Path planning obstacle avoidance method and system for mobile transparent display equipment

Abstract

The invention relates to the technical field of path planning, and discloses a path planning obstacle avoidance method and system for mobile transparent display equipment. The method comprises the following steps: acquiring environment perception data of the transparent display equipment; distance field and sight field fusion is carried out on the current position and the sight orientation of the visitor according to the environment perception data, and a visitor behavior intention label is obtained; performing trajectory prediction on the visitors based on the visitor behavior intention labels to obtain predicted trajectory data; and performing obstacle avoidance path planning on the transparent display device based on the predicted trajectory data to obtain a candidate moving path, and generating a motion control instruction based on the candidate moving path, thereby not only ensuring the optimality of a global path, but also ensuring the real-time performance of local obstacle avoidance, improving the overall navigation performance, and reducing the navigation cost. Therefore, the applicability and safety of the mobile transparent display device in the exhibition hall environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of path planning, in particular to a path planning obstacle avoidance method and system for a mobile transparent display device. BACKGROUND

[0002] The mobile transparent display device can provide dynamic information display services for visitors during movement. Existing mobile device obstacle avoidance systems generally use simple obstacle avoidance strategies based on distance and speed, lacking a deep understanding of visitor behavior intentions. In an exhibition hall environment, visitors have two typical behavior patterns: stopping to watch and passing through the channel. However, traditional methods cannot effectively distinguish between these two states, leading to inappropriate path planning decisions. When visitors are in the state of stopping to watch, the mobile transparent display device may perform unnecessary obstacle avoidance actions. When visitors are quickly passing through, the mobile transparent display device may not be able to respond in time, affecting the visitor experience and safety. SUMMARY

[0003] The present application provides a path planning obstacle avoidance method and system for a mobile transparent display device, which ensures the optimality of the global path and the real-time performance of local obstacle avoidance, improves the overall navigation performance, and further improves the applicability and safety of the mobile transparent display device in an exhibition hall environment.

[0004] In a first aspect, the present application provides a path planning obstacle avoidance method for a mobile transparent display device, which includes: acquiring environmental perception data of the transparent display device; fusing a distance field and a line-of-sight field based on the current position and line-of-sight orientation of the visitor according to the environmental perception data to obtain a visitor behavior intention label; performing trajectory prediction on the visitor based on the visitor behavior intention label to obtain predicted trajectory data; performing obstacle avoidance path planning on the transparent display device based on the predicted trajectory data to obtain a candidate movement path, and generating a motion control instruction based on the candidate movement path.

[0005] In combination with the first aspect, in a first implementation manner of the first aspect of the present application, the environmental perception data of the transparent display device is acquired, including: performing transparent material optical compensation on a first signal collected by a laser radar in the transparent display device to obtain a first corrected signal; performing depth deviation compensation on a second signal collected by a depth camera in the transparent display device according to the optical parameters of the transparent glass material to obtain a second corrected signal; The first and second correction signals are fused with inertial measurement unit data and encoder position data using multimodal sensor fusion to obtain environmental perception data.

[0006] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of fusing the distance field and line-of-sight field of the visitor's current position and gaze orientation based on the environmental perception data to obtain a visitor's behavioral intention label includes: Based on the environmental perception data, calculate the spatial distance between the visitor's current location and the transparent display device; Based on the spatial distance data, distance potential energy modeling is performed on the exhibition hall space to obtain distance field distribution data; Calculate the angle between the visitor's line of sight and the orientation angle of the transparent display device, and combine the distance field distribution data and the angle to perform line of sight analysis to obtain line of sight field distribution data; Visitor behavioral intent labels are generated based on the distance field distribution data and the line-of-sight field distribution data.

[0007] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of performing distance potential energy modeling on the exhibition hall space based on the spatial distance data to obtain distance field distribution data includes: The exhibition hall space is discretized into a grid based on the spatial distance data to obtain a grid node matrix. The distance matrix data between each grid coordinate point in the grid node matrix and the current position of the visitor is then calculated. The reciprocal potential energy transformation is performed on each distance value in the distance matrix data to obtain distance potential energy distribution data, and distance field distribution data is generated based on the distance potential energy distribution data of all visitors in the exhibition hall space.

[0008] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the step of predicting the visitor's trajectory based on the visitor's behavioral intent tag to obtain predicted trajectory data includes: Based on the visitor behavioral intent tags, the visitor's historical location data is classified to obtain stop-and-watch behavior data and passage-walking behavior data; Based on the standing and watching behavior data, slow movement feature encoding is performed to obtain the first feature code, and the straight-line crossing behavior data is performed to obtain the second feature code. Trajectory prediction is performed on the first feature code and the second feature code to obtain a predicted position coordinate sequence; The predicted position coordinate sequence is corrected by standing still and by passage speed correction to obtain the predicted trajectory data.

[0009] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, the step of planning an obstacle avoidance path for the transparent display device based on the predicted trajectory data to obtain a candidate movement path, and generating motion control commands based on the candidate movement path, includes: Extract the predicted location coordinates and movement speed of each visitor from the predicted trajectory data; Based on the movement speed, the risk impact radius of each visitor is dynamically adjusted to obtain risk radius data; Calculate the risk weight coefficient based on the predicted location coordinates and the movement speed; Construct a risk distribution map of the exhibition hall based on the risk radius data and the risk weight coefficient; Based on the exhibition hall risk distribution map, obstacle avoidance path planning is performed on the transparent display device to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

[0010] In conjunction with the first aspect, in a sixth implementation of the first aspect of the present invention, the step of calculating the risk weight coefficient based on the predicted position coordinates and the movement speed includes: Based on the predicted position coordinates and the movement speed, construct a distance feature vector, a motion feature vector, and an interaction feature vector; Cross-modal feature interaction calculations are performed on the distance feature vector, the motion feature vector, and the interaction feature vector to obtain the attention feature vector; Deep feature extraction is performed on the attention feature vector to obtain a deep feature encoding vector, and multi-scale weight calculation is performed on the deep feature encoding vector to obtain the risk weight coefficient.

[0011] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of planning obstacle avoidance paths for the transparent display device based on the exhibition hall risk distribution map to obtain candidate movement paths, and generating motion control commands based on the candidate movement paths, includes: Based on the exhibition hall risk distribution map, a Riemannian manifold model was performed on the transparent display device to obtain a global reference trajectory; Based on the global reference trajectory, the target attraction potential field and the obstacle repulsion potential field are superimposed to calculate the local optimized trajectory. Collision detection is performed on the local optimized trajectory to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

[0012] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the step of performing Riemannian manifold modeling on the transparent display device based on the exhibition hall risk distribution map to obtain a global reference trajectory includes: The initial configuration space is constructed by combining the three-dimensional position coordinates and three-dimensional attitude angles of the transparent display device, and the Riemannian manifold geometry model is performed on the initial configuration space based on the exhibition hall risk distribution map to obtain the Riemannian manifold configuration space. The weighted metric tensor of the Riemannian manifold configuration space is calculated based on the cost of position movement and attitude change of the transparent display device. Based on the weighted metric tensor, geodesic distance integration is performed on the transparent display device to obtain the optimal geodesic path, and trajectory point interpolation is performed on the optimal geodesic path to obtain the global reference trajectory.

