Smart bracelet position tracking and navigation method, device and equipment and storage medium

Abstract

The invention relates to the technical field of smart bracelet positioning, and discloses a smart bracelet position tracking and navigation method, device and equipment and a storage medium, and the method comprises the steps: collecting multi-source positioning data through a built-in multi-sensor system of a smart bracelet, and carrying out the preprocessing, thereby obtaining the preprocessed multi-source data; performing feature extraction and space-time embedding processing to obtain an embedded multi-source feature sequence; inputting the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; constructing an adaptive pose map based on the initial position estimation result and the position uncertainty, and optimizing the adaptive pose map to obtain optimized position information; and a multi-level environment map is generated according to the optimized position information, and dynamic path planning is performed to obtain a navigation path and a navigation instruction, so that high-precision and low-power-consumption position tracking and navigation functions are realized, and smooth user experience is provided at the same time.

Description

Technical Field

[0001] The present application relates to the field of smart bracelet positioning technology, and in particular to a smart bracelet position tracking and navigation method, device, equipment and storage medium. Background Art

[0002] Traditional smart bracelets face huge challenges in accurate positioning and navigation. Existing positioning technologies such as GPS often perform poorly in complex environments such as indoors and urban canyons, while single-sensor positioning methods are difficult to cope with diverse application scenarios.

[0003] In addition, as a resource-constrained device, smart bracelets often face bottlenecks in computing power and battery life when processing complex positioning algorithms. Traditional positioning methods often require a lot of computing resources, which is contrary to the lightweight design concept of smart bracelets. At the same time, most existing navigation systems are developed for devices such as smartphones, and fail to fully consider the unique interaction methods and display limitations of smart bracelets. Summary of the invention

[0004] The present application provides a method, device, equipment and storage medium for smart bracelet location tracking and navigation, thereby realizing high-precision, low-power location tracking and navigation functions while providing a smooth user experience.

[0005] The first aspect of the present application provides a smart wristband position tracking and navigation method, the smart wristband position tracking and navigation method comprising: The multi-source positioning data collected by the multi-sensor system built into the smart bracelet is pre-processed to obtain pre-processed multi-source data; Performing feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence; Inputting the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; Constructing an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimizing the adaptive pose graph to obtain optimized position information; A multi-level environment map is generated according to the optimized position information, and dynamic path planning is performed to obtain a navigation path and navigation instructions.

[0006] A second aspect of the present application provides a smart wristband position tracking and navigation device, the smart wristband position tracking and navigation device comprising: The acquisition module is used to collect multi-source positioning data through the multi-sensor system built into the smart bracelet, and pre-process it to obtain pre-processed multi-source data; An extraction module is used to perform feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence; A processing module, used for inputting the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; A construction module, used to construct an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimize the adaptive pose graph to obtain optimized position information; The planning module is used to generate a multi-level environment map according to the optimized location information, and perform dynamic path planning to obtain a navigation path and navigation instructions.

[0007] The third aspect of the present application provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the electronic device executes the above-mentioned smart bracelet position tracking and navigation method.

[0008] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned smart bracelet position tracking and navigation method.

