Tourist interpersonal interaction body language automatic generation system based on deep learning

Abstract

This invention relates to the field of intelligent service robot perception technology, and discloses an automatic generation system for tourist interpersonal interaction body language based on deep learning. The system acquires scene images through an image acquisition and skeleton extraction module and extracts the coordinate sequences of all two-dimensional skeletons of people, storing them in a database. A joint motion feature extraction module calculates the changes in joint coordinates frame by frame and inputs them into a trained deep learning model. An action category differentiation module classifies continuous actions into discrete actions such as waving, approaching, standing, leaving, and passing by. A depth information acquisition module simultaneously acquires distance information of people. A target service priority calculation module integrates action category, depth information, and action sequence to calculate and sort the service priority score for each person, generating a target service priority queue. This invention can automatically parse tourist interaction intentions and intelligently prioritize services.

Description

Technical Field

[0001] This invention relates to the field of intelligent service robot perception technology, specifically to a deep learning-based automatic generation system for tourist interpersonal interaction body language. Background Technology

[0002] Existing tourist service or human-computer interaction systems often employ general human motion recognition technologies based on computer vision. These technologies can detect basic movements such as walking and running, or locate key points on the human skeleton. However, the recognition categories defined by these technologies are typically broad, lacking a detailed characterization of interpersonal interaction intentions in specific service scenarios. General models struggle to accurately distinguish between "intentional approach" and "unintentional passing by," and cannot parse continuous sequences of limb movements into discrete events with clear interactive semantics. This results in the system being insensitive to service request signals initiated by tourists, leading to delayed responses.

[0003] In determining service priority in multi-user scenarios, existing solutions often rely on single-dimensional information. Some methods attempt to combine simple trajectory tracking, but fail to deeply integrate and calculate the semantic meaning of behavior, precise spatial distance, and the temporal logic of actions. The judgments generated by these methods are one-sided and static, unable to dynamically quantify and prioritize the urgency of service needs that change in real time for different individuals, and thus difficult to directly drive efficient and reasonable proactive allocation of service resources. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based automatic generation system for tourist interpersonal interaction body language to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a deep learning-based automatic generation system for tourist interpersonal interaction body language, the system comprising: The image acquisition and skeleton extraction module is used to acquire target scene images containing multiple people, convert all the images of people in the target scene images into two-dimensional skeleton coordinate sequences through a skeleton estimation model, and store all skeleton coordinate sequences in a skeleton database. The joint motion feature extraction module is used to extract the coordinate changes of all joints in each skeleton coordinate sequence frame by frame from the skeleton database, and input the coordinate changes of all joints into the trained deep learning model in chronological order. The action category differentiation module is used to process the input joint coordinate changes based on the deep learning model, and to classify the continuous actions of the person in the target scene image into a number of preset discrete action categories, including waving, approaching, standing, leaving and passing by. The depth information acquisition module is used to synchronously acquire depth information of the target scene image, wherein the depth information includes the distance of each person in the target scene image relative to the acquisition device; The target service priority calculation module is used to calculate the service priority score of each character based on the discrete action category, the depth information and the time order of the action, and sort all characters according to the service priority score to generate a target service priority queue.

[0006] Preferably, the image acquisition and skeleton extraction module includes: The image acquisition unit is used to continuously capture video streams within a single service environment to form raw video stream data. The person detection unit is used to detect people in each frame of the original video stream data, locate the bounding box positions of all target people within the image range, and form multi-person position information. The skeleton transformation unit is used to input each frame of image containing the position information of the multiple characters into the skeleton estimation model, and the skeleton estimation model outputs two-dimensional skeleton data of each character containing the coordinates of multiple joint points; The skeleton storage unit is used to organize the two-dimensional skeleton data of all characters in each frame output by the skeleton conversion unit according to the timestamp and character identifier, and store it in the skeleton database.

[0007] Preferably, the joint motion feature extraction module includes: The data retrieval unit is used to sequentially read multiple consecutive frames of two-dimensional skeleton data of the same person identifier under different timestamps from the skeleton database in chronological order. The change calculation unit is used to calculate the coordinate difference of each joint point between adjacent frames of continuous multi-frame two-dimensional skeleton data read by the data retrieval unit, and generate a joint coordinate change vector for each frame. The sequence construction unit is used to concatenate multiple joint coordinate change vectors belonging to the same person and arranged in chronological order, generated by the change calculation unit, into a time series feature vector. The feature input unit is used to input the time series feature vector into the trained deep learning model.

[0008] Preferably, the target service priority calculation module includes: The action sequence parsing unit is used to receive the discrete action categories output by the action category differentiation module, and construct an action category sequence with timestamps for each character according to the time order in which the actions occur. The spatial distance mapping unit is used to receive the depth information output by the depth information acquisition module and associate the distance value of each character in the depth information with the discrete action category of the character at the corresponding time. The priority score evaluation unit is used to combine the action category sequence provided by the action sequence parsing unit and the associated distance value provided by the spatial distance mapping unit to calculate the service priority score for each person according to preset rules. Among them, waving actions and approaching actions are given higher base score weights, the smaller the distance value, the higher the score bonus, and the closer the action occurs, the higher the score bonus. The sorting queue generation unit is used to sort all people in descending order based on the priority score calculated by the priority score evaluation unit, and generate the target service priority queue.