[0013] Secondly, the present invention provides a path planning and obstacle avoidance system for a mobile transparent display device, the path planning and obstacle avoidance system for the mobile transparent display device comprising: The acquisition module is used to acquire environmental perception data of the transparent display device; The distance field and line-of-sight field fusion module is used to fuse the distance field and line-of-sight field based on the environmental perception data of the visitor’s current position and line of sight orientation to obtain the visitor’s behavioral intention label. The trajectory prediction module is used to predict the trajectory of visitors based on the visitor's behavioral intention tags, and obtain predicted trajectory data. The obstacle avoidance path planning module is used to plan obstacle avoidance paths for the transparent display device based on the predicted trajectory data, obtain candidate movement paths, and generate motion control commands based on the candidate movement paths.

[0014] The technical solution provided by this invention effectively eliminates the impact of the transparent display screen on the ranging accuracy of sensors by using transparent material optical compensation for LiDAR signals and depth camera signals, thereby improving the accuracy and reliability of environmental perception data. By employing a distance field and line-of-sight field fusion method, it can accurately distinguish between visitors' pausing and viewing behaviors and passageway walking behaviors, avoiding unnecessary obstacle avoidance actions. Based on visitor behavioral intention tags, it performs classification trajectory prediction, employing corresponding prediction algorithms for different behavioral patterns to improve the accuracy and reliability of future trajectory prediction. By dynamically adjusting the risk impact radius according to the visitor's movement speed and combining it with an intention weight network to calculate the risk weight coefficient, it achieves more accurate and intelligent risk field modeling. Through Riemannian manifold modeling and transparent display smoothness potential field constraints, it effectively controls the attitude changes and speed jumps during device movement, ensuring the visual continuity of screen content and viewing comfort. A two-layer path planning architecture is adopted, ensuring both the optimization of the global path and the real-time performance of local obstacle avoidance, improving overall navigation performance. By organically integrating transparent display stability constraints, safety obstacle avoidance constraints, and social norm constraints, it generates motion control commands that meet the special requirements of the exhibition hall environment, thereby improving the applicability and safety of mobile transparent display devices in the exhibition hall environment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the path planning and obstacle avoidance method for a mobile transparent display device provided in this application embodiment; Figure 2 A schematic block diagram of the path planning and obstacle avoidance system for a mobile transparent display device provided in an embodiment of this application. Detailed Implementation

[0017] 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, not all, of the embodiments of the present invention. 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.

[0018] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change based on the actual situation.

[0019] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0020] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0021] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating the path planning and obstacle avoidance method for a mobile transparent display device provided in an embodiment of this application. Figure 1As shown in the embodiments of this application, the path planning and obstacle avoidance method for a mobile transparent display device includes: Step S100: Obtain environmental perception data of the transparent display device; Specifically, the original first signal acquired by the LiDAR in the transparent display device undergoes optical compensation processing tailored to the transparent material. This compensation process is achieved through a nonlinear correction function constructed by the transparent material optical compensation module. This nonlinear correction function utilizes the nonlinear attenuation characteristics of the laser signal and the refraction interference caused by the transparent medium, combined with the nonlinear mapping capabilities of the tanh and sigmoid functions. Based on the optical parameters of the transparent glass material, such as refractive index, thickness, and transmittance, it corrects the distance error in the LiDAR signal caused by penetrating the transparent screen in real time, outputting a first correction signal. For the second signal acquired by the depth camera, a depth deviation compensation mechanism based on the transparent glass material parameters is introduced. This mechanism dynamically adjusts the depth value of the near-field region in the depth image captured by the camera, correcting the ranging error caused by glass transmission, reflection, and multipath interference, generating a second correction signal. The first and second correction signals are synchronously input into the fusion module, and together with the triaxial acceleration and angular velocity data provided by the inertial measurement unit and the real-time position information recorded by the encoder, they form a multi-source sensing input vector. The Kalman filter performs state estimation and error compensation on the data from these four types of heterogeneous sensors in the time domain, and performs attitude and position fusion in the space domain, outputting an environmental sensing data stream with high temporal resolution and high spatial consistency.

[0023] Step S200: Based on the environmental perception data, perform distance field and line-of-sight field fusion on the visitor's current position and line of sight orientation to obtain the visitor's behavioral intention label; Specifically, based on 3D spatial point cloud and image data from environmental perception data, the current location of visitors is spatially reconstructed. Combined with the posture information of the transparent display device and the relative motion state of the visitors, Euclidean spatial distance data between the visitor's current position and the device is calculated. Based on this spatial distance data, a visitor distance potential energy function is constructed across the entire exhibition hall space. This function transforms the spatial position of each visitor into an influencing potential energy point, ensuring that the distance field value of each point in space represents the cumulative effect of its inverse distance to all visitors. This generates distance field distribution data containing spatial distance gradient information, showing a significant steep potential energy region when the device approaches the crowd. An eye-tracking module acquires the visitor's current gaze orientation, and the angle between this gaze orientation and the orientation angle of the transparent display device is calculated. By weighting the cosine value of the angle between the visitor's gaze angle and the device's orientation, and combining this with the distance field value at the corresponding location, a gaze field distribution data reflecting the coupling relationship between gaze intensity and spatial distance is constructed. In this gaze field, the spatial region closer to the device and with a gaze angle closer to the front of the device has higher gaze potential energy. Joint analysis is performed based on distance field distribution data and line-of-sight field distribution data. A fusion function is used to weight and map the two data to form a gaze behavior feature vector for each visitor in a multi-dimensional feature space. A pre-trained deep classification network is then used to perform pattern recognition on the gaze behavior feature vector and output behavioral intent labels, including various types such as "stopping to watch", "walking through the passage", or "stopping briefly".