[0009] Compared with the prior art, the present application has the following beneficial effects: by integrating multiple sensor data such as GPS, IMU, Wi-Fi and BLE, the positioning accuracy and robustness are improved, and reliable location information can be continuously provided in different indoor and outdoor environments. The sensor sampling frequency is dynamically adjusted according to the user's motion state, which effectively reduces energy consumption while ensuring positioning accuracy and prolongs the use time of the smart bracelet. Targeted feature extraction methods, such as principal component analysis of Wi-Fi data and compressed sensing algorithm for BLE data, are adopted to greatly reduce the data dimension and reduce the computational complexity of subsequent processing. A multi-source information fusion model including feature processing layer, self-attention mechanism and cross-attention mechanism is designed, which can effectively capture the spatiotemporal correlation between different data sources and improve positioning accuracy. Through nonlinear factor recovery and dynamic covariance adjustment, an adaptive pose graph is constructed, which improves optimization efficiency and positioning accuracy, while reducing storage requirements. Combined with the global topology map and the local occupancy grid map, the macro and micro environmental representation is realized, providing comprehensive environmental information for subsequent path planning. The combination of A* algorithm and dynamic window method not only ensures global optimality, but also adapts to the local dynamic environment, and improves the intelligence and flexibility of navigation. By generating hierarchical navigation instructions and optimizing instruction combinations, the interactive characteristics of smart bracelets are fully considered to provide an intuitive and easy-to-understand navigation experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0011] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0012] Figure 1 It is a flowchart of a smart bracelet position tracking and navigation method provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of a smart wristband position tracking and navigation device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0015] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0016] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1, an embodiment of the smart bracelet position tracking and navigation method in the embodiment of the present application includes: Step 100: collect multi-source positioning data through the multi-sensor system built into the smart bracelet, and pre-process it to obtain pre-processed multi-source data; It is understandable that the execution subject of the present application may be a smart bracelet position tracking and navigation device, or a terminal or a server, which is not limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0017] Specifically, the smart bracelet collects raw GPS data, IMU data, Wi-Fi data and BLE data through multiple sensors such as built-in GPS receiver, inertial measurement unit (IMU), wireless communication module and Bluetooth low energy (BLE) module. During the GPS data collection process, the smart bracelet records the strength of the satellite signal, the number of visible satellites and pseudo-range information to obtain fine raw GPS data, which reflects the geographic information of the current location of the smart bracelet and the quality of the satellite signal, and helps to judge the reliability and accuracy of the positioning information. At the same time, the inertial measurement unit collects data from the three-axis accelerometer and the three-axis gyroscope to form raw IMU data, reflecting the dynamic information such as acceleration and angular velocity of the smart bracelet, thereby assisting positioning and attitude estimation. The smart bracelet scans the service set identifier (SSID) and received signal strength indication (RSSI) of the surrounding Wi-Fi access points through the wireless communication module to generate raw Wi-Fi data. This data is used to identify the environmental characteristics of the smart bracelet, such as indoor location or signal distribution in a specific area. Similarly, the Bluetooth low energy module continuously detects the presence of surrounding BLE beacons and records their received signal strength indications to generate raw BLE data. During the data collection process, the sampling frequency of each sensor is dynamically adjusted based on the user's motion state to improve the efficiency and accuracy of data collection. When the user is detected to be in a stationary state, the sampling frequency of GPS and IMU is reduced. At the same time, the time interval of Wi-Fi and BLE scanning is increased, because in a stationary state, the wireless signal characteristics of the environment are usually relatively stable, and less frequent scanning can provide sufficient information. When the user is in motion, the smart bracelet will increase the sampling frequency of GPS and IMU to capture the user's dynamic location information and movement trajectory more timely; at the same time, the scanning interval of Wi-Fi and BLE is reduced, and the frequency of data collection is increased, so as to better capture the changes in environmental signals during the user's movement and ensure the continuity and accuracy of positioning. By dynamically adjusting the sampling frequency of the sensor, the acquisition of adaptive sampling data is achieved. Outliers are detected and removed for raw GPS data, raw IMU data, raw Wi-Fi data, and raw BLE data. For example, for GPS data, satellite number threshold method and pseudo-range residual analysis are used to detect and remove data points with large errors; for IMU data, kinematic models and statistical analysis methods are used to identify and remove abnormal acceleration and angular velocity data that do not conform to the laws of motion; and for Wi-Fi and BLE data, by comparing the historical signal strength distribution at the same location, outliers that deviate from the normal range are excluded. After processing, the multi-source data after preliminary cleaning is obtained. The multi-source data after preliminary cleaning is processed by filtering and smoothing algorithm to remove noise and data fluctuations, making the data smoother and more continuous.In the filtering process, Kalman filtering, particle filtering and other methods are used to perform filtering operations to different degrees according to the characteristics of various types of data to obtain more accurate filtered multi-source data. After the filtering process is completed, the data of different sensors are unified to the same time base, and time alignment algorithms such as interpolation and extrapolation are used to unify the data of all sensors to a common time base to form pre-processed multi-source data.

