossovers exceeds the preset value and the amplitude is greater than the threshold, it is determined that there is reciprocating motion. The static state interruption determination determines whether the hardware carrier has changed from a long-term static state to a moving state, and if the duration is abnormally long, it is determined that the static state has been interrupted. For the extracted data from multiple consecutive sampling windows, the above four judgment methods are executed respectively, and the result of each judgment method (satisfied or not satisfied) is recorded to provide a basis for the final judgment of subsequent triggering conditions.

[0043] After summarizing the above four judgment results, if any of the following conditions exist, it is determined that the preset triggering condition is met: the change in acceleration magnitude of any sampling window is greater than the first threshold; the change in angular velocity magnitude of any sampling window is greater than the second threshold; there is a periodic rhythmic change; there is reciprocating motion; or the static state is interrupted. The first and second thresholds are reasonable values ​​preset based on a large amount of experimental data, used to distinguish between active actions and minor disturbances. If none of the above conditions exist, it is determined that the preset triggering condition is not met, and low-power monitoring continues.

[0044] This embodiment analyzes multiple dimensions such as magnitude changes and rhythm changes of acceleration and angular velocity to accurately distinguish between user-initiated play behaviors and passive interference (such as slight vibrations or accidental device contact), effectively reducing invalid high-frequency sampling triggers, further improving the accuracy and reliability of trigger judgment, providing effective protection for subsequent high-frequency sampling switching and motion mode recognition, and avoiding energy waste caused by invalid high-frequency sampling.

[0045] In one embodiment, step S203 includes: The second inertial sensing data is segmented according to a preset time window to obtain several data segments; Within each data segment, time-domain features, frequency-domain features, attitude features, and rhythm features are extracted respectively; Based on the timestamp of each data segment, the task status and application foreground / background status of the target intelligent agent are obtained, and context features are extracted. The temporal, frequency, pose, rhythm, and contextual features of each data segment are spliced ​​and fused to obtain the fused motion features to be identified.

[0046] In this embodiment, the second inertial sensing data obtained by high-frequency sampling is segmented according to a preset time window, preferably 1 to 5 seconds, to balance the time resolution and computational complexity of feature calculation. For example, the second inertial sensing data can be segmented into sliding segments with a preset time window of 1 second and a sliding step size of 0.5 seconds to ensure overlap between adjacent segments and avoid breakage of motion features. The specific window size and sliding step size can be flexibly set according to requirements, and this embodiment does not limit them. Segmentation processing through a preset time window helps to capture local motion features, reduce the influence of noise, and facilitate the recognition of continuous motion patterns.

[0047] Subsequently, core features for characterizing motion patterns are extracted from different dimensions within each data segment, including time-domain features, frequency-domain features, attitude features, and rhythm features. This ensures the comprehensiveness of the fused motion features by achieving mutual complementarity among the features. Specifically, time-domain features reflect the statistical characteristics of inertial sensing data over time, including mean, variance, standard deviation, peak value, peak-to-peak value, root mean square, skewness, kurtosis, and zero-crossing rate. Frequency-domain features reflect the distribution characteristics of inertial sensing data over frequency, extracted through FFT transformation, including dominant frequency, band energy, energy ratio, spectral peak value, spectral entropy, and periodic stability. Attitude features reflect the attitude changes of the hardware carrier, which can be calculated based on accelerometer and gyroscope data, including attitude change angle, attitude change rate, number of flips, peak angular velocity, and degree of directional change. Rhythm features reflect the rhythmicity and periodicity of motion, including peak spacing, step frequency, period length, and periodic consistency.

[0048] Furthermore, in addition to extracting motion features, contextual features are also extracted to supplement the scene information for motion pattern recognition, avoiding misjudgment of the same motion pattern in different scenarios and improving recognition accuracy. Contextual features include at least the task state of the target intelligent agent and the foreground and background states of the application, and may also include, for example, the connection status between the hardware terminal and the mobile terminal, historical interaction states, etc.