[0024] Step S300: Predict the trajectory of visitors based on visitor behavior intention tags to obtain predicted trajectory data; Specifically, based on visitor behavioral intent tags, semantic classification is performed on the historical location data of visitors within the current observation period, dividing all visitors into two categories: "stopping to watch" and "walking through the passageway." The "stopping to watch" behavior data is characterized by a short-term trajectory with small displacement amplitude, near-zero speed, and stable head posture. The "walking through the passageway" behavior data is characterized by a trajectory segment showing continuous movement along the exhibition hall path with a clear linear trend in position change. Differential feature encoding is then applied to these two types of behavioral trajectories. For "stopping to watch" behavior, features such as micro-displacement patterns, micro-variations in head orientation, and the distribution of dwell time during short-distance slow movement are extracted, and the slow movement features are extracted using a low-speed feature convolution module to obtain the first feature code. For "walking through the passageway" behavior, the linear directionality of the trajectory, velocity consistency, and acceleration stability are extracted, and the second feature code is generated using a directional convolution and velocity nested encoding module. The first and second feature codes are input into the trajectory prediction network. The network uses a Transformer structure to process historical trajectory information in a time-series manner and employs a multi-head attention mechanism to mine the correlation between behavioral intent and trajectory patterns, outputting a sequence of predicted position coordinates for several future time steps. For the behavior of pausing to watch, a static correction mechanism is introduced. By identifying segments in the predicted sequence whose position changes are less than a static threshold, these segments are forced to remain stationary, enhancing the semantic consistency of trajectory prediction. For the behavior of traversing a passage, a velocity correction is performed. This involves recalibrating the predicted position spacing based on a forward velocity estimation model to ensure the temporal consistency and physical feasibility of the trajectory, generating predicted trajectory data.

[0025] Step S400: Based on the predicted trajectory data, perform obstacle avoidance path planning for the transparent display device to obtain candidate movement paths, and generate motion control commands based on the candidate movement paths.

[0026] Specifically, the predicted position coordinates and velocity vectors of each visitor within a certain number of time steps in the future are extracted from the predicted visitor trajectory data, and these are used as the basic input for dynamic obstacle modeling. After obtaining the predicted velocity of each visitor, the risk impact radius is adaptively expanded according to their instantaneous velocity magnitude to construct a risk radius function. Visitors in a low-speed, stationary or standing-and-watching state are assigned a smaller local risk coverage area, while visitors in high-speed continuous movement are assigned a larger dynamic safety impact area, generating risk radius data for each visitor and fully considering the interference range of visitor behavior on path planning. Based on multi-dimensional factors such as the shortest distance between the predicted position and the predetermined path of the equipment, the magnitude of the predicted velocity, the degree of acceleration change, and the angle between the visitor's line of sight and the front of the equipment, the risk level of each visitor is assessed through an intent weight network, and a risk weight coefficient is output, so that visitors with a high tendency to cross and close to the center area of ​​the equipment path are given a higher avoidance priority. By spatially superimposing risk radius data and risk weight coefficients, a dynamic risk distribution map covering the exhibition hall space is constructed. Centered on the predicted trajectory point of each visitor, the risk distribution map forms a Gaussian decaying risk potential energy field within its risk radius based on the risk weight, achieving a visualized and quantitative description of potential collision areas. Based on the exhibition hall risk distribution map, the path planning module uses a multi-objective optimization strategy to perform obstacle avoidance path search. While ensuring the consistency of the device's global movement target, it minimizes path energy consumption and the integral value of the risk function, generating a candidate movement path that dynamically avoids crowd interference. Based on the candidate path sequence, combined with the device's current attitude, acceleration limits, and display stability constraints, motion control commands are generated to drive the underlying actuators of the transparent display device.

[0027] In one specific embodiment, the process of performing step S100 may specifically include the following steps: The first signal collected by the lidar in the transparent display device is optically compensated using transparent material to obtain the first correction signal; Based on the optical parameters of the transparent glass material, a second correction signal is obtained by performing depth deviation compensation on the second signal collected by the depth camera in the transparent display device. The first and second correction signals are fused with inertial measurement unit data and encoder position data using a multimodal sensor to obtain environmental perception data.

[0028] Specifically, the system models and corrects in real time the refraction, reflection, and multipath interference issues that occur when lidar signals pass through glass. After the lidar on the transparent display device receives the original echo signal, the built-in transparent material optical compensation module corrects the optical distortion of the first signal. The transparent material optical compensation module takes the refractive index, thickness, surface angle, and scattering coefficient of the transparent glass as input parameters, and uses a dual-channel nonlinear transformation function to jointly model the laser signal intensity, phase, time delay, and ranging error. The nonlinear part includes the hyperbolic tangent function tanh modeling the nonlinear attenuation of the laser signal, and the sigmoid function edge repairing the displacement error caused by the refraction path below the glass critical angle, generating a first correction signal after physically correcting the original distance information of the lidar. The second signal acquired by the depth camera contains depth map data calculated by infrared projection structured light or binocular parallax. However, under the interference of transparent materials, depth drift will occur due to light refraction or surface reflection. Based on the optical parameters of the glass, including its transmittance, surface roughness, reflectivity and optical density, the second signal is compensated for in a targeted manner. The compensation method adopts a pixel-level depth repair method based on a dynamic filtering model of transparent regions. Specifically, by calculating the depth consistency and texture change of each pixel between the current frame and the reference frame, and combining the refraction influence probability of the transparent material, local convolution reestimation and boundary enhancement interpolation are performed to obtain the second correction signal. The first and second correction signals are used as the main data sources for spatial perception. Together with the acceleration and angular velocity data collected by the inertial measurement unit of the device and the motion position change data obtained by the encoder, they form a fusion input set. A Kalman filter (or extended Kalman filter EKF) is constructed to perform temporal alignment and cooperative state estimation on various heterogeneous data sources. In each frame fusion calculation cycle, the filter maps the multi-source inputs to a unified state vector space. At the same time, the iterative mechanism of state prediction and observation update is used to continuously optimize the environmental perception state. The short-term high-frequency motion trend provided by the inertial measurement unit is used to compensate for the data sampling delay of the laser and camera when they rotate or vibrate rapidly. The position ranging provided by the encoder is used to correct the pose drift error. The first correction signal provides the high-precision contour boundary of the obstacle, and the second correction signal provides the dense area filling of the environmental depth. After the fusion of these four types of data, a structurally complete, spatiotemporally consistent and error-converged environmental perception dataset is formed. It includes the precise position, attitude and velocity vector of the transparent display device relative to the world coordinate system, as well as high-dimensional field information such as the geometric distribution, relative distance and motion trend of key entities such as visitors, static obstacles and transparent interfaces around the device.

[0029] In one specific embodiment, the process of performing step S200 may specifically include the following steps: Calculate the spatial distance between the visitor's current location and the transparent display device based on environmental perception data; Based on spatial distance data, distance potential energy modeling is performed on the exhibition hall space to obtain distance field distribution data; Calculate the angle between the visitor's line of sight and the orientation of the transparent display device, and combine the distance field distribution data and the angle to perform line of sight analysis, thereby obtaining line of sight field distribution data; Visitor behavioral intent labels are generated based on distance field distribution data and line-of-sight field distribution data.