[0018] Step 200, performing feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence; Specifically, for the preprocessed GPS data, the spatial relationship between the satellite and the smart bracelet is measured by extracting the geometric distribution characteristics of the satellite, and the horizontal precision factor and vertical precision factor are calculated based on this. These factors represent the uncertainty and accuracy of GPS positioning in the horizontal and vertical directions, respectively, and can effectively characterize the reliability and accuracy of the current GPS positioning. After combining the geometric distribution characteristics with the precision factor, a GPS feature vector containing rich positioning information is obtained. At the same time, for the preprocessed IMU data, the time domain signal is converted to the frequency domain space through Fourier transform to extract the frequency domain features related to the motion state. The accelerometer and gyroscope data in the IMU data contain a lot of time-varying information. Through Fourier transform, the amplitude and phase of different frequency components are decomposed to reveal the periodicity and change characteristics in the user's motion pattern. The extracted frequency domain features can help better understand the user's dynamic behavior state and generate an IMU feature vector containing the main frequency information and motion pattern. At the same time, the RSSI fingerprint vector is constructed for the preprocessed Wi-Fi data. RSSI fingerprint vector refers to the signal feature representation obtained by sampling the received signal strength (RSSI) of multiple Wi-Fi access points around, reflecting the strength distribution of Wi-Fi signals in the current environment. The RSSI fingerprint vector is reduced in dimension using principal component analysis. The feature dimension is reduced while retaining the main signal features to obtain the Wi-Fi feature vector. For the preprocessed BLE data, the compressed sensing algorithm is used for processing. In the sparsely represented RSSI fingerprint space, the compressed sensing algorithm can effectively extract representative significant features and avoid the information loss problem caused by signal sparsity. BLE beacons are used for short-distance positioning and environmental perception. Their signal strength is greatly affected by the complexity of the environment. The BLE features are extracted through sparse representation. While retaining important signal information, irrelevant noise interference is filtered out to obtain a BLE feature vector containing significant location features. The GPS feature vector, IMU feature vector, Wi-Fi feature vector and BLE feature vector are spliced ​​to form a multi-source feature vector, which reflects the motion state and environmental perception of the current smart bracelet. The time embedding algorithm is applied to the multi-source feature vector. The absolute time information is converted into a periodic representation using the sinusoidal position coding method. Sinusoidal position coding maps the time information into a high-dimensional space, allowing the model to capture the periodic laws of time changes and obtain a time embedding vector containing the time change pattern. While performing time embedding, a spatial embedding algorithm is applied to the multi-source feature vector. The grid coding method is used to discretize the prior position information into a grid index with a fixed resolution, so that the originally continuous spatial information is discretized into grid units, forming a spatial embedding vector containing the position discretization information.The multi-source feature vector is concatenated with the time embedding vector and the space embedding vector to generate an embedding vector representation that comprehensively considers the time, space and multi-source sensor information. The embedding vector sequence is divided into several time windows of fixed length, and each time window contains the information changes of all sensors within a period of time. Through the construction of spatiotemporal embedding and feature sequence, the embedded multi-source feature sequence is obtained.

[0019] Step 300: input the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; It should be noted that the embedded multi-source feature sequences are respectively input into the independent neural network submodules of the feature processing layer. For the GPS and IMU feature sequences, since they have strong correlation in the time dimension, the long short-term memory network (LSTM) is used for processing. LSTM can capture the long-term and short-term dependencies in time series data, and is suitable for processing location information and motion state data with time continuity. Through the memory unit and forgetting mechanism of LSTM, the time series pattern in GPS and IMU features is effectively extracted. At the same time, for Wi-Fi and BLE feature sequences, since they have a certain topological structure in the spatial dimension and the signal strength distribution has a certain spatial correlation, the graph convolutional neural network (GCN) is used for processing. GCN can capture the relationship and feature transfer between nodes in the graph structure data, and is suitable for processing the RSSI fingerprint data of Wi-Fi and BLE signals. Through the convolution operation of GCN, the spatial relationship of signal strength is established between the Wi-Fi access point and the BLE beacon node, and the spatial feature representation of each node is extracted. After processing by LSTM and GCN, the preliminarily processed GPS, IMU, Wi-Fi and BLE feature representations are obtained. The self-attention mechanism is applied to the feature representation after preliminary processing. The self-attention mechanism identifies the intrinsic correlation between features by calculating the similarity matrix between different features in the same data source. By weighted summing the feature representation, the self-attention mechanism can highlight important feature information, suppress irrelevant or redundant features, and obtain enhanced single-source feature representation. In order to achieve interaction and fusion between multi-source data, the cross-attention mechanism is applied to the enhanced single-source feature representation. The cross-attention mechanism captures the potential connection and complementary information between different modal data by calculating the correlation matrix between feature representations of different data sources. It can interact information between features of different data sources and achieve cross-modal feature fusion. In this way, GPS, IMU, Wi-Fi and BLE features can share each other's context information to obtain the fused feature representation. The fused feature representation is input into the fully connected layer and mapped into a unified latent space to obtain the latent space feature vector. The fully connected layer includes an input layer, a hidden layer and an output layer, where the hidden layer uses the ReLU activation function. The ReLU function can effectively alleviate the gradient vanishing problem and performs well in nonlinear mapping. Through layer-by-layer mapping and nonlinear transformation of a multi-layer fully connected network, the latent space feature vector can abstract and condense multi-source feature information. The latent space feature vector is input into the regression head network to output the initial position coordinates and obtain the initial position estimation result. The regression head network can map the latent space features to the actual position information, train the model by minimizing the positioning error, and make the initial position estimation result more accurate.At the same time, in order to evaluate the reliability of the initial position estimation results, the classification head network is applied to the latent space feature vector to output a discretized probability distribution to obtain the position uncertainty. The classification head network can give the confidence and possibility distribution of each position prediction based on the diversity of features and the uncertainty of model prediction.