[0009] Preferably, the system further includes: The reinforcement learning decision module is used to select the target person with the highest service priority in the target service priority queue as the current interaction object, input the current environment state, the action category sequence of the target person and the depth information into the action decision model based on Q learning, and deduce the type of body language response that the system should generate. The body language generation module is used to generate control instructions that drive virtual characters or service robots to perform corresponding body movements by calling the corresponding preset three-dimensional motion template based on the body language response type derived by the reinforcement learning decision module.

[0010] Preferably, the trained deep learning model includes: A feature encoding network is used to receive the time series feature vector and extract the spatiotemporal features of joint motion implicit in the time series feature vector through multiple convolutional layers, and output a high-level feature map. A temporal context modeling network is used to receive high-level feature maps output by the feature encoding network, and to model the dependencies of the high-level feature maps in the temporal dimension through a recurrent neural network layer, outputting feature vectors containing temporal context information. The classification output layer receives the feature vector containing temporal context information output by the temporal context modeling network, and maps the feature vector to a probability distribution prediction of the multiple discrete action categories through a fully connected layer and a normalization function. The category with the highest probability is output as the discrete action category at the current time.

[0011] Preferably, the Q-learning-based action decision model includes: The state construction unit is used to construct the current environment state vector by combining the identity of the first person in the target service priority queue at the current moment, the discrete action categories of the first person at the most recent moments, the current depth distance value of the first person, and the body language response type executed by the system at the previous moment. A motion value assessment network is used to receive the current environmental state vector and calculate the Q value corresponding to each preset body language response type that the system can execute through a deep neural network. The decision output unit is used to select the preset body language response type with the largest Q value from all Q values ​​calculated by the action value evaluation network, as the optimal response action derived by the reinforcement learning decision module.

[0012] Preferably, the system further includes: An experience replay buffer is used to store state transition samples generated by the reinforcement learning decision-making module during historical interactions. Each sample contains a historical environmental state vector, the type of body language response executed, the immediate reward obtained, a new environmental state vector, and a flag indicating whether the interaction has terminated. The target network parameter update module is used to periodically copy network parameters from the action value evaluation network to generate a target evaluation network with the same structure as the action value evaluation network but with frozen parameters. The target evaluation network is used to calculate the target Q value corresponding to the new environmental state vector in the state transition sample.

[0013] Preferably, the system further includes: The model training module is used to randomly sample a batch of state transition samples from the experience replay buffer pool, and use the target evaluation network and the action value evaluation network provided by the target network parameter update module to calculate the time difference error between the predicted Q value and the target Q value. The network parameter optimization unit is used to update the parameters of the action value evaluation network using the backpropagation algorithm based on the time difference error calculated by the model training module, so as to minimize the time difference loss.

[0014] Preferably, the body language generation module includes: The action type matching unit is used to receive the body language response type derived by the reinforcement learning decision module, and retrieve a preset three-dimensional motion template that completely matches the body language response type from the preset action template library. The preset three-dimensional motion template defines the motion trajectory of each joint of the virtual character or service robot within a fixed period. The control instruction synthesis unit is used to convert the joint motion trajectory data in the retrieved preset three-dimensional motion template into low-level control protocol instructions to drive the target virtual character or service robot. The instruction output unit is used to send the control protocol instructions generated by the control instruction synthesis unit to the virtual character animation engine or the service robot motion controller according to a preset execution sequence.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By defining a set of discrete action categories directly related to service interaction scenarios, such as waving, approaching, standing, leaving, and passing by, and using a deep learning model to analyze the time series of joint coordinate changes, the system can parse continuous actions in the original video stream into semantic units with clear interpersonal interaction intentions. This allows the system to go beyond general pose estimation, accurately identifying the behavioral patterns of tourists actively seeking interaction, and recognizing action sequences as a complete expression of a service request intention, rather than fragmented action segments.

[0016] A dynamic priority calculation model is constructed by synchronously fusing three dimensions: the behavioral semantics represented by discrete action categories, the precise spatial distance provided by depth information, and the temporal order of action occurrence. The model outputs a quantified service priority score for each individual in the scene and generates an executable service queue accordingly. This achieves the transformation from multimodal perception information to direct decision-making instructions, enabling the service system to automatically determine the optimal order of service responses based on a dynamic standard that integrates factors such as "who has intent," "who is closer," and "who requested earlier." Attached Figure Description