[0030] Specifically, based on environmental perception data, in a unified coordinate system integrating LiDAR correction signals, depth camera-corrected depth maps, inertial measurement unit attitude data, and encoder position data, the three-dimensional position coordinates of all visitors are extracted. These coordinates are then used to calculate the Euclidean distance between each visitor and the transparent display device's current spatial position vector, yielding the spatial distance data between each visitor and the device. Based on each visitor's current spatial position, a potential field model of the exhibition hall space is created using a reverse distance potential energy function centered on that visitor. Specifically, in a two-dimensional or three-dimensional spatial grid, the weighted sum of the reciprocal distances between each spatial point and all visitors is calculated to generate distance field distribution data. This distance field exhibits a high gradient and high potential energy near crowds, while showing a gradually decreasing distribution in open areas, effectively describing the degree to which each point in the space is affected by crowds. Simultaneously, facial recognition and eye-tracking modules are used to obtain each visitor's current head posture and gaze orientation vector, which are then combined with the transparent display device's own posture model. The orientation vector output by the block is used to calculate the angle between the viewing direction and the device orientation through three-dimensional vector dot product operation. The closer the cosine value of the angle is to 1, the more the visitor's gaze direction tends to be towards the front of the device. Based on this, the system establishes a gaze attention function. By fusing the cosine value of the angle with the potential energy value of the visitor's position in the distance field, a set of gaze field distribution data with gaze weights is constructed. In the gaze potential energy distribution map, the area closer to the device and looking directly at it has high gaze potential energy, while the gaze potential energy of the distant or backward viewing area is close to zero. The gaze analysis model can effectively capture the density distribution of the visitor's gaze towards the transparent display device in space. After obtaining the distance field and gaze field, the behavioral intent discrimination module uses the two types of potential energy fields as input feature vectors. Pattern recognition and behavior classification are performed through a fusion network. The fusion network is based on a multilayer perceptron structure and combines residual connections and Dropout regularization mechanism to prevent overfitting. By extracting the spatial distance sequence, gaze angle sequence and path features in the potential energy field of each visitor within the time window, specific behavioral intent labels are output, including multiple semantic levels such as stopping to watch, walking through the passage, and short stay. The label generation process not only considers the single frame state, but also integrates temporal information and spatial semantics. The classification robustness of the labels is ensured by cross-validating the discrimination accuracy between the training set and the test set, forming a high-confidence visitor behavioral semantic label.

[0031] In one specific embodiment, the process of performing distance potential energy modeling on the exhibition hall space based on spatial distance data to obtain distance field distribution data can specifically include the following steps: The exhibition hall space is discretized into a grid based on spatial distance data to obtain a grid node matrix. The distance matrix data between each grid coordinate point in the grid node matrix and the current position of the visitor is then calculated. The reciprocal potential energy transformation is performed on each distance value in the distance matrix data to obtain the distance potential energy distribution data, and the distance field distribution data is generated based on the distance potential energy distribution data of all visitors in the exhibition space.

[0032] Specifically, a unified spatial coordinate system covering the entire exhibition hall area is established. Based on the boundary dimensions of the spatial coordinate system, a spatial resolution is set to regularly divide the three-dimensional or two-dimensional space. A gridding method is used to divide the exhibition hall into several equally spaced cubic or rectangular units. The center coordinates of each unit are defined as a grid node, thus constructing a grid node matrix for the exhibition hall space. Each element in the matrix corresponds to a discrete spatial coordinate point, forming a discrete environmental representation structure with spatial topological relationships. The current position coordinates of all visitors are extracted from the environmental perception module and mapped to the same spatial reference system. Using each grid node coordinate point as the reference origin, Euclidean distance calculations are performed between the reference origin and all visitors, forming a many-to-many distance calculation structure from grid nodes to visitors. That is, distance matrix data is established in the matrix dimension, where each element represents the geometric distance value from a certain grid node to a certain visitor. For each distance value in the distance matrix, a reciprocal transformation is performed to construct a distance potential energy function, expressed in the form of 1 / d, meaning the potential energy value is greater the closer the distance and decays rapidly with increasing distance. Exponential decay or higher-order reciprocal transformations are introduced as needed to enhance the gradient representation of densely populated areas, generating distance potential energy distribution data that represents the degree of spatial oppression or behavioral interference perceived by each grid point under the influence of a single visitor. In the case of multiple visitors, the distance potential energy distribution data corresponding to all visitors are superimposed and merged. An element-wise summation or maximum value superposition method is used to construct a global distance field distribution map. Based on the exhibition hall grid, this global distance field distribution map superimposes and expresses the potential energy influence of all visitors on every location point in the entire space, forming a spatial semantic map that can reflect the real-time distribution of visitor density, the range of behavioral interference, and the risk level of equipment avoidance. After the potential energy map is generated, a Gaussian filter is introduced to perform spatial smoothing on the field map to eliminate abrupt changes at the grid edges and improve the consistency of gradient transition. This allows the path selection direction and travel speed to be dynamically adjusted using the distance field as part of the cost function during subsequent path planning of the transparent display device, thereby enabling proactive avoidance of high potential energy regions and priority selection of paths for low potential energy regions.

[0033] In one specific embodiment, the process of performing step S300 may specifically include the following steps: Based on visitor behavioral intent tags, visitor historical location data is categorized to obtain data on stopping and watching behavior and data on passageway movement behavior. Based on the data of pausing and watching behavior, slow movement feature encoding is performed to obtain the first feature encoding, and straight-line crossing feature encoding is performed on the data of passage crossing behavior to obtain the second feature encoding; Trajectory prediction is performed on the first and second feature codes to obtain a predicted position coordinate sequence; The predicted position coordinate sequence is corrected by standing still and by traveling through the channel to obtain the predicted trajectory data.