[0020] Step 400: construct an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimize the adaptive pose graph to obtain optimized position information; Specifically, a factor graph model is established based on the initial position estimation results. The model obtains an initial pose graph by representing each position estimation in a continuous time period as a node in the graph. In the factor graph model, the position node is the basic element of the graph, which is used to represent the position information of the smart bracelet at different times, while the observation factor and the motion model factor are the edges connecting these nodes, respectively representing the observation information of each sensor and the constraint relationship of the kinematic model. When establishing the initial pose graph, an absolute position factor is established for the GPS observation data in the graph, and the GPS observation information is linked to the position information node at the corresponding time to represent the position of the smart bracelet in the global coordinate system. For IMU data, since it mainly reflects the relative motion relationship between the bracelet at different time steps, a relative pose factor is constructed to describe the motion state between adjacent time nodes in the form of acceleration and angular velocity, providing a smooth transition between consecutive time steps. Pseudo-range factors are established for Wi-Fi and BLE observations. These factors simulate the distance relationship between the smart bracelet and the surrounding wireless signal source through the received signal strength indication (RSSI) information, adding additional constraints to the pose graph. Combining all the above factors, a multi-source observation factor graph is formed, which includes GPS factor, IMU factor, Wi-Fi factor and BLE factor. All kinds of factors work together to describe the position and motion characteristics of the bracelet in space. In order to make the multi-source observation factor graph more concise and improve the computational efficiency, a nonlinear factor recovery algorithm is applied to it. The algorithm extracts key position information from the dense prior factors generated by the marginalization process, merges or deletes the redundant observation factors in the graph, and obtains a sparse pose graph. In this process, the sparse processing significantly reduces the number of nodes and edges in the graph, thereby reducing the computational complexity and storage requirements. At the same time, the covariance matrix of each factor in the sparse pose graph is dynamically adjusted according to the position uncertainty. The adjustment is based on the confidence of the observation data of each sensor. The factor with a larger covariance matrix represents a higher uncertainty of the observation, and vice versa, it means that the observation is more reliable. Through the dynamic adjustment of the covariance matrix, the adaptive pose graph can better adapt to the changes in sensor data in different environments and motion states, and obtain a more reasonable and accurate adaptive pose graph. In order to optimize the adaptive pose graph, the incremental smoothing and mapping (iSAM) algorithm is applied to construct a sparse information matrix. While constructing the sparse information matrix, a Levenberg-Marquardt optimizer is constructed to perform the optimization task. The Levenberg-Marquardt algorithm is a nonlinear least squares optimization method that can provide stable optimization results when the uncertainty of multi-source data is large. By setting an appropriate damping factor and number of iterations, the step size and convergence speed of the optimization process are controlled to avoid divergence problems caused by initial value deviation or data inconsistency.This iterative optimization strategy combines efficient matrix decomposition with a gradual update mechanism, so that the amount of computation in each iteration can be controlled and can quickly converge to the local optimal solution. After completing the iterative optimization of the adaptive pose graph, the optimization strategy is iterated for multiple rounds until the positions and edge covariances of all nodes converge to a stable state, and the optimized node positions and edge covariance matrices are obtained. Based on the optimized node positions and edge covariances, the latest position information and the complete covariance matrix are extracted to obtain the final optimized position information. The optimized position information reflects the precise position of the smart bracelet in the environment, and also includes a global position uncertainty measure based on multi-source data fusion.

[0021] Step 500: Generate a multi-level environment map based on the optimized location information, and perform dynamic path planning to obtain a navigation path and navigation instructions.