[0017] Figure 1 This is a timing diagram of the deep learning-based automatic generation system for tourist interpersonal interaction body language as described in this invention. Figure 2 This is a flowchart of the image acquisition and skeleton extraction module; Figure 3 A flowchart for the target service priority calculation module; Figure 4 A graph showing the average target Q-value variation during the optimization phase under different discount factors; Figure 5 Heatmap of execution time for each stage of the body language generation module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a deep learning-based automatic generation system for tourist interpersonal interaction body language. The system includes: an image acquisition and skeleton extraction module, a joint motion feature extraction module, an action category differentiation module, a depth information acquisition module, and a target service priority calculation module. The image acquisition and skeleton extraction module acquires target scene images containing multiple people, converts all person images in the target scene images into a two-dimensional skeleton coordinate sequence using a skeleton estimation model, and stores it in a skeleton database. The joint motion feature extraction module extracts the coordinate changes of all joints in each skeleton coordinate sequence frame by frame from the skeleton database, and inputs the coordinate changes of all joints in chronological order into a trained system. The deep learning model and action category differentiation module, based on the processing of input joint coordinate changes by the deep learning model, classify the continuous actions of people in the target scene image into a variety of preset discrete action categories. Discrete action categories include waving, approaching, standing, leaving, and passing by. The depth information acquisition module synchronously collects the depth information of the target scene image, which includes the distance of each person in the target scene image relative to the acquisition device. The target service priority calculation module calculates the service priority score of each person according to the discrete action category, depth information, and the time sequence of the action occurrence. Based on the service priority score, all people are sorted to generate a target service priority queue.

[0020] In one embodiment of the present invention, see [reference] Figure 2 The image acquisition and skeleton extraction module includes an image acquisition unit, a person detection unit, a skeleton conversion unit, and a skeleton storage unit. The image acquisition unit continuously captures video streams within a single service environment to form raw video stream data. The person detection unit performs person detection on each frame of the raw video stream data, locates the bounding box positions of all target persons within the image range, and forms multi-person position information. The skeleton conversion unit inputs each frame of the image containing multi-person position information into the skeleton estimation model. The skeleton estimation model outputs two-dimensional skeleton data of each person containing multiple key point coordinates. The skeleton storage unit organizes the two-dimensional skeleton data of all persons in each frame output by the skeleton conversion unit according to timestamps and person identifiers, and stores it in the skeleton database. The joint motion feature extraction module includes a data retrieval unit, a change calculation unit, a sequence construction unit, and a feature input unit. The data retrieval unit sequentially reads multiple consecutive frames of two-dimensional skeleton data of the same person at different timestamps from the skeleton database according to time order. The change calculation unit calculates the coordinate difference of each joint point between adjacent frames of the consecutive frames of two-dimensional skeleton data read by the data retrieval unit, generating a joint coordinate change vector for each frame. The sequence construction unit concatenates multiple joint coordinate change vectors of the same person that are arranged in time order and generated by the change calculation unit into a time series feature vector. The feature input unit inputs the time series feature vector into the trained deep learning model.

[0021] The trained deep learning model includes a feature encoding network, a temporal context modeling network, and a classification output layer. The feature encoding network receives time-series feature vectors and extracts the spatiotemporal features of joint motion implicit in the time-series feature vectors through multiple convolutional layers, outputting a high-level feature map. The temporal context modeling network receives the high-level feature map output by the feature encoding network and models the dependencies of the high-level feature map in the temporal dimension through recurrent neural network layers, outputting a feature vector containing temporal context information. The classification output layer receives the feature vector containing temporal context information output by the temporal context modeling network and maps the feature vector to a probability distribution prediction of multiple discrete action categories through fully connected layers and a normalization function. The category with the highest probability is output as the discrete action category at the current moment.

[0022] In practical implementation, the image acquisition unit in the image acquisition and skeleton extraction module continuously captures video streams within a single service environment, forming raw video stream data. This raw video stream data is input to the person detection unit at a rate of 30 frames per second. The person detection unit performs person detection on each frame of the raw video stream data, locating the bounding boxes of all target persons within the image range, forming multi-person position information. The bounding box position information includes the pixel coordinates of the upper left and lower right corners of the rectangle. In practical implementation, the skeleton conversion unit receives each frame containing multi-person position information and inputs the person sub-images cropped from each bounding box region one by one into the skeleton estimation model. The skeleton estimation model outputs two-dimensional skeleton data for each person, containing coordinates of multiple joint points. The two-dimensional skeleton data includes the horizontal and vertical coordinates of key human joints such as the top of the head, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. In some embodiments, the skeleton storage unit organizes the two-dimensional skeleton data of all characters in each frame output by the skeleton conversion unit according to the timestamp and character identifier, and stores it in the skeleton database. Each record in the database contains a timestamp field, a character identifier field, and a key point coordinate sequence field composed of multiple floating-point numbers.

[0023] In practical implementation, the data retrieval unit of the joint motion feature extraction module sequentially reads multiple consecutive frames of two-dimensional skeleton data of the same character at different timestamps from the skeleton database according to time sequence, with a fixed time window length of 16 frames. The change calculation unit calculates the coordinate difference of each joint point between adjacent frames of the consecutive frames of two-dimensional skeleton data read by the data retrieval unit, generating a joint coordinate change vector for each frame. For frame t, the joint coordinate change vector is obtained by element-wise subtracting the joint coordinates of frame t from those of frame t-1. The sequence construction unit concatenates multiple joint coordinate change vectors generated by the change calculation unit, belonging to the same character and arranged in time sequence, along the feature dimension to form a time-series feature vector, representing the character's motion pattern over a period of time. The feature input unit inputs the time-series feature vector into the trained deep learning model.