[0034] Specifically, the semantic labels output by the visitor intent recognition module are used as the basic input, and high-frequency collected historical location data is used as the data source for dynamic behavior modeling. Based on the behavioral intent label of each visitor at the current moment, the trajectory data within the corresponding time window for each visitor is classified into two groups: lingering and viewing behavior and corridor walking behavior. Lingering and viewing is characterized by small positional changes per unit time, low speed, near-zero acceleration, stable orientation, and a direct alignment with the transparent display device. Corridor walking, on the other hand, is characterized by continuous linear positional migration, small but greater speed fluctuations than lingering behavior, stable path orientation, and a path orientation that is basically parallel to the exhibition hall corridor. During classification, four features are considered simultaneously: position increment, speed magnitude, angular velocity change, and the angle between the gaze direction and the device orientation. The lingering and viewing behavior data undergoes slow-movement feature encoding. By extracting information such as the visitor's micro-displacement amplitude changes, head micro-swing trajectory, and the drift trend of the center point of the lingering area per unit time, a slowly varying feature curve is constructed and encoded using a temporal convolutional coding network to generate the first feature code. The first feature code has temporal characteristics such as low frequency, low momentum, and strong positional stability. Straight-line crossing feature extraction is performed on the channel crossing behavior data. By calculating the root mean square error of the motion path deviating from the straight-line fitting direction formed by multiple historical position points, the rate of change of the angle between the instantaneous velocity and the direction vector, and the continuity of linear acceleration, a crossing stability index is established. The second feature code is obtained by encoding using a directional nested convolutional structure, reflecting the momentum intensity of continuous linear displacement, the consistency of crossing direction, and the rhythmic coherence. The first and second feature codes are respectively input into the trajectory prediction model. The model adopts a time-series-based Transformer structure or a bidirectional LSTM structure. At the encoding layer, a multi-head attention mechanism or a gated recurrent unit is used to capture the temporal dependency and state transition pattern between the trajectory history and future prediction. At the decoding layer, the predicted position coordinate sequence for the next T time moments is output. The first type of predicted trajectory is a relatively stationary or slightly disturbed trajectory, while the second type of predicted trajectory is a long linear trajectory that progresses along a linear direction and gradually stabilizes at a rate. A semantic-level correction mechanism is introduced into the predicted position coordinate sequence. For the behavior of standing and watching, if the displacement of multiple consecutive predicted points is lower than the preset static threshold, the trajectory is forcibly corrected to remain static, so that its coordinates remain constant to simulate the real standing state and prevent weak disturbances from being misinterpreted as intentional movement. For the behavior of walking through the passage, the spacing between predicted points is adjusted according to the actual average speed and direction components in the historical trajectory to eliminate speed jitter caused by distraction or uneven sampling intervals, thereby realizing speed correction and direction optimization of the movement rhythm. The predicted trajectory data output can reflect the future spatial dynamic trend of visitors.

[0035] In one specific embodiment, the process of performing step S400 may specifically include the following steps: Extract the predicted location coordinates and movement speed of each visitor from the predicted trajectory data; The risk radius data is obtained by dynamically adjusting the risk impact radius of each visitor based on their movement speed. Calculate the risk weighting coefficient based on the predicted location coordinates and movement speed; A risk distribution map of the exhibition hall is constructed based on risk radius data and risk weight coefficients; Based on the risk distribution map of the exhibition hall, obstacle avoidance path planning is performed on the transparent display device to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

[0036] Specifically, based on the multi-moment predicted trajectory data output by the trajectory prediction module, a point-by-point analysis is performed on each visitor's trajectory. The predicted position coordinates of each visitor within the next T time moments are extracted in time series form. The local velocity is calculated by the ratio of the spatial displacement difference between adjacent predicted points to the time step, resulting in a high-dimensional behavioral state sequence containing visitor ID, predicted time, predicted coordinates, and predicted velocity. The risk impact radius for each visitor is dynamically adjusted based on the extracted predicted velocity modulus. By constructing a linear or nonlinear velocity-radius mapping function, visitors in low-speed or stationary states are assigned a smaller local risk impact range, while visitors exhibiting high-speed passage or rapid turning behaviors are assigned a larger dynamic expansion radius. The risk radius not only changes dynamically with speed but is further modified by considering the visitor's current behavioral intention, head orientation, and relative positional potential with the equipment, thus forming risk radius data reflecting the aggressiveness of the current behavior and the possibility of spatial interference. A risk weighting model is constructed based on the predicted location coordinates and predicted speed information of visitors. By introducing characteristic parameters such as intent intensity, speed mutation rate, acceleration continuity, line-of-sight angle, and the probability of intersection with the transparent display device's path, these are used as inputs to a neural network or multivariate regression model. The output is a risk weighting coefficient describing the potential collision risk level of visitors. Individuals with high intersection with the device's movement path or sudden speed increases are given higher obstacle avoidance priority, while individuals stationary or far from the device's path are given lower weights. The risk impact range of all visitors is spatially superimposed. A gridded risk map is constructed in the global coordinate system of the exhibition hall, with each grid node serving as an evaluation unit. The Gaussian risk field function of all visitors is mapped to this grid area, and multiple risk sources within the same area are weighted and summed to generate an exhibition hall risk distribution map. The risk value of each spatial point in the map represents the probability intensity of interaction conflict or visual occlusion of the transparent display device at the current location. The exhibition hall risk distribution map is used as a spatial penalty term in the cost function during device path planning. During the path planning phase, a global path search algorithm with a risk penalty term (such as improved A*, RRT*, or geodesic path search based on Riemannian manifolds) is employed. A model predictive control (MPC) structure is embedded in the local path adjustment. By minimizing a weighted combination of path length, attitude change cost, display stability function, and spatial risk function over multiple time steps, a set of candidate movement paths is generated that avoids high-risk areas while satisfying spatial continuity and control stability. Based on the spatial location and orientation information of the candidate movement paths, combined with the device's kinematic model and control execution constraints, a specific sequence of motion control commands is generated, including multi-dimensional quantities such as linear velocity, angular velocity, acceleration limits, and target pose. This sequence is then sent to the underlying driver unit of the transparent display device via a control interface.

[0037] In one specific embodiment, the process of calculating the risk weight coefficient based on the predicted position coordinates and movement speed may specifically include the following steps: Construct distance feature vectors, motion feature vectors, and interaction feature vectors based on predicted location coordinates and motion speed; Cross-modal feature interaction calculations are performed on distance feature vectors, motion feature vectors, and interaction feature vectors to obtain attention feature vectors; Deep feature extraction is performed on the attention feature vector to obtain the deep feature encoding vector, and multi-scale weight calculation is performed on the deep feature encoding vector to obtain the risk weight coefficient.