[0022] Specifically, a global topology map is constructed based on the optimized location information, and the connection relationship between different areas is represented by a graph structure. The global topology map consists of nodes, edges and weights, where nodes represent key locations or landmarks, such as rooms, floor entrances or important intersections; edges represent the connection relationship between these nodes, and weights reflect the distance between nodes or the difficulty of passage. The macro-level environmental representation can effectively describe the overall layout structure of the entire navigation area. At the same time, the local environmental features are extracted from the optimized location information to construct an occupancy grid map. The occupancy grid map divides the local environment into multiple grid units, each unit corresponds to a spatial area, and the occupancy probability of the grid unit indicates whether the area is occupied by obstacles or is passable. Through real-time updating rules, the occupancy grid map dynamically reflects changes in the environment, such as newly added obstacles, dynamic crowds, etc. The map describes the geometric characteristics and traffic conditions of the local environment, forming a micro-level environmental representation. The global topology map of the macro layer and the occupancy grid map of the micro layer are merged to construct a multi-level environmental map. The multi-level environmental map consists of a topological layer, a metric layer, and a semantic layer. The topological layer mainly represents the connection relationship between different areas, the metric layer accurately describes the geometric characteristics of the local environment through the occupancy grid map, and the semantic layer contains a high-level understanding of the environment, such as room type, regional function, etc. Through the multi-level environmental representation, the smart bracelet can understand and perceive the user's current location and its surrounding environment at different levels. Based on the multi-level environmental map, the A algorithm is applied to the global topological map for rough path search to obtain the global navigation direction. The A algorithm provides users with a rough direction of travel and path planning by searching the shortest path from the starting position to the target position in the topological map. The dynamic window method is applied to the local occupancy grid map. The dynamic window method can perform local fine path planning based on the user's motion characteristics and the distribution of environmental obstacles. This method selects the optimal direction of travel by evaluating the possible paths under different speed and direction combinations, thereby avoiding obstacles and finding the safest and optimal local path. In order to make the navigation path smoother, the cubic spline interpolation algorithm is used to smooth the local path. Cubic spline interpolation can generate smooth continuous curves between original path points, eliminating unnecessary sharp turns and mutations, making the path more in line with the user's natural walking habits. Smooth navigation paths can improve user followability and comfort. Based on the smooth navigation path, hierarchical navigation instructions are generated, covering direction instructions at the macro level and detailed guidance at the micro level. At the macro level, navigation instructions mainly provide guidance in the general direction, while at the micro level, the instructions are more specific and detailed. The generated multi-level navigation instructions include text instructions, and are also presented in a multimodal manner through voice prompts and vibration feedback. Text instructions are displayed directly on the bracelet's screen, voice prompts are broadcast through the bracelet's speakers, and vibration feedback provides users with silent tactile prompts, which are suitable for environments where it is noisy or inconvenient to view the bracelet screen.In order to improve the navigation experience, the multi-level navigation instructions are optimized and combined, and the instruction density is dynamically adjusted according to the user's preferences and the complexity of the environment. For example, in an environment familiar to the user, the navigation instructions are simpler, and in a complex environment they are more detailed, so that the user can better understand and follow the navigation instructions. The smart bracelet can generate accurate navigation paths and guidance information based on multi-level environmental maps, optimized path planning and personalized navigation instructions.

[0023] Based on Bluetooth mesh technology, a communication network between multiple smart bracelets is established to achieve interconnection between smart bracelets. In order to achieve effective location information exchange, motion state sharing and navigation intention broadcasting in this network, a set of special communication protocols is defined. The protocol clarifies the structure, transmission method and broadcasting rules of the data packet, and constructs a distributed multi-smart bracelet communication architecture. In this architecture, each smart bracelet acts as both a data source and a relay node. It broadcasts its own location, motion state and navigation intention to other bracelets through the Bluetooth mesh network, and receives the status information of other bracelets, forming a group system of multiple smart bracelets working together. The location information, motion state and navigation intention of each bracelet are encoded. A state encoder is constructed to convert high-dimensional raw data into low-dimensional feature vectors, which contain the current important state information of each bracelet, such as location coordinates, motion direction, speed and user's navigation target. Based on the topological structure in the communication architecture, a graph attention network is constructed. The network can perform weighted aggregation of state information between different smart bracelets, and automatically assign the importance weight of each smart bracelet state through the attention mechanism, so as to extract key group state features in the network. The obtained group state features can comprehensively reflect the collective motion trends of each smart bracelet in the entire bracelet network and the relative position relationship between individuals. The group state features are input into the state decoder, decoded and reconstructed, and the optimized position estimation and consistent motion state are calculated for each smart bracelet to obtain the collaborative positioning result. This collaborative positioning result effectively reduces the positioning error of a single bracelet, and achieves more accurate collective positioning through information sharing between smart bracelets. Group consistency constraints are applied to the collaborative positioning results, that is, by minimizing the relative position and speed differences between adjacent smart bracelets, all bracelets in the group show consistent motion states, ensuring that the entire bracelet group has good coordination and consistency during the navigation process, and obtaining a globally consistent group state. Based on the globally consistent group state, a multi-agent reinforcement learning environment is constructed, and each smart bracelet is regarded as an independent agent. The state space of each agent is defined as its position information, speed, and relative position of neighboring bracelets, and the action space includes the direction of travel, speed adjustment, etc. The reward function comprehensively considers the navigation accuracy, the degree of proximity to the target, and the group consistency, and obtains a complete group navigation learning framework. In this group navigation learning framework, a distributed proximal policy optimization algorithm is applied to train the navigation policy network of each smart bracelet. Through the distributed algorithm, each smart bracelet can learn strategies to work with other bracelets while maintaining individual navigation independence, thus obtaining a more robust individual navigation strategy. In order to optimize these strategies while protecting user privacy, based on the principle of federated learning, the navigation experience of multiple smart bracelets is aggregated, the global navigation policy model is updated, and the initial group navigation strategy is obtained.The federated learning method only exchanges model parameters between wristbands without transferring user data, so it can effectively protect user privacy. Based on the initial group navigation strategy, the user's historical trajectory and current navigation intention are used to identify potential group gathering points, such as common destinations, possible gathering places, etc., and a group gathering point recommendation list is generated to realize the social navigation function. This function can help users more conveniently meet with other members of the group and enhance the social experience. In order to meet the personalized needs of different users, the social navigation function is personalized. According to the user's preferences, dynamic changes in the environment, and real-time navigation needs, the smart bracelet can adaptively update the navigation strategy, thereby providing users with the best navigation suggestions in different scenarios, and finally obtaining the navigation strategy for the target group.