[0049] Specifically, each data segment contains a corresponding timestamp, i.e. the collection time. Based on the timestamp, the task status of the current target agent and the front-end and back-end status of the agent application can be queried. Of course, other data such as connection status and historical interaction status can also be collected. The collected status data is converted into binary code and used as the context feature of the data segment. Together with time domain, frequency domain, attitude, and rhythm features, it constitutes a multi-dimensional feature.

[0050] For each data segment, the time domain features, frequency domain features, posture features, rhythm features, and context features are sequentially concatenated to integrate features of different dimensions into a fixed-dimensional feature vector. Then, global average pooling is performed on the feature vectors of all data segments to obtain a fused motion feature vector, which serves as the final input to the motion pattern recognition model. This vector comprehensively contains all feature information from the time domain, frequency domain, posture, rhythm, and context, and can accurately represent the current motion state.

[0051] This embodiment extracts features from time domain, frequency domain, pose, rhythm, and context in segments and then integrates them to obtain fused motion features. This ensures the comprehensiveness and representativeness of the fused motion features, effectively avoids the recognition bias caused by the one-sidedness of single-dimensional features, provides high-quality and high-recognition input for the motion pattern recognition model, and further improves the reliability of motion pattern recognition.

[0052] In one embodiment, the motion pattern recognition model includes at least a convolutional layer, an attention layer, and a classification output layer, and step S204 includes: The fused motion features are input into the convolutional layer to extract local time series patterns, thereby obtaining the corresponding local time series features. The local time series features are input into the attention layer, and attention weights are assigned and filtered through the attention mechanism to obtain attention-weighted features. The attention-weighted features are input into the classification output layer for classification output mapping to obtain the probability distribution of each motion mode, and the motion mode with the highest probability value is identified as the current motion mode.

[0053] In this embodiment, the pre-built and trained motion pattern recognition model includes at least a convolutional layer, an attention layer, and a classification output layer. The convolutional layer extracts local time-series patterns from the fused motion features, captures local detailed features of the motion, and filters redundant information. Specifically, the fused motion feature vector is input into the convolutional layer, and the feature vector is convolved by convolution kernels corresponding to local feature patterns to extract local time-series features. Preferably, after the convolution operation, a non-linear transformation is performed using the ReLU activation function to avoid gradient vanishing while retaining effective local features, resulting in the final local time-series feature vector.

[0054] The local time-series features are then input into the attention layer. An attention mechanism is used to assign and filter these features based on their attention weights. Specifically, the attention layer employs a self-attention mechanism to weight the local time-series features, focusing on key features relevant to motion pattern recognition and minimizing the influence of irrelevant features. The local time-series features are input into the attention layer, and the feature vectors are first linearly transformed to obtain a query vector (Q), a key vector (K), and a value vector (V). Then, the similarity between the query vector and the key vector is calculated using self-attention to obtain an attention weight matrix. This weight matrix is ​​normalized and then weighted and summed with the value vector to obtain an attention-weighted feature vector. This vector highlights the role of key motion features, thereby improving the ability to recognize actions.

[0055] Finally, the attention-weighted features are input into the classification output layer for classification output mapping to obtain the probability distribution of each motion mode. This classification output layer can adopt a structure of fully connected layer and Softmax function. The fully connected layer is used to map the attention-weighted features to the category dimension of the motion mode. Specific motion modes include stationary mode, portable mode, deliberate shaking mode, walking mode, running mode, jumping mode, riding mode, mixed mode, invalid mode, unknown mode, etc. The Softmax function is used to convert the output result into the probability distribution of each motion mode.

[0056] Specifically, attention-weighted features are input into a fully connected layer, and the feature vectors are mapped to the category dimension through a linear transformation, corresponding to different motion modes. Then, the multidimensional output is converted into the probability distribution of each motion mode through the Softmax function. The probability distribution is traversed to find the motion mode with the highest probability value, which is taken as the current motion mode. At the same time, the confidence level (i.e. the highest probability value) of the motion mode is output, which is used as the basis for confirming whether the target companion play event is triggered.