[0038] Specifically, the predicted position coordinates and predicted movement speed at several future moments are extracted from the predicted trajectory data of each visitor. Based on this, three independent feature vectors are constructed. The distance feature vector describes the visitor's closest distance, minimum lateral offset, arrival time delay, and distance change slope relative to the center line of the current path of the transparent display device. The motion feature vector characterizes the velocity magnitude, velocity change rate, angular velocity change trend, positive and negative distribution of acceleration, and the curvature radius fluctuation of the trajectory. The interaction feature vector is constructed based on the visitor's head orientation, gaze direction, behavioral intention label, path overlap area length, and viewing angle, reflecting the degree of possible interaction interference with the device. The three types of features mentioned above are used as heterogeneous modal input vectors and input into the cross-modal feature fusion module. The cross-modal feature fusion module adopts a multimodal attention fusion structure to realize the saliency modeling relationship between distance modality, motion modality and interaction modality. By constructing three independent attention channels, the dynamic correlation weights within each modality are extracted using a self-attention mechanism. At the same time, a cross-attention mechanism is used to model the cross-response between different modalities. Then, a unified attention feature vector is generated through multi-channel weighted fusion. In the attention feature vector, high-risk actions (such as rapid traversal, sudden turning, and accelerated approach towards the device) are given higher attention activation responses, thus forming a high gradient representation in the feature space. The attention feature vector is input into a deep feature extraction network. The deep feature extraction network adopts a multi-layer residual convolutional structure or a stacked Transformer encoder architecture. The high-dimensional vector is transformed layer by layer through stacked nonlinear activation functions, batch normalization and residual connections to capture complex nonlinear behavioral features and potential risk causes. Dropout mechanism and L2 regularization constraint are introduced into the network to prevent overfitting and output a deep feature encoding vector with stable representation ability. Multi-scale weight calculation is performed on the deep feature encoding vector. Attention channels with multiple receptive fields at different scales are constructed to focus on local behavioral changes, short-term velocity fluctuations, and global trajectory trends. The response features output by each scale channel are weighted, normalized, and linearly weighted to generate risk weight coefficients, which represent the degree of impact of visitors on the safety of equipment paths and dynamic obstacle avoidance within the predicted time range.

[0039] In one specific embodiment, the process of performing obstacle avoidance path planning for the transparent display device based on the exhibition hall risk distribution map, obtaining candidate movement paths, and generating motion control commands based on the candidate movement paths can specifically include the following steps: Based on the risk distribution map of the exhibition hall, a Riemannian manifold model of the transparent display device is performed to obtain the global reference trajectory; The target attraction potential field and the obstacle repulsion potential field are superimposed and calculated based on the global reference trajectory to obtain the local optimized trajectory; Collision detection is performed on the locally optimized trajectory to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

[0040] Specifically, the exhibition hall risk distribution map serves as the environmental cost basis for path modeling. Each discrete spatial point in the exhibition hall risk distribution map carries risk density information related to the dynamic behavior of visitors. Therefore, before constructing the path geometry, the traditional Euclidean path space is transformed into a Riemannian manifold configuration space with a non-uniform cost structure. Using the six-dimensional attitude variables (including position and attitude) of the transparent display device as the domain, a Riemannian manifold containing position and angle dimensions is constructed. By introducing a metric tensor, the movement cost and risk density in each direction in space are coupled and expressed, thus lengthening the Riemann distance corresponding to higher-risk areas, naturally tending towards lower-risk areas in the path shortestification solution. The elements in the metric tensor are dynamically assigned by the risk value of the corresponding spatial position in the exhibition hall risk distribution map through a mapping function, enabling path search to not only consider spatial distance but also internalize spatial risk encoding. Then, within the Riemannian manifold space, a geodesic search algorithm is used to numerically solve for the shortest path connecting the starting point and the target point, generating a global reference trajectory that considers multi-dimensional risk costs. Based on the global reference trajectory and the actual operating status of the current equipment and local environmental disturbance information, a local optimization mechanism is initiated. This mechanism adjusts the global trajectory locally by constructing a joint potential energy function of the gravitational and repulsive potential fields. The target attraction potential field, originating from the target point on the current global trajectory, attracts the equipment towards the target. The obstacle repulsion potential field, constructed based on the high-risk point set in the risk distribution map, generates a strong reverse thrust near high-risk areas, forming an obstacle avoidance driving vector. The gradient superposition of the target attraction potential field and the obstacle repulsion potential field is calculated to form the resultant force vector in the total potential energy field. Then, based on the vector direction, local disturbance adjustments are made to the original trajectory points, generating a locally optimized trajectory that combines risk avoidance capabilities with target orientation. Multi-stage collision detection based on dynamic obstacles and static structures is performed on the locally optimized trajectory. A 3D spatial octree index structure is constructed to quickly determine whether trajectory points are located in static obstacle areas. Then, spatial overlap analysis synchronized with the predicted trajectory of visitors is used to determine whether potential trajectory conflict points exist within future time windows. Trajectory segments that do not meet constraints are pruned and reconstructed to select a set of candidate movement paths with feasibility and physical continuity. Based on each trajectory point in the candidate paths and the changes in velocity and orientation between them, combined with the kinematic model of the transparent display device and the dynamic constraint parameters of the actuators, specific motion control commands are generated. These commands include linear velocity, angular velocity, acceleration, and target attitude information, and are input to the device control layer after dynamic filtering.

[0041] The process involves superimposing the target attraction potential field and the obstacle repulsion potential field based on the global reference trajectory to obtain a locally optimized trajectory. This includes: constructing a target attraction potential field in quadratic form based on the target position coordinates in the global reference trajectory; obtaining the target attraction potential field function that drives the device to move towards the target by calculating the square of the Euclidean distance between the current position of the transparent display device and the target position and multiplying it by the attraction coefficient; constructing a single repulsion potential field in exponential decay form for each risk source based on the risk distribution position of each visitor in the exhibition hall risk distribution map; obtaining the obstacle repulsion potential field function that avoids collisions by calculating the square of the inverse of the distance between the transparent display device and each risk source minus the inverse of the safety distance threshold and multiplying it by the repulsion coefficient; and innovatively constructing a display smoothness potential field specifically for moving transparent display devices by calculating the square of the velocity term and acceleration of the transparent display device. The squared term and the integral term of attitude angle change are weighted and combined, and weighted coefficients for velocity stability, acceleration stability, and angle stability are set respectively to obtain a transparent display smoothness potential field function that suppresses display content jitter. The gradient of the target attraction potential field function, the negative gradient of the obstacle repulsion potential field function, and the weighted gradient of the transparent display smoothness potential field function are vector superimposed, and the superimposed vector field is parameterized by a sensory motion neural network to obtain a sensory motion vector field that integrates triple constraints. Based on the sensory motion vector field, the transparent display stability constraint manifold is applied to restrict the angular velocity magnitude and linear acceleration magnitude. By projecting the motion state that violates the constraints into the feasible region and keeping the display jitter evaluation function value below the set threshold, a locally optimized trajectory that simultaneously meets navigation requirements and display stability requirements is obtained.