[0024] In the embodiment of the present application, by integrating multiple sensor data such as GPS, IMU, Wi-Fi and BLE, the positioning accuracy and robustness are improved, and reliable location information can be continuously provided in different indoor and outdoor environments. The sensor sampling frequency is dynamically adjusted according to the user's motion state, which effectively reduces energy consumption while ensuring positioning accuracy and prolongs the use time of the smart bracelet. Targeted feature extraction methods, such as principal component analysis of Wi-Fi data and compressed sensing algorithm for BLE data, greatly reduce the data dimension and reduce the computational complexity of subsequent processing. A multi-source information fusion model including feature processing layer, self-attention mechanism and cross-attention mechanism is designed, which can effectively capture the spatiotemporal correlation between different data sources and improve positioning accuracy. Through nonlinear factor recovery and dynamic covariance adjustment, an adaptive pose graph is constructed, which improves optimization efficiency and positioning accuracy, while reducing storage requirements. Combined with the global topology map and the local occupancy grid map, the macro and micro levels of environmental representation are realized, providing comprehensive environmental information for subsequent path planning. The combination of A* algorithm and dynamic window method not only ensures global optimality, but also adapts to the local dynamic environment, and improves the intelligence and flexibility of navigation. By generating hierarchical navigation instructions and optimizing instruction combinations, the interactive characteristics of smart bracelets are fully considered to provide an intuitive and easy-to-understand navigation experience.

[0025] In a specific embodiment, the process of executing step 100 may specifically include the following steps: The satellite signal strength, number of visible satellites and pseudo-range information collected by the GPS receiver are recorded to obtain the original GPS data; the accelerometer and gyroscope three-axis data collected by the inertial measurement unit are collected to obtain the original IMU data; the wireless communication module scans the service set identifier and received signal strength indication of the surrounding Wi-Fi access points to obtain the original Wi-Fi data; the Bluetooth low energy module detects the surrounding BLE beacons and records their received signal strength indication to obtain the original BLE data; The sampling frequency of each sensor is dynamically adjusted based on the user's motion state. When the user is detected to be stationary, the sampling frequency of GPS and IMU is reduced, and the Wi-Fi and BLE scanning intervals are increased. When the user is in motion, the sampling frequency of GPS and IMU is increased, and the Wi-Fi and BLE scanning intervals are reduced to obtain adaptive sampling data. Perform outlier removal processing on the original GPS data, original IMU data, original Wi-Fi data, and original BLE data to obtain multi-source data after preliminary cleaning; A filtering and smoothing algorithm is performed on the multi-source data after preliminary cleaning to obtain filtered multi-source data, and the filtered multi-source data are unified to the same time base to obtain preprocessed multi-source data.