[0024] In practical implementation, the feature encoding network of the trained deep learning model receives time-series feature vectors and extracts the spatiotemporal features of joint motion implicit in the time-series feature vectors through multiple convolutional layers, outputting a high-level feature map. The operations of the convolutional layers can capture the local correlations of joint motion in the spatial and temporal dimensions. The temporal context modeling network receives the high-level feature map output by the feature encoding network and models the temporal dependencies of the high-level feature map through a long short-term memory network layer, outputting a feature vector containing temporal context information. The classification output layer receives the feature vector containing temporal context information output by the temporal context modeling network and maps the feature vector to a probability distribution prediction for multiple discrete action categories through a fully connected layer and a softmax normalization function. The category with the highest probability is output as the discrete action category at the current moment. It can be understood that the feature encoding network and the temporal context modeling network use skeleton sequence data labeled with action categories for supervised learning during the training phase. In some embodiments, the time-series feature vectors need to be standardized before being input into the model. Standardization ensures that the input data meets the distribution characteristics of zero mean and unit variance. Optionally, the formula for standardization is: ; in: This represents the vector of joint coordinate changes for the object identified as i at time step t. This represents the mean of the vector of changes in all joint coordinates of the character over a period of time. This represents the corresponding standard deviation. This is the standardized vector of joint coordinate changes.

[0025] In one embodiment of the present invention, see [reference] Figure 3The target service priority calculation module includes an action sequence parsing unit, a spatial distance mapping unit, a priority score evaluation unit, and a sorting queue generation unit. The action sequence parsing unit receives discrete action categories output by the action category differentiation module and constructs a timestamped action category sequence for each person according to the time sequence of action occurrence. The spatial distance mapping unit receives depth information output by the depth information acquisition module and associates the distance value of each person in the depth information with the discrete action category of the person at the corresponding time. The priority score evaluation unit combines the action category sequence provided by the action sequence parsing unit and the associated distance value provided by the spatial distance mapping unit, and calculates the service priority score for each person according to preset rules. Among them, waving and approaching actions are given higher base score weights. The smaller the distance value, the higher the score bonus. The closer the action occurrence time, the higher the score bonus. The sorting queue generation unit sorts all people in descending order according to the service priority score calculated by the priority score evaluation unit to generate a target service priority queue.

[0026] In specific implementations, the action sequence parsing unit receives discrete action categories output by the action category differentiation module. These discrete action categories include waving, approaching, standing, leaving, and passing by. The action sequence parsing unit constructs a timestamped action category sequence for each person according to the chronological order of the actions. In specific implementations, the spatial distance mapping unit receives depth information output by the depth information acquisition module. This depth information includes the distance value of each person in the target scene image relative to the acquisition device. The spatial distance mapping unit associates the distance value of each person in the depth information with the person's discrete action category at the corresponding time, forming an action record with distance annotations. In some embodiments, the output data of the action sequence parsing unit and the spatial distance mapping unit are cached in a temporary data structure. This structure uses the person's identifier as the primary key and stores the latest records for that person, each with a timestamp, action category, and corresponding distance value.

[0027] In practice, the priority score evaluation unit combines the action category sequence provided by the action sequence analysis unit and the associated distance value provided by the spatial distance mapping unit to calculate the service priority score for each person according to preset rules. Waving and approaching actions are assigned higher base score weights; the smaller the distance value, the higher the score bonus; and the closer the action occurred, the higher the score bonus. The priority score evaluation unit's calculation process iterates through all persons to be evaluated, comprehensively evaluating the most recent N action records for each person. The preset rules are implemented through a scoring function that comprehensively considers the base weight of the action type, the distance decay factor, and the time decay factor. The scoring function is designed based on the timestamped action category sequence provided by the action sequence analysis unit and the associated distance value provided by the spatial distance mapping unit to calculate the service priority score for each person to be evaluated. Specifically, the scoring function iterates through several of the most recent action records for each person; each record contains a discrete action category, the corresponding distance value, and the timestamp of the occurrence. For each record, the function first assigns a base score weight based on the action category; for example, waving and approaching actions receive higher weights due to their clear interaction intent. Then, a distance decay factor is introduced, giving records with smaller distance values ​​a higher bonus in the calculation, reflecting the principle of prioritizing proximity. Simultaneously, a time decay factor is applied to ensure that records with more recent actions contribute more, reflecting the timeliness of the demand. Finally, the overall priority score for that person is obtained by summing the results of multiplying the base weight of each record by the distance and time decay factors.

[0028] Optionally, for a given person identified as i, its service priority score P i The calculation formula is as follows: ; Where: P i This represents the priority score of the serviced action, where N represents the number of furthest action records considered, and a k This represents the discrete action category in the k-th record. d represents the preset base weight coefficient corresponding to this action category. k The distance value associated with the k-th record. It is a distance attenuation coefficient between 0 and 1, t k It is the timestamp of the k-th record, t current It is the current time. This is a time decay coefficient between 0 and 1. In some embodiments, the base weight coefficient for waving is set to 1.0, for approaching actions to 0.8, for standing actions to 0.3, and for leaving and passing actions to 0.1. Distance decay coefficient Set to 0.95, time decay coefficient The coefficients are set to 0.9, and their specific values ​​are determined through configuration before system deployment.