[0042] In one specific embodiment, the process of performing Riemannian manifold modeling on the transparent display device based on the exhibition hall risk distribution map to obtain the global reference trajectory can specifically include the following steps: The initial configuration space is constructed by combining the three-dimensional position coordinates and three-dimensional attitude angles of the transparent display device, and the Riemannian manifold geometry model is performed on the initial configuration space based on the exhibition hall risk distribution map to obtain the Riemannian manifold configuration space. Calculate the weighted metric tensor of the Riemannian manifold configuration space based on the cost of position movement and attitude change of the transparent display device; Based on the weighted metric tensor, geodesic distance integration is performed on the transparent display device to obtain the optimal geodesic path. Then, trajectory point interpolation is performed on the optimal geodesic path to obtain the global reference trajectory.

[0043] Specifically, a high-dimensional configuration space is established for the motion state of the transparent display device. This high-dimensional configuration space simultaneously describes the device's three-dimensional position coordinates and three-dimensional attitude angles within the exhibition hall environment. The three-dimensional position determines the device's spatial occupancy in the translational dimension, while the three-dimensional attitude angles represent its rotational state around the pitch, roll, and yaw axes. This six-dimensional joint space constitutes the complete configuration state of the device at any given moment. Therefore, all possible configuration sets are defined as the initial configuration space. The initial configuration space is then transformed into a Riemannian manifold configuration space with spatial cost distribution information. A risk potential energy field constructed from the exhibition hall risk distribution map is introduced as the source of metric weights. The exhibition hall risk distribution map is a multi-channel environmental cost map composed of visitor predicted behavior, dynamic obstacle avoidance probability, environmental constraint boundaries, and visual occlusion risk. Each spatial point contains cost density information in different directions. This cost information is not only related to the translational path but also to the device's orientation, attitude, and the visibility of the expected display content at that location. Therefore, based on the initial configuration space, a six-dimensional weighted metric tensor is assigned to each configuration state to describe the trend of path cost changes at the current position and attitude. The position component of the six-dimensional weighted metric tensor is weighted by the spatial risk density of that point in the exhibition hall risk map, while the orientation component is weighted based on the impact of the device's content display angle on the visitor's field of vision. If the device's orientation is unfavorable for frontal content presentation, the movement cost in that orientation direction is increased, thus automatically avoiding inappropriate orientation configurations during path planning. Optimal path search is performed on the Riemannian manifold. The geodesic distance between the device's starting state and target state is solved by integrating the weighted metric tensor. The solution process is a variational problem of minimum path calculation. The solution method uses numerical approximation to represent the path as a discrete state sequence, and at each state point, the local increment of the path length is calculated based on the current tensor. Cost integration is accumulated over the entire path segment to obtain the shortest geodesic path in a high-dimensional non-Euclidean geometric space. The optimal geodesic path is processed by trajectory point interpolation. The discrete configuration sequence is smoothly connected by high-order spline interpolation, polynomial fitting or B-spline interpolation, so that the path satisfies the dynamic smoothness constraints such as continuous first derivative and bounded second derivative in six-dimensional space, and generates a global reference trajectory that meets the requirements of the kinematic model.

[0044] The process involves collision detection on locally optimized trajectories to obtain candidate movement paths, and generating motion control commands based on these paths. This includes: inputting the locally optimized trajectory into an octree spatial index for hierarchical fast collision detection pre-screening; recursively segmenting the 3D exhibition space into eight sub-cubes and storing visitor positions and obstacle information at each leaf node; performing hierarchical matching of trajectory points and octree nodes with logarithmic time complexity to obtain trajectory segments to be verified after excluding obvious safe areas; constructing a multi-objective constraint optimization problem based on the trajectory segments to be verified, including kinematic equality constraints, safety distance inequality constraints, and transparent display stability constraints; solving the constraint optimization problem using a precise trajectory verifier through sequential convex optimization; linearizing non-convex safety constraints near the current trajectory point to obtain feasible trajectory paths that pass safety verification; and inputting the feasible trajectory paths into a transparent display physical information neural network for physical constraint encoding processing, by constructing a system containing task tracking... A comprehensive loss function is derived, incorporating tracking loss, physical consistency loss, display stability loss, and social norms loss. The optical properties of the transparent material and display jitter constraints are encoded as physical constraints in the neural network, resulting in an optimization objective function that integrates the physical characteristics of the transparent display device. A scene complexity evaluation function is constructed based on obstacle density, visitor numbers, and average movement speed in the exhibition hall environment. An adaptive weight scheduler is used to exponentially decay the weight coefficients of each constraint in the comprehensive loss function, automatically reducing the weights of physical constraints in complex scenarios to enhance emergency response capabilities, thus obtaining an adaptive weighted loss function adapted to the current scene complexity. This adaptive weighted loss function is then input into a real-time optimization solver for rapid solution using an improved interior-point method. A pre-conditional conjugate gradient method is used to accelerate the calculation of linear equations and maintain the obstacle function for Newton direction updates. Path optimization calculations are completed within a 20-millisecond time limit, yielding motion control commands that meet the real-time requirements of the transparent display device.

[0045] Please see Figure 2 , Figure 2 This is a schematic block diagram of the path planning and obstacle avoidance system 200 for a mobile transparent display device provided in an embodiment of this application. Figure 2 As shown, the path planning and obstacle avoidance system 200 for a mobile transparent display device includes: The acquisition module 210 is used to acquire environmental perception data of the transparent display device; The distance field and line-of-sight field fusion module 220 is used to fuse the distance field and line-of-sight field based on the visitor's current position and line of sight orientation according to environmental perception data, so as to obtain the visitor's behavioral intention label. The trajectory prediction module 230 is used to predict the trajectory of visitors based on visitor behavioral intention tags to obtain predicted trajectory data. The obstacle avoidance path planning module 240 is used to plan obstacle avoidance paths for the transparent display device based on predicted trajectory data, obtain candidate movement paths, and generate motion control commands based on the candidate movement paths.

[0046] Through the collaborative efforts of the aforementioned components, and by applying transparent material optical compensation to LiDAR and depth camera signals, the impact of the transparent display on sensor ranging accuracy is effectively eliminated, improving the accuracy and reliability of environmental perception data. Employing a distance field and line-of-sight field fusion method, it can accurately distinguish between visitor pausing and viewing behaviors and passageway movement, avoiding unnecessary obstacle avoidance actions. Based on visitor behavioral intent labels, trajectory prediction is performed, employing corresponding prediction algorithms for different behavioral patterns to improve the accuracy and reliability of future trajectory prediction. By dynamically adjusting the risk impact radius according to visitor movement speed and combining it with an intent weight network to calculate risk weight coefficients, more accurate and intelligent risk field modeling is achieved. Through Riemannian manifold modeling and transparent display smoothness potential field constraints, attitude changes and speed jumps during device movement are effectively controlled, ensuring visual continuity and viewing comfort of the screen content. A two-layer path planning architecture is adopted, ensuring both global path optimization and real-time local obstacle avoidance, improving overall navigation performance. By organically integrating transparent display stability constraints, safety obstacle avoidance constraints, and social norm constraints, motion control commands that meet the special requirements of the exhibition hall environment are generated, thereby improving the applicability and safety of mobile transparent display devices in the exhibition hall environment.