[0057] It should be noted that during the training phase of the motion pattern recognition model, the training data consists of a large amount of fused motion feature data labeled with motion pattern tags, and the training data is enhanced by random noise superposition and time window shifting. Then, the training data is divided into training set, validation set and test set in a 7:2:1 ratio. The loss function is the cross-entropy loss function, and the optimizer is the Adam optimizer (initial learning rate 0.001). The training cycle is 50 to 100 rounds. The motion pattern recognition performance is measured by accuracy, recall and F1 score to ensure the model's recognition accuracy.

[0058] This embodiment leverages the synergistic effect of convolutional layers extracting local time-series features and attention layers filtering key motion features to enhance key features related to companionship behavior and weaken redundant interference features. This not only improves the model's recognition efficiency and accuracy but also enhances the model's adaptability to different companionship scenarios (such as shaking and walking), ensuring accurate recognition of motion patterns even in complex scenarios.

[0059] In one embodiment, step S205 includes: Obtain a preset event mapping table, which stores the mapping relationship between different movement modes and different play-along events; Based on the mapping relationship in the event mapping table, the current movement mode is mapped to the target play-along event and a trigger weight is generated; When the trigger weight is greater than the weight threshold, the target play-along event is triggered, and the state parameters of the target agent are updated according to the trigger result of the target play-along event.

[0060] In this embodiment, the event mapping table is a pre-defined association table stored on the server. It is used to establish a one-to-one correspondence between movement modes and target companion events. The mapping rules can be dynamically adjusted according to user needs or application scenarios to improve interaction flexibility. The mapping table contains information such as movement mode, companion event type, event description, and basic trigger weight. The basic trigger weight can be used to characterize the basic priority of the companion event triggered by the movement mode, reflecting the closeness of the association between the companion event and the movement mode.

[0061] Specifically, when triggering a corresponding target play event based on the current movement pattern, the latest event mapping table is first retrieved. For example, the corresponding relationships in the event mapping table may include: The "Carry-on Mode" corresponds to companionship and interaction events. Event description: The target AI agent follows the user, updates the AI ​​agent's location and following status, and increases intimacy. The base trigger weight is 0.6. The intentional shaking mode corresponds to playful interaction events. Event description: The target AI interacts with the user by shaking, triggering virtual actions and increasing emotional value and intimacy. The base trigger weight is 0.8. Walking mode corresponds to short-distance exploration events. Event description: The target agent follows the user on a short-distance exploration, advancing the map exploration progress and increasing growth value and exploration experience. The basic trigger weight is 0.7. Running mode corresponds to high-intensity adventure events. Event description: The target AI agent follows the user on high-intensity adventures, triggering task challenges and increasing experience points. The base trigger weight is 0.9. The ride-sharing mode corresponds to long-distance travel events. Event description: The target intelligent agent follows the user on a ride-sharing adventure, triggering updates to the map location and travel task status. The basic trigger weight is 0.6. The static mode corresponds to the rest and dwell event. Event description: The target agent enters a resting state, restores its stamina or updates its dwelling status. The base trigger weight is 0.5. The jump mode corresponds to a special reward event. Event description: The target agent follows the user agreement and receives a growth or intimacy bonus. The base trigger weight is 0.9.

[0062] Understandably, the contents of the event mapping table can be updated in the background according to actual interaction needs, such as adding or editing, to improve the scalability of the system.

[0063] Then, based on the currently identified motion pattern, the corresponding target play-along event and basic trigger weight are matched from the event mapping table. Combined with the confidence score output by the motion pattern recognition model, the trigger weight is calculated. Specifically, the trigger weight is used to determine whether the target play-along event needs to be triggered. It can be calculated jointly from the basic trigger weight and the motion pattern confidence score. For example, trigger weight = basic trigger weight × motion pattern confidence score. Optionally, if the confidence score of the motion pattern is lower than a preset lower limit, such as 0.5, the trigger weight is set to 0 regardless of the magnitude of the basic trigger weight to avoid false triggers due to insufficient reliability of motion pattern recognition.