[0029] In practice, the sorting queue generation unit sorts all personnel in descending order based on the service priority score calculated by the priority score evaluation unit, generating a target service priority queue. The target service priority queue is an ordered list, where each element contains a personnel identifier and its corresponding service priority score. The head of the queue represents the personnel with the highest current service priority score. The sorting queue generation unit updates the target service priority queue at the end of each calculation cycle and sends the queue data to subsequent modules.

[0030] In one embodiment of the present invention, the system further includes a reinforcement learning decision module and a body language generation module. The reinforcement learning decision module selects the target person with the highest service priority in the target service priority queue as the current interaction object, inputs the current environmental state, the action category sequence of the target person and depth information into the action decision model based on Q-learning, and derives the body language response type that the system should generate. The body language generation module calls the corresponding preset three-dimensional motion template according to the body language response type derived by the reinforcement learning decision module, and generates control instructions to drive the virtual character or service robot to perform corresponding body movements. The Q-learning-based action decision-making model includes a state construction unit, an action value evaluation network, and a decision output unit. The state construction unit constructs the current environment state vector by combining the identity of the first person in the target service priority queue at the current moment, the discrete action categories of the first person at the current moment, the current depth distance value of the first person, and the body language response type executed by the system at the previous moment. The action value evaluation network receives the current environment state vector and calculates the Q value corresponding to each preset body language response type that the system can execute through a deep neural network. The decision output unit selects the preset body language response type with the largest Q value from all the Q values ​​calculated by the action value evaluation network as the optimal response action derived by the reinforcement learning decision module.

[0031] In practical implementation, the reinforcement learning decision-making module selects the target person with the highest service priority in the target service priority queue as the current interaction object. The module inputs the current environmental state, the target person's action category sequence, and depth information into a Q-learning-based action decision model. This model then derives the type of body language response the system should generate. The body language generation module, based on the body language response type derived by the reinforcement learning decision-making module, calls the corresponding preset 3D motion template to generate control commands that drive the virtual character or service robot to perform the corresponding body language actions. In practical implementation, the state construction unit of the Q-learning-based action decision model constructs the current environmental state vector by combining the identity of the first person in the target service priority queue at the current moment, the discrete action categories of the first person at the current moment, the current depth distance value of the first person, and the body language response type executed by the system at the previous moment.

[0032] In specific implementation, the action value evaluation network receives the current environment state vector output by the state construction unit and calculates the Q-value corresponding to each preset body language response type that the system can execute through a deep neural network. The preset body language response types include actions such as "waving in response," "nodding in greeting," "turning to guide," "waving to refuse," and "waiting still" that the system can execute. The deep neural network consists of multiple fully connected layers, with the number of neurons in its output layer being the same as the number of preset body language response types. The output value of each neuron represents the estimated cumulative reward for executing the corresponding type of body language response given the current environment state vector. The decision output unit selects the preset body language response type with the largest Q-value from all the Q-values ​​calculated by the action value evaluation network, as the optimal response action derived by the reinforcement learning decision module. It can be understood that the decision output unit can adopt an ε-greedy strategy when selecting actions, that is, selecting the action with the largest Q-value most of the time and randomly selecting an action a small portion of the time to facilitate exploration. In some embodiments, the calculation process of the action value evaluation network can be expressed as a function mapping the state vector to the action value vector; optionally, its core relationship can be represented as: ; in: This represents the current environment state vector generated by the state building unit. The parameter is The forward propagation computation process of a deep neural network, This represents the output action value vector, where each component... The Q value corresponds to selecting a specific body language response type a in state s.

[0033] In some embodiments, when constructing the current environment state vector, the state construction unit performs one-hot encoding on the discrete action categories of the first character at recent moments, normalizes the depth distance values, and also performs one-hot encoding on the body language response type executed in the previous moment. Finally, all encoded values ​​are concatenated with the character identifier index to form a one-dimensional vector. It can be understood that the specific dimensions and composition order of the current environment state vector are defined and maintained fixed during system initialization.

[0034] In one embodiment of the present invention, the system further includes an experience replay buffer and a target network parameter update module. The experience replay buffer stores state transition samples generated by the reinforcement learning decision-making module during historical interactions. Each sample includes a historical environment state vector, the type of body language response executed, the immediate reward obtained, a new environment state vector, and a flag indicating whether the interaction has terminated. The target network parameter update module periodically copies network parameters from the action value evaluation network to generate a target evaluation network with the same structure as the action value evaluation network but with frozen parameters. The target evaluation network is used to calculate the target Q-value corresponding to the new environment state vector in the state transition sample. The system also includes a model training module and a network parameter optimization unit. The model training module randomly samples a batch of state transition samples from the experience replay buffer and uses the target evaluation network and action value evaluation network provided by the target network parameter update module to calculate the temporal difference error between the predicted Q-value and the target Q-value. The network parameter optimization unit uses the backpropagation algorithm to update the parameters of the action value evaluation network based on the temporal difference error calculated by the model training module to minimize the temporal difference loss.