[0047] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0048] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0049] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application.

Claims

1. A path planning and obstacle avoidance method for a mobile transparent display device, characterized in that, include: Acquire environmental perception data for transparent display devices; Based on the environmental perception data, distance field and line-of-sight field are fused to obtain visitor behavioral intention labels by analyzing the visitor's current position and line of sight orientation. Based on the visitor's behavioral intent tags, the visitor's trajectory is predicted to obtain predicted trajectory data; Based on the predicted trajectory data, obstacle avoidance path planning is performed on the transparent display device to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

2. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 1, characterized in that, The acquisition of environmental perception data of the transparent display device includes: The first signal collected by the lidar in the transparent display device is optically compensated using transparent material to obtain the first correction signal; Based on the optical parameters of the transparent glass material, a second correction signal is obtained by performing depth deviation compensation on the second signal collected by the depth camera in the transparent display device. The first and second correction signals are fused with inertial measurement unit data and encoder position data using multimodal sensor fusion to obtain environmental perception data.

3. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 1, characterized in that, The step of fusing the distance field and line-of-sight field based on the visitor's current position and gaze orientation using the environmental perception data to obtain a visitor's behavioral intent label includes: Based on the environmental perception data, calculate the spatial distance between the visitor's current location and the transparent display device; Based on the spatial distance data, distance potential energy modeling is performed on the exhibition hall space to obtain distance field distribution data; Calculate the angle between the visitor's line of sight and the orientation angle of the transparent display device, and combine the distance field distribution data and the angle to perform line of sight analysis to obtain line of sight field distribution data; Visitor behavioral intent labels are generated based on the distance field distribution data and the line-of-sight field distribution data.

4. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 3, characterized in that, The step of modeling the distance potential energy of the exhibition hall space based on the spatial distance data to obtain distance field distribution data includes: The exhibition hall space is discretized into a grid based on the spatial distance data to obtain a grid node matrix. The distance matrix data between each grid coordinate point in the grid node matrix and the current position of the visitor is then calculated. The reciprocal potential energy transformation is performed on each distance value in the distance matrix data to obtain distance potential energy distribution data, and distance field distribution data is generated based on the distance potential energy distribution data of all visitors in the exhibition hall space.

5. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 1, characterized in that, The step of predicting the visitor's trajectory based on the visitor's behavioral intent label to obtain predicted trajectory data includes: Based on the visitor behavioral intent tags, the visitor's historical location data is classified to obtain stop-and-watch behavior data and passage-walking behavior data; Based on the standing and watching behavior data, slow movement feature encoding is performed to obtain the first feature code, and the straight-line crossing behavior data is performed to obtain the second feature code. Trajectory prediction is performed on the first feature code and the second feature code to obtain a predicted position coordinate sequence; The predicted position coordinate sequence is corrected by standing still and by passage speed correction to obtain the predicted trajectory data.

6. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 1, characterized in that, The step of planning obstacle avoidance paths for the transparent display device based on the predicted trajectory data to obtain candidate movement paths, and generating motion control commands based on the candidate movement paths, includes: Extract the predicted location coordinates and movement speed of each visitor from the predicted trajectory data; Based on the movement speed, the risk impact radius of each visitor is dynamically adjusted to obtain risk radius data; Calculate the risk weight coefficient based on the predicted location coordinates and the movement speed; Construct a risk distribution map of the exhibition hall based on the risk radius data and the risk weight coefficient; Based on the exhibition hall risk distribution map, obstacle avoidance path planning is performed on the transparent display device to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

7. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 6, characterized in that, The step of calculating the risk weight coefficient based on the predicted position coordinates and the movement speed includes: Based on the predicted position coordinates and the movement speed, construct a distance feature vector, a motion feature vector, and an interaction feature vector; Cross-modal feature interaction calculations are performed on the distance feature vector, the motion feature vector, and the interaction feature vector to obtain the attention feature vector; Deep feature extraction is performed on the attention feature vector to obtain a deep feature encoding vector, and multi-scale weight calculation is performed on the deep feature encoding vector to obtain the risk weight coefficient.

8. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 7, characterized in that, The step of planning obstacle avoidance paths for the transparent display device based on the exhibition hall risk distribution map to obtain candidate movement paths, and generating motion control commands based on the candidate movement paths, includes: Based on the exhibition hall risk distribution map, a Riemannian manifold model was performed on the transparent display device to obtain a global reference trajectory; Based on the global reference trajectory, the target attraction potential field and the obstacle repulsion potential field are superimposed to calculate the local optimized trajectory. Collision detection is performed on the local optimized trajectory to obtain candidate movement paths, and motion control commands are generated based on the candidate movement paths.

9. The path planning and obstacle avoidance method for a mobile transparent display device according to claim 8, characterized in that, The process of performing Riemannian manifold modeling on the transparent display device based on the exhibition hall risk distribution map to obtain a global reference trajectory includes: The initial configuration space is constructed by combining the three-dimensional position coordinates and three-dimensional attitude angles of the transparent display device, and the Riemannian manifold geometry model is performed on the initial configuration space based on the exhibition hall risk distribution map to obtain the Riemannian manifold configuration space. The weighted metric tensor of the Riemannian manifold configuration space is calculated based on the cost of position movement and attitude change of the transparent display device. Based on the weighted metric tensor, geodesic distance integration is performed on the transparent display device to obtain the optimal geodesic path, and trajectory point interpolation is performed on the optimal geodesic path to obtain the global reference trajectory.

10. A path planning and obstacle avoidance system for a mobile transparent display device, characterized in that, A path planning and obstacle avoidance method for a mobile transparent display device as described in any one of claims 1-9, comprising: The acquisition module is used to acquire environmental perception data of the transparent display device; The distance field and line-of-sight field fusion module is used to fuse the distance field and line-of-sight field based on the environmental perception data of the visitor’s current position and line of sight orientation to obtain the visitor’s behavioral intention label. The trajectory prediction module is used to predict the trajectory of visitors based on the visitor's behavioral intention tags, and obtain predicted trajectory data. The obstacle avoidance path planning module is used to plan obstacle avoidance paths for the transparent display device based on the predicted trajectory data, obtain candidate movement paths, and generate motion control commands based on the candidate movement paths.

Images