[0026] Specifically, raw data is obtained from multiple different sensors, and these data are preprocessed and optimized. For GPS receivers, the strength of satellite signals, the number of visible satellites, and pseudorange information are recorded. Satellite signal strength refers to the signal strength value received from GPS satellites, which is used to measure the quality of the received signal. Signal strength is usually expressed in dB-Hz (decibel Hertz). The higher the signal strength, the better the signal quality, which improves positioning accuracy. The number of visible satellites indicates the total number of satellites whose signals the receiver can receive at the current moment. More visible satellites usually mean improved positioning accuracy, and more satellites can provide more spatial constraints. Pseudorange information is the distance obtained by multiplying the signal propagation time measured by the receiver by the speed of light. It is the basic data for measuring the distance between the receiver and the satellite. Suppose at a certain moment, the recorded pseudorange of a certain satellite is , the satellite signal strength is , the number of visible satellites is , the calculation formula of pseudorange is: ; in, Indicates the receiving time of the receiver, Indicates the satellite transmission time, is the receiver clock deviation, is the satellite clock bias, The pseudo-range data helps calculate the position of the receiver relative to each visible satellite, thereby obtaining the positioning result. For the inertial measurement unit (IMU), the three-axis data of the accelerometer and gyroscope are collected. The accelerometer measures the acceleration changes on the X, Y, and Z axes, which are expressed as and These data reflect the linear acceleration of the smart bracelet in all directions. Similarly, the gyroscope measures the angular velocity around three axes, which are expressed as ,and , represents the change of the rotation speed of the bracelet in each direction. Assume that the acceleration and angular velocity of the bracelet at a certain moment are: ; These data are integrated to obtain speed and attitude information. For the wireless communication module, data is obtained by scanning the service set identifier (SSID) and received signal strength indicator (RSSI) of the surrounding Wi-Fi access points. SSID is the name used to uniquely identify a Wi-Fi network, and RSSI indicates the strength of the received signal. Assume that the RSSI of a Wi-Fi signal detected at a certain location is , its value is generally negative, indicating the ratio of received signal power to transmitted power. The closer RSSl is to zero, the stronger the signal strength. Similarly, the Bluetooth Low Energy (BLE) module detects surrounding BLE beacons and records their RSSI values. Assume that the RSSI value of a BLE beacon at a certain moment is , then its value represents the signal strength and distance relationship between the beacon and the bracelet. These data are used for further indoor positioning or short-range device communication application scenarios. The sampling frequency of each sensor is dynamically adjusted based on the user's motion state. Assume that a motion state variable Represents the current state of the user, where Indicates a stationary state. Indicates the motion state. When the user is detected to be stationary, , reduce the sampling frequency of GPS and IMU, and set the sampling intervals to and , increase them to and ,in and represent the sampling intervals in static and moving states respectively. For Wi-Fi and BLE scanning intervals, they are increased to and , thereby reducing energy consumption. When the user is in motion, , then increase the sampling frequency of GPS and IMU to reduce the sampling interval to and , while reducing the scanning interval of Wi-Fi and BLE to and , obtain more dynamic data and improve the accuracy of positioning and motion estimation. After completing dynamic sampling, the collected raw GPS data, IMU data, Wi-Fi data and BLE data are processed for outlier removal. Suppose a data set is defined ,in , each data Corresponding to different sensor information. Use statistical methods, such as the 3σ principle, to determine whether the data is abnormal. For a certain feature , calculate its mean and standard deviation , if a data point satisfy , then the data is considered an outlier and should be removed. After processing, the preliminarily cleaned multi-source data is obtained. The filtering and smoothing algorithm is applied to the cleaned multi-source data to eliminate high-frequency noise and fluctuations in the data and obtain the filtered multi-source data. Assuming that the Kalman filter is used to smooth the data, for the state estimation at a certain moment , the state update formula of Kalman filter is: ; in, is the state estimate at the previous moment, is the Kalman gain, is the observed value at the current moment, is the observation matrix. Through the iterative update of the filter, the high-frequency noise in the data is smoothed out to obtain a more stable state estimate. The filtered multi-source data is time synchronized to unify the data of different sensors to the same time base. Assume that the timestamps of all sensors are and , through linear interpolation or time alignment algorithm, all data are mapped to a unified time point , and obtain the preprocessed multi-source data.

[0027] In a specific embodiment, the process of executing step 200 may specifically include the following steps: Extract satellite geometric distribution features from preprocessed GPS data, calculate horizontal precision factor and vertical precision factor, and obtain GPS feature vector; perform Fourier transform on preprocessed IMU data, extract frequency domain features, and obtain IMU feature vector; construct RSSI fingerprint vector from preprocessed Wi-Fi data, and use principal component analysis to reduce the dimension of RSSI fingerprint vector to obtain Wi-Fi feature vector after dimension reduction; apply compressed sensing algorithm to preprocessed BLE data, extract significant features in sparse RSSI fingerprint space, and obtain BLE feature vector; The GPS feature vector, IMU feature vector, Wi-Fi feature vector and BLE feature vector are concatenated to obtain a multi-source feature vector; Applying a time embedding algorithm to the multi-source feature vector, using sinusoidal position coding to convert absolute time into a periodic representation, and obtaining a time embedding vector, and applying a space embedding algorithm to the multi-source feature vector, using grid coding to discretize the prior position information into a grid index of a fixed resolution, and obtaining a space embedding vector; The multi-source feature vector is concatenated with the time embedding vector and the space embedding vector, and divided into time windows of fixed length to obtain the embedded multi-source feature sequence.