[0035] In specific implementations, the system also includes an experience replay buffer and a target network parameter update module. The experience replay buffer stores state transition samples generated by the reinforcement learning decision-making module during historical interactions. In each implementation, each state transition sample includes a historical environment state vector, the type of body language response executed, the immediate reward received, a new environment state vector, and a flag indicating whether the interaction has terminated. The immediate reward is provided by the environment based on the interaction effect; for example, when the body language response type generated by the system receives a positive response from the target person, a positive reward value is awarded. The experience replay buffer uses a first-in, first-out queue data structure with a fixed capacity. When new samples are added and the capacity is exceeded, the earliest added samples are removed to maintain the buffer size. In some embodiments, refer to Table 1 for the specific data structure of the state transition samples stored in the experience replay buffer.

[0036] Table 1: Data structure table of state transition samples in the experience replay buffer.

[0037]

[0038] In practice, the target network parameter update module periodically copies network parameters from the action value evaluation network to generate a target evaluation network with the same structure but frozen parameters. Parameter freezing means that the parameters of the target evaluation network remain fixed within an update cycle and do not change in real time during training. The target evaluation network is used to calculate the target Q-value corresponding to the new environmental state vector in the state transition samples. The calculation of the target Q-value provides a stable target signal for temporal difference learning.

[0039] In specific implementations, the system also includes a model training module and a network parameter optimization unit. The model training module randomly samples a batch of state transition samples from the experience replay buffer, with the batch size fixed at B. The model training module uses the target evaluation network and action value evaluation network provided by the target network parameter update module to calculate the time difference error between the predicted Q-value and the target Q-value. For a sampled state transition sample, its time difference error reflects the accuracy of the current action value evaluation network in estimating the state-action value. In some embodiments, the process of calculating the time difference error by the model training module is based on the Bellman optimal equation. Optionally, for the time difference error of a single sample... The calculation is as follows: ; in: Let s represent the temporal difference error of a single sample, s represent the historical environmental state vector, a represent the type of body language response performed, r represent the immediate reward obtained, s' represent the new environmental state vector, and Q(s,a) represent the value prediction of the action value evaluation network for the historical state-action pair. target (s', a') represents the Q-value calculation of the target evaluation network for all actions a' under the new environmental state vector. Take the largest Q value. It is a discount factor between 0 and 1, used to measure the present value of future returns.

[0040] In practice, the network parameter optimization unit uses the backpropagation algorithm to update the parameters of the action value evaluation network based on the temporal difference error calculated by the model training module, in order to minimize the temporal difference loss. The temporal difference loss is typically defined as the squared mean of the temporal difference errors of all samples within a batch. The network parameter optimization unit calls an optimizer, such as a stochastic gradient descent optimizer or an Adam optimizer, to calculate the gradient and perform the parameter updates.

[0041] See Figure 4In the Q-value calculation optimization stage, the variation characteristics of the average target Q-value are directly related to the value of the discount factor (γ). Specifically, the average target Q-value shows a trend of first increasing and then decreasing with the change of the discount factor: when the discount factor is in the range of 0.1 to 0.6, the average target Q-value shows a continuous growth trend, reaching a peak at γ=0.6; when the discount factor exceeds 0.6, the average target Q-value gradually decreases. The shaded area in the figure represents the fluctuation range of the Q-value. Combined with the target Q-value calculation logic of Q-learning (the target Q-value is composed of the discount sum of the immediate reward and the maximum Q-value of the future state), this trend reflects the balancing effect of the discount factor on the "immediate reward weight" and the "present value of future returns": a smaller discount factor weakens the impact of future returns, resulting in limited Q-value growth; a moderate discount factor (such as 0.6) achieves the optimal trade-off between immediate and future returns, pushing the Q-value to its peak; while an excessively large discount factor over-amplifies the uncertainty of future states, causing a decrease in the estimated stability of the target Q-value. In terms of parameters, the experimental configuration corresponding to the data in the figure is as follows: the batch size of the state transition sample is fixed, the update cycle of the target network parameters remains consistent, and the discount factor is adjusted from 0.1 to 0.9 in steps of 0.1.

[0042] In one embodiment of the present invention, the body language generation module includes an action type matching unit, a control instruction synthesis unit, and an instruction output unit. The action type matching unit receives the body language response type derived by the reinforcement learning decision module and retrieves a preset three-dimensional motion template that perfectly matches the body language response type from a preset action template library. The preset three-dimensional motion template defines the motion trajectory of each joint of the virtual character or service robot within a fixed period. The control instruction synthesis unit converts the joint motion trajectory data in the retrieved preset three-dimensional motion template into underlying control protocol instructions that drive the target virtual character or service robot. The instruction output unit sends the control protocol instructions generated by the control instruction synthesis unit to the virtual character animation engine or service robot motion controller according to a preset execution sequence.