[0064] The calculated trigger weight is compared with a weight threshold (e.g., 0.7 or other thresholds). The weight threshold is a preset trigger judgment standard used to filter out false triggering events with low weight. If the trigger weight is greater than the weight threshold, the target play event corresponding to the current motion mode is triggered. The corresponding animation feedback (such as play interaction animation) can be displayed on the mobile terminal. At the same time, the state parameters of the target agent are updated based on the state parameter update rules of the event. After the update is completed, the latest state parameters are synchronized to the server, mobile terminal, and hardware carrier to ensure consistency across multiple devices. If the trigger weight is less than or equal to the weight threshold, the target play event is not triggered, and the system returns to the low-power monitoring mode.

[0065] This embodiment establishes a mapping relationship between movement patterns and companion events. By combining trigger weights with the confidence level of movement patterns, it accurately and reliably filters valid companion events, avoids false triggering of events by low-confidence identification results, ensures the rationality and accuracy of companion event triggering, and achieves precise linkage between movement patterns and virtual events, further enhancing the user's experience of interacting with the intelligent agent.

[0066] In one embodiment, after mapping the current motion pattern to a target play-along event according to the mapping relationship in the event mapping table, the method further includes: If the current motion pattern is mapped to a specified misjudged candidate event, then trigger restriction processing is applied to the misjudged candidate event, either by filtering out the misjudged candidate event or reducing the trigger weight of the misjudged candidate event.

[0067] In this embodiment, to further improve the reliability of the identification and triggering of companion play events, the mapping results of the motion mode are further subjected to misjudgment suppression processing. That is, if the current motion mode is mapped to a specified misjudgment candidate event, the trigger restriction processing is performed on the misjudgment candidate event. The misjudgment candidate event refers to companion play events that are easily triggered by mistake, corresponding to motion modes that are easily misjudged, including companion play events corresponding to the carry-on mode and the ride mode. The motion modes corresponding to these events (passive carry, ride vibration) are easily confused with active companion play behavior, so trigger restriction processing is required to reduce the misjudgment rate.

[0068] Specifically, a list of potential misjudged events is pre-defined, including "companionship interaction events" (corresponding to the portable mode) and "long-distance travel events" (corresponding to the car travel mode). When the target companionship event mapped by the current sports mode is a potential misjudged event, trigger restriction processing is performed. The specific restriction processing includes two types: one is filtering processing, for example, if the confidence level of the sports mode is preset to a lower limit, such as 0.5, and the context features show that the intelligent agent application is in the background, then the potential misjudged event is directly filtered out, and no companionship event is triggered; the other is weight reduction processing, for example, if the confidence level of the sports mode is between 0.6 and 0.8, then the trigger weight is reduced (such as reduced by 20% or other values) before being compared with the weight threshold to make a trigger judgment, in order to reduce the probability of false triggering.

[0069] This embodiment formulates a targeted trigger restriction strategy for candidate events that are prone to misjudgment. By filtering out low-confidence misjudgment events and reducing the trigger weight of suspicious events, it effectively reduces the misjudgment rate in scenarios such as passive carrying, vibration during rides, and accidental external contact, further improving the accuracy and reliability of companion recognition and avoiding the decline in user experience and waste of system resources caused by invalid event triggers.

[0070] In one embodiment, after step S205, the method further includes: Statistics are compiled on all target companion play events triggered within a preset time period to obtain event type statistics and trigger time statistics. Based on the event type statistics and trigger time statistics, identify whether there is an abnormal trigger. If there is an abnormal trigger, perform attenuation, marking and / or review processing on the status change data corresponding to the abnormal trigger.