[0043] In practical implementation, the body language generation module includes an action type matching unit, a control instruction synthesis unit, and an instruction output unit. The action type matching unit receives the body language response type derived by the reinforcement learning decision module. The action type matching unit retrieves a preset 3D motion template from a preset action template library that perfectly matches the body language response type. The preset 3D motion template defines the motion trajectory of each joint of the virtual character or service robot within a fixed period. In practical implementation, the preset action template library is stored in the system's non-volatile memory. Each preset 3D motion template is bound to a unique body language response type identifier. The data structure of the preset 3D motion template includes joint identifiers, time series, and the rotation angle or position coordinates of each joint at each time point.

[0044] The control command synthesis unit converts the joint motion trajectory data retrieved from the preset 3D motion template into low-level control protocol commands to drive the target virtual character or service robot. The conversion process varies depending on the target execution platform's protocol. For a virtual character animation engine, the control protocol command is a set of keyframe animation data or bone transformation commands; for a service robot motion controller, the control protocol command is the target angle sequence or speed command for each joint servo motor. In some embodiments, the data conversion performed by the control command synthesis unit is a mapping process; optionally, its core relationship can be represented as: ; Where: T represents the retrieved preset 3D motion template data, This represents the set of control protocol parameters related to the target platform, and g represents the conversion function from template data to control commands. This represents the generated control protocol command data stream. The specific implementation of the transformation function g depends on the application programming interface specification of the target virtual character animation engine or service robot motion controller.

[0045] The instruction output unit sends the control protocol instructions generated by the control instruction synthesis unit to the virtual character animation engine or service robot motion controller according to a preset execution sequence. The preset execution sequence is determined by the timeline defined in the preset 3D motion template, and the instruction output unit ensures that the control instructions at each time point are accurately delivered within the corresponding system clock cycle. In some embodiments, the instruction output unit streams the control protocol instructions to the target device via a high-speed data bus or network socket. It is understood that the instruction output unit encapsulates the instruction data before sending the instructions, adding necessary frame headers, checksums, or sequence number information to ensure reliable transmission. In a specific implementation, when the body language response type is "waving response," the action type matching unit retrieves the corresponding waving action preset 3D motion template, and the control instruction synthesis unit generates control instructions based on the template to make the robot's right arm complete a trajectory from drooping to raising and swinging several times. The instruction output unit sends these instructions to the robot's joint controller at millisecond intervals. It is understood that the motion trajectory data of the preset 3D motion template is usually pre-made and optimized offline using motion capture or keyframe animation design software.

[0046] See Figure 5In the execution flow of the body language generation module, the time consumption distribution of each functional unit in different sub-steps is quantified using this heatmap: the vertical dimension of the graph represents the three core units of action type matching, control instruction synthesis, and instruction output; the horizontal dimension corresponds to the five sub-steps of template retrieval, data conversion, instruction encapsulation, network transmission, and execution feedback; and the color depth (and the color level on the right) corresponds to the time consumption of each step (unit: ms). Specifically, the control instruction synthesis unit has the longest time consumption in the "data conversion" sub-step (8.7ms), making it the core time-consuming step of the entire module; the time consumption of the action type matching unit is concentrated in the "data conversion" step (5.2ms); and the time consumption of the instruction output unit is mainly distributed in the "network transmission" (4.2ms) and "instruction encapsulation" (3.5ms) steps. The time consumption of other sub-steps (such as template retrieval) in each unit is at a relatively low level (≤0.9ms). This time consumption distribution characteristic reflects that data conversion and network transmission are key nodes for performance optimization of the body language generation module, and targeted speed-ups can be made for the algorithms or transmission methods of these two steps in the future.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based automatic generation system for tourist interpersonal interaction body language, characterized in that, The system includes: The image acquisition and skeleton extraction module is used to acquire target scene images containing multiple people, convert all the images of people in the target scene images into two-dimensional skeleton coordinate sequences through a skeleton estimation model, and store all skeleton coordinate sequences in a skeleton database. The joint motion feature extraction module is used to extract the coordinate changes of all joints in each skeleton coordinate sequence frame by frame from the skeleton database, and input the coordinate changes of all joints into the trained deep learning model in chronological order. The action category differentiation module is used to process the input joint coordinate changes based on the deep learning model, and to classify the continuous actions of the person in the target scene image into a number of preset discrete action categories, including waving, approaching, standing, leaving and passing by. The depth information acquisition module is used to synchronously acquire depth information of the target scene image, wherein the depth information includes the distance of each person in the target scene image relative to the acquisition device; The target service priority calculation module is used to calculate the service priority score of each character based on the discrete action category, the depth information and the time order of the action, and sort all characters according to the service priority score to generate a target service priority queue.

2. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 1, characterized in that, The image acquisition and skeleton extraction module includes: The image acquisition unit is used to continuously capture video streams within a single service environment to form raw video stream data. The person detection unit is used to detect people in each frame of the original video stream data, locate the bounding box positions of all target people within the image range, and form multi-person position information. The skeleton transformation unit is used to input each frame of image containing the position information of the multiple characters into the skeleton estimation model, and the skeleton estimation model outputs two-dimensional skeleton data of each character containing the coordinates of multiple joint points; The skeleton storage unit is used to organize the two-dimensional skeleton data of all characters in each frame output by the skeleton conversion unit according to the timestamp and character identifier, and store it in the skeleton database.

3. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 2, characterized in that, The joint motion feature extraction module includes: The data retrieval unit is used to sequentially read multiple consecutive frames of two-dimensional skeleton data of the same person identifier under different timestamps from the skeleton database in chronological order. The change calculation unit is used to calculate the coordinate difference of each joint point between adjacent frames of continuous multi-frame two-dimensional skeleton data read by the data retrieval unit, and generate a joint coordinate change vector for each frame. The sequence construction unit is used to concatenate multiple joint coordinate change vectors belonging to the same person and arranged in chronological order, generated by the change calculation unit, into a time series feature vector. The feature input unit is used to input the time series feature vector into the trained deep learning model.

4. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 1, characterized in that, The target service priority calculation module includes: The action sequence parsing unit is used to receive the discrete action categories output by the action category differentiation module, and construct an action category sequence with timestamps for each character according to the time order in which the actions occur. The spatial distance mapping unit is used to receive the depth information output by the depth information acquisition module and associate the distance value of each character in the depth information with the discrete action category of the character at the corresponding time. The priority score evaluation unit is used to combine the action category sequence provided by the action sequence parsing unit and the associated distance value provided by the spatial distance mapping unit to calculate the service priority score for each person according to preset rules. Among them, waving actions and approaching actions are given higher base score weights, the smaller the distance value, the higher the score bonus, and the closer the action occurs, the higher the score bonus. The sorting queue generation unit is used to sort all people in descending order based on the priority score calculated by the priority score evaluation unit, and generate the target service priority queue.

5. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 1, characterized in that, The system also includes: The reinforcement learning decision module is used to select the target person with the highest service priority in the target service priority queue as the current interaction object, input the current environment state, the action category sequence of the target person and the depth information into the action decision model based on Q learning, and deduce the type of body language response that the system should generate. The body language generation module is used to generate control instructions that drive virtual characters or service robots to perform corresponding body movements by calling the corresponding preset three-dimensional motion template based on the body language response type derived by the reinforcement learning decision module.

6. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 3, characterized in that, The trained deep learning model includes: A feature encoding network is used to receive the time series feature vector and extract the spatiotemporal features of joint motion implicit in the time series feature vector through multiple convolutional layers, and output a high-level feature map. A temporal context modeling network is used to receive high-level feature maps output by the feature encoding network, and to model the dependencies of the high-level feature maps in the temporal dimension through a recurrent neural network layer, outputting feature vectors containing temporal context information. The classification output layer receives the feature vector containing temporal context information output by the temporal context modeling network, and maps the feature vector to a probability distribution prediction of the multiple discrete action categories through a fully connected layer and a normalization function. The category with the highest probability is output as the discrete action category at the current time.

7. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 5, characterized in that, The Q-learning-based action decision model includes: The state construction unit is used to construct the current environment state vector by combining the identity of the first person in the target service priority queue at the current moment, the discrete action categories of the first person at the most recent moments, the current depth distance value of the first person, and the body language response type executed by the system at the previous moment. A motion value assessment network is used to receive the current environmental state vector and calculate the Q value corresponding to each preset body language response type that the system can execute through a deep neural network. The decision output unit is used to select the preset body language response type with the largest Q value from all Q values ​​calculated by the action value evaluation network, as the optimal response action derived by the reinforcement learning decision module.

8. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 7, characterized in that, The system also includes: An experience replay buffer is used to store state transition samples generated by the reinforcement learning decision-making module during historical interactions. Each sample contains a historical environmental state vector, the type of body language response executed, the immediate reward obtained, a new environmental state vector, and a flag indicating whether the interaction has terminated. The target network parameter update module is used to periodically copy network parameters from the action value evaluation network to generate a target evaluation network with the same structure as the action value evaluation network but with frozen parameters. The target evaluation network is used to calculate the target Q value corresponding to the new environmental state vector in the state transition sample.

9. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 8, characterized in that, The system also includes: The model training module is used to randomly sample a batch of state transition samples from the experience replay buffer pool, and use the target evaluation network and the action value evaluation network provided by the target network parameter update module to calculate the time difference error between the predicted Q value and the target Q value. The network parameter optimization unit is used to update the parameters of the action value evaluation network using the backpropagation algorithm based on the time difference error calculated by the model training module, so as to minimize the time difference loss.

10. The deep learning-based automatic generation system for tourist interpersonal interaction body language according to claim 5, characterized in that, The body language generation module includes: The action type matching unit is used to receive the body language response type derived by the reinforcement learning decision module, and retrieve a preset three-dimensional motion template that completely matches the body language response type from the preset action template library. The preset three-dimensional motion template defines the motion trajectory of each joint of the virtual character or service robot within a fixed period. The control instruction synthesis unit is used to convert the joint motion trajectory data in the retrieved preset three-dimensional motion template into low-level control protocol instructions to drive the target virtual character or service robot. The instruction output unit is used to send the control protocol instructions generated by the control instruction synthesis unit to the virtual character animation engine or the service robot motion controller according to a preset execution sequence.

Images