[0071] In this embodiment, the identification and triggering results of the play-along event are further processed for anomaly identification. All target play-along events triggered within a preset time period are statistically analyzed in real time, for example, every 10 minutes or other times. The type, trigger time, trigger weight, and corresponding movement mode of each event within the preset time period are recorded, and event type statistics are generated, such as "play interaction event" being triggered 15 times within 10 minutes, "short adventure event" being triggered 3 times, etc. Trigger time statistics are also generated, such as the trigger time of each time and the time interval between two adjacent times.

[0072] Based on event type statistics and trigger time statistics, abnormal triggers are identified. Abnormal triggers refer to behaviors such as repeatedly triggering the same type of companion game event within a short period, triggering intervals that are too short, or triggering an abnormally large number of times. These behaviors may be due to user cheating (such as deliberately shaking the device repeatedly to farm rewards) or accidental triggering caused by device malfunction. Appropriate handling is required to avoid affecting the growth balance of the intelligent agent. Specifically, abnormal trigger judgment criteria can be preset, including: the number of times the same type of companion game event is triggered exceeds a specified number (e.g., 10 times) within a preset time period (e.g., 10 minutes); the time interval between two consecutive events of the same type of companion game event is less than a preset interval (e.g., 15 seconds); the trigger weight of a certain companion game event is continuously lower than the weight threshold but is still triggered, possibly indicating device malfunction, etc. If any judgment criterion is met, it is identified as an abnormal trigger.

[0073] For abnormal trigger times, corresponding exception handling is implemented, including attenuation, marking, and / or review of the status change data corresponding to the abnormal trigger. Specifically, reward attenuation reduces the update amount of the corresponding status parameters for abnormally triggered play-along events by a certain percentage, such as 50% to 80%, to prevent users from cheating and obtaining excessive rewards. Exception marking involves marking the abnormal trigger behavior on the server, recording the trigger time, user ID, and device ID for subsequent tracing and analysis. Review handling can trigger manual server-side review for users whose abnormal trigger count exceeds a preset number to confirm whether it constitutes cheating. If cheating is confirmed, the user's play-along event triggering privileges can be suspended for the corresponding duration.

[0074] This embodiment accurately identifies abnormal triggering behaviors caused by user cheating, equipment failure, etc., through statistical analysis of play-along events within a preset time period. It also avoids abnormal state updates from disrupting the growth balance of the intelligent agent and improves anti-cheating capabilities and operational stability through targeted processing such as reward decay, abnormal marking, and manual review.

[0075] It should be noted that there is no necessary order between the above steps. Those skilled in the art will understand from the description of the embodiments of the present invention that the above steps may have different execution orders in different embodiments, that is, they may be executed in parallel or in turn, etc.

[0076] Further reference Figure 3 As a response to the above Figure 2 The present invention provides an embodiment of an intelligent agent play-along recognition device based on multi-feature joint implementation, which is similar to the method shown. Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0077] like Figure 3 As shown, the intelligent agent companion recognition device 30 based on multi-feature joint analysis described in this embodiment includes: The data sampling module 301 is used to collect inertial sensing data from the hardware carrier bound to the target intelligent agent at a first sampling frequency, and to obtain the first inertial sensing data. The sampling switching module 302 is used to perform feature analysis on the first inertial sensing data, confirm whether the current preset triggering condition is met, and if it is met, switch to the second sampling frequency to collect inertial sensing data from the hardware carrier and obtain the second inertial sensing data. The second sampling frequency is greater than the first sampling frequency. The feature extraction module 303 is used to extract and fuse multi-dimensional features from the second inertial sensing data to obtain the fused motion features to be identified. The motion recognition module 304 is used to input the fused motion features into a pre-trained motion pattern recognition model to identify and obtain the current motion pattern; The companion play event mapping module 305 is used to trigger corresponding target companion play events based on the current movement mode and update the state parameters of the target intelligent agent.

[0078] The module referred to in this invention is a series of computer program instruction segments that can perform specific functions. It is more suitable than a program for describing the intelligent agent's play-play recognition and execution process. For the specific implementation of each module, please refer to the corresponding method embodiments above, which will not be repeated here.

[0079] In one embodiment, the sampling switching module 302 includes: The data extraction unit is used to extract acceleration data and angular velocity data from the first inertial sensing data within multiple consecutive sampling windows; The feature analysis unit is used to perform modulus change analysis, rhythm change analysis, reciprocating motion determination, and static state interruption determination on the acceleration data and angular velocity data within the multiple consecutive sampling windows. The condition confirmation unit is used to confirm whether there is at least one of the following: acceleration magnitude change greater than a first threshold, angular velocity magnitude change greater than a second threshold, periodic rhythmic change, reciprocating motion, and interruption of the stationary state. If so, the preset triggering condition is met.

[0080] In one embodiment, the feature extraction module 303 includes: The data segmentation unit is used to segment the second inertial sensing data according to a preset time window to obtain several data segments; The motion feature extraction unit is used to extract time-domain features, frequency-domain features, posture features, and rhythm features in each data segment, respectively. The context feature extraction unit is used to obtain the task state and application foreground / background state of the target intelligent agent based on the timestamp of each data segment, and extract the context features. The feature fusion unit is used to splice and fuse the temporal features, frequency features, posture features, rhythm features, and context features of each data segment to obtain the fused motion features to be identified.

[0081] In one embodiment, the motion recognition module 304 includes: A convolutional processing unit is used to input the fused motion features into the convolutional layer to extract local time series patterns and obtain corresponding local time series features. An attention unit is used to input the local time series features into the attention layer, and to perform attention weight allocation and filtering on the local time series features through an attention mechanism to obtain attention-weighted features; The recognition output unit is used to input the attention-weighted features into the classification output layer for classification output mapping, obtain the probability distribution of each motion mode, and identify the motion mode with the highest probability value as the current motion mode.

[0082] In one embodiment, the companion play event mapping module 305 includes: The mapping relationship acquisition unit is used to acquire a preset event mapping table, which stores the mapping relationship between different movement modes and different play-along events; The event mapping unit is used to map the current motion mode to a target play-along event and generate trigger weights according to the mapping relationship in the event mapping table; The event triggering unit is used to trigger the target play-along event when the trigger weight is greater than the weight threshold, and to update the state parameters of the target agent according to the triggering result of the target play-along event.

[0083] In one embodiment, the companion play event mapping module 305 further includes: The misjudgment suppression unit is used to perform trigger restriction processing on the misjudgment candidate event if the current motion mode is mapped to a specified misjudgment candidate event, thereby filtering out the misjudgment candidate event or reducing the trigger weight of the misjudgment candidate event.

[0084] In one embodiment, the device 30 further includes: The event statistics module is used to collect statistics on all target companion events triggered within a preset time period, and obtain event type statistics and trigger time statistics. The exception handling module is used to identify whether there is an exception trigger based on the event type statistics and trigger time statistics. If there is an exception trigger, the status change data corresponding to the exception trigger is attenuated, marked and / or reviewed.

[0085] In the above embodiments, this invention discloses an intelligent agent companion game recognition device based on multi-feature fusion. It acquires first inertial sensing data by collecting inertial sensing data from a hardware carrier bound to a target intelligent agent at a first sampling frequency; performs feature analysis on the first inertial sensing data to confirm whether a preset triggering condition is met; if so, it switches to a second sampling frequency to collect inertial sensing data from the hardware carrier, acquiring second inertial sensing data, where the second sampling frequency is greater than the first sampling frequency; performs multi-dimensional feature extraction and fusion on the second inertial sensing data to obtain fused motion features to be identified; inputs the fused motion features into a pre-trained motion pattern recognition model to identify the current motion pattern; triggers a corresponding target companion game event based on the current motion pattern and updates the state parameters of the target intelligent agent. This invention achieves precise linkage between real-world motion and companion game events by triggering corresponding target companion game events after motion pattern recognition through hierarchical sampling and multi-feature fusion, while considering sampling power consumption.

[0086] Specific limitations regarding the intelligent companion gaming identification device based on multi-feature combination can be found in the limitations of the intelligent companion gaming identification method based on multi-feature combination mentioned above, and will not be repeated here. Each module in the aforementioned intelligent companion gaming identification device based on multi-feature combination can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0088] Another embodiment of the present invention provides a computer device, such as... Figure 4 As shown, computer device 40 includes: One or more processors 401 and memory 402, Figure 4 The following section uses a processor 401 as an example. The processor 401 and the memory 402 can be connected via a bus or other means. Figure 4 Taking the example of a connection between China and Israel via a bus.

[0089] The processor 401 is used to perform various control logics of the computer device 40. It can be any conventional processor, microprocessor, state machine, general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), microcontroller, ARM (Acorn RISC Machine) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination of these components.

[0090] The memory 402, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions corresponding to the intelligent agent companion recognition method based on multi-feature combination in this embodiment of the invention. The processor 401 executes various functional applications and data processing of the computer device 40 by running the non-volatile software programs, instructions, and units stored in the memory 402, thereby implementing the intelligent agent companion recognition method based on multi-feature combination in the above method embodiment.

[0091] Another embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by one or more processors, perform the steps of the intelligent agent companion recognition method based on multi-feature joint in any of the above method embodiments.

[0092] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0093] Based on the above description of the embodiments, those skilled in the art will understand that the methods described in the embodiments can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0094] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The computer program can be stored in a non-volatile, computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The storage medium can be a memory, magnetic disk, floppy disk, flash memory, optical storage, etc.

[0095] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying intelligent agents for companionship based on multi-feature combination, characterized in that, include: Inertial sensing data is collected from the hardware carrier bound to the target intelligent agent at a first sampling frequency to obtain the first inertial sensing data. The first inertial sensing data is subjected to feature analysis to confirm whether the preset triggering condition is met. If it is met, the sampling frequency is switched to the second sampling frequency to collect inertial sensing data from the hardware carrier and obtain the second inertial sensing data. The second sampling frequency is greater than the first sampling frequency. Multi-dimensional feature extraction and fusion are performed on the second inertial sensing data to obtain the fused motion features to be identified; The fused motion features are input into a pre-trained motion pattern recognition model to identify the current motion pattern. Trigger the corresponding target play-along event based on the current movement pattern, and update the state parameters of the target agent; The step of performing feature analysis on the first inertial sensing data to confirm whether the preset triggering conditions are met includes: Acceleration and angular velocity data from multiple consecutive sampling windows are extracted from the first inertial sensing data; The acceleration and angular velocity data within the multiple consecutive sampling windows are analyzed for magnitude variation, rhythm variation, reciprocating motion determination, and static state interruption determination. Confirm whether there is at least one of the following: acceleration magnitude change greater than the first threshold, angular velocity magnitude change greater than the second threshold, periodic rhythmic change, reciprocating motion, and interruption of the stationary state. If so, the preset triggering condition is met. The step of extracting and fusing multi-dimensional features from the second inertial sensing data to obtain the fused motion features to be identified includes: The second inertial sensing data is segmented according to a preset time window to obtain several data segments; Within each data segment, time-domain features, frequency-domain features, attitude features, and rhythm features are extracted respectively; Based on the timestamp of each data segment, the task status and application foreground / background status of the target intelligent agent are obtained, and context features are extracted. The temporal, frequency, pose, rhythm, and contextual features of each data segment are spliced ​​and fused to obtain the fused motion features to be identified.

2. The intelligent agent companion recognition method based on multi-feature joint as described in claim 1, characterized in that, The motion pattern recognition model includes at least a convolutional layer, an attention layer, and a classification output layer. The step of inputting the fused motion features into the pre-trained motion pattern recognition model to identify the current motion pattern includes: The fused motion features are input into the convolutional layer to extract local time series patterns, thereby obtaining the corresponding local time series features. The local time series features are input into the attention layer, and attention weights are assigned and filtered through the attention mechanism to obtain attention-weighted features. The attention-weighted features are input into the classification output layer for classification output mapping to obtain the probability distribution of each motion mode, and the motion mode with the highest probability value is identified as the current motion mode.

3. The intelligent agent companion recognition method based on multi-feature joint analysis according to claim 1, characterized in that, The process of triggering a corresponding target play-along event based on the current movement pattern and updating the state parameters of the target agent includes: Obtain a preset event mapping table, which stores the mapping relationship between different movement modes and different play-along events; Based on the mapping relationship in the event mapping table, the current movement mode is mapped to the target play-along event and a trigger weight is generated; When the trigger weight is greater than the weight threshold, the target play-along event is triggered, and the state parameters of the target agent are updated according to the trigger result of the target play-along event.

4. The intelligent agent companion recognition method based on multi-feature joint analysis according to claim 3, characterized in that, After mapping the current movement mode to the target play-along event according to the mapping relationship in the event mapping table, the method further includes: If the current motion pattern is mapped to a specified misjudged candidate event, then trigger restriction processing is applied to the misjudged candidate event, either by filtering out the misjudged candidate event or reducing the trigger weight of the misjudged candidate event.

5. The intelligent agent companion recognition method based on multi-feature joint analysis according to claim 1, characterized in that, After triggering the corresponding target play-along event based on the current movement pattern and updating the state of the target agent, the method further includes: Statistics are compiled on all target companion play events triggered within a preset time period to obtain event type statistics and trigger time statistics. Based on the event type statistics and trigger time statistics, identify whether there is an abnormal trigger. If there is an abnormal trigger, perform attenuation, marking and / or review processing on the status change data corresponding to the abnormal trigger.

6. A smart agent companion recognition device based on multi-feature combination, characterized in that, include: The data sampling module is used to collect inertial sensing data from the hardware carrier bound to the target intelligent agent at a first sampling frequency, and to obtain the first inertial sensing data. The sampling switching module is used to perform feature analysis on the first inertial sensing data, confirm whether the preset triggering conditions are met, and if so, switch to the second sampling frequency to collect inertial sensing data from the hardware carrier to obtain the second inertial sensing data. The second sampling frequency is greater than the first sampling frequency. The feature extraction module is used to extract and fuse multi-dimensional features from the second inertial sensing data to obtain the fused motion features to be identified. The motion recognition module is used to input the fused motion features into a pre-trained motion pattern recognition model to identify and obtain the current motion pattern; The companion play event mapping module is used to trigger corresponding target companion play events based on the current movement mode and update the state parameters of the target intelligent agent; The sampling switching module includes: The data extraction unit is used to extract acceleration data and angular velocity data from the first inertial sensing data within multiple consecutive sampling windows; The feature analysis unit is used to perform modulus change analysis, rhythm change analysis, reciprocating motion determination, and static state interruption determination on the acceleration data and angular velocity data within the multiple consecutive sampling windows. The condition confirmation unit is used to confirm whether there is at least one of the following: acceleration modulus change greater than the first threshold, angular velocity modulus change greater than the second threshold, periodic rhythmic change, reciprocating motion, and interruption of the stationary state. If so, the preset triggering condition is met. The feature extraction module includes: The data segmentation unit is used to segment the second inertial sensing data according to a preset time window to obtain several data segments; The motion feature extraction unit is used to extract time-domain features, frequency-domain features, posture features, and rhythm features in each data segment, respectively. The context feature extraction unit is used to obtain the task state and application foreground / background state of the target intelligent agent based on the timestamp of each data segment, and extract the context features. The feature fusion unit is used to splice and fuse the temporal features, frequency features, posture features, rhythm features, and context features of each data segment to obtain the fused motion features to be identified.

7. A computer device, characterized in that, Includes at least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the intelligent agent companion recognition method based on multi-feature joint as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by one or more processors, cause the one or more processors to perform the intelligent agent companion recognition method based on multi-feature combination as described in any one of claims 1-5.

Images