[0028] Specifically, feature extraction is performed on each type of preprocessed data. For GPS data, the geometric distribution features of the satellites are extracted, and the horizontal dilution of precision (HDOP) and vertical dilution of precision (VDOP) are calculated to obtain a feature vector that comprehensively represents the quality of the GPS signal. The satellite geometric distribution features are reflected by the distribution of the positions of each satellite in the sky. Assume that the position of each visible satellite is , the receiver location is , then the unit vector from the receiver to each satellite is It can be expressed as: ; Constructing the geometric distribution matrix , each line of which is the unit vector of each satellite, in the form of: ; According to the matrix , calculate the position accuracy matrix ; in, for The transposed matrix of The diagonal elements of represent the variance of the position estimate. The horizontal precision factor HDOP and the vertical precision factor VDOP are: ; in, and They are The first and second diagonal elements of represent the horizontal position error. is the third diagonal element, indicating the vertical error. These precision factors comprehensively reflect the current GPS positioning quality. HDOP and VDOP are combined with other geometric distribution characteristics (such as signal-to-noise ratio, number of satellites, etc.) to form the GPS feature vector For the preprocessed IMU data, including the three-axis accelerometer and three-axis gyroscope data, Fourier transform is applied to extract the frequency domain features. Suppose the three-axis acceleration data at a certain moment is , and perform a discrete Fourier transform (DFT): ; in, is the acceleration component in the frequency domain, is the number of sampling points, is the frequency, is an imaginary unit. Fourier transform converts time domain signals into frequency domain signals, which can reflect the amplitude and phase information of different frequency components. The main frequency components and corresponding amplitude characteristics are extracted from them, such as the main peak frequency and its amplitude , combine multiple features to form an IMU feature vector , thus reflecting the user's motion status and change trend. For the pre-processed Wi-Fi data, the RSSI fingerprint vector is constructed. Suppose at a certain moment, the bracelet scans Wi-Fi access points, and the RSSI values ​​of each access point are . Arrange multiple RSSI values ​​into a vector: ; Since the RSSI vector has a high dimension, in order to reduce the computational complexity, principal component analysis is used to reduce its dimension and find a set of orthogonal principal components to maximize the projection variance of the data on these principal components. Let the original vector be , the dimension reduction vector after principal component analysis is: ; Among them, the matrix It is composed of principal component eigenvectors, is the Wi-Fi feature vector after dimensionality reduction. It can retain the main information of the original data at a lower dimension. For the preprocessed BLE data, since the BLE signal is sparse and susceptible to environmental interference, the compressed sensing algorithm is used for feature extraction. In the sparse RSSI fingerprint space, the RSSI value of the BLE beacon is represented by a sparse vector. Let the original vector of the RSSI fingerprint be , define a compressed sensing matrix , sparse features are extracted by solving the following optimization problem: ; in, is a sparse representation vector, express of The norm is the sum of the absolute values ​​of all its elements. Solving this optimization problem yields the sparsest feature representation , as the BLE feature vector , which can effectively extract significant signal features in a high-noise environment. , IMU feature vector , Wi-Fi feature vector after dimensionality reduction and BLE feature vector Perform concatenation to obtain multi-source feature vectors: ; The multi-source feature vector integrates multi-dimensional information from different sensors and can fully reflect various information in the current environment. The time embedding algorithm is applied to the multi-source feature vector. Time embedding uses sinusoidal position encoding to convert absolute time into periodic representation. Assume that the absolute time is , then its time embedding vector for: ; in, is the dimension of the embedding vector, is the index of the feature. In this way, the temporal information is mapped into a high-dimensional periodic space, so that the model can capture the law of temporal changes. At the same time, a spatial embedding algorithm is applied to the multi-source feature vector, and the prior position information is discretized into a grid index of fixed resolution using grid coding. Assume that the environment is divided into Each grid cell has a unique index , then a certain position The grid index is: ; in, and is the side length of the grid unit. Each position is mapped to a grid index, which is then converted into a one-hot encoding to obtain a spatial embedding vector The multi-source feature vector is concatenated with the time embedding vector and the space embedding vector to form the final embedding vector. The embedding vector is divided into time windows of fixed length, each of which contains the multi-source embedded data of a time series segment, and the embedded multi-source feature sequence is obtained.

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: The embedded multi-source feature sequences are respectively input into the independent neural network submodules of the feature processing layer, and the GPS and IMU feature sequences are processed using the long short-term memory network, and the Wi-Fi and BLE feature sequences are processed using the graph convolutional neural network to obtain the feature representation after preliminary processing; the feature processing layer includes: GPS feature processing submodule, IMU feature processing submodule, Wi-Fi feature processing submodule and BLE feature processing submodule; The self-attention mechanism is applied to the feature representation after preliminary processing to calculate the similarity matrix of features in the same data source, and the important features are highlighted through weighted summation to obtain enhanced single-source feature representation; A cross-attention mechanism is applied to the enhanced single-source feature representation to calculate the correlation matrix between features from different data sources, realize cross-modal information interaction, and obtain the fused feature representation; The fused feature representation is input into the fully connected layer and mapped to a unified latent space to obtain a latent space feature vector; the fully connected layer includes: input layer, hidden layer and output layer, where the hidden layer uses the ReLU activation function; A regression head network is applied to the latent space feature vector to output the initial position coordinates to obtain the initial position estimation result, and a classification head network is applied to the latent space feature vector to output the discretized probability distribution to obtain the position uncertainty.

[0030] Specifically, we design corresponding neural network structures for the characteristics of different data sources. For GPS and IMU feature sequences, since they have strong correlation in the time dimension and can reflect the changing trend of the user's position and motion state, we choose to use the long short-term memory network (LSTM) to process these two types of features. The LSTM network is a recursive neural network structure that can capture the long-term and short-term dependencies in time series data. Through memory units and forget gate mechanisms, it effectively solves the problems of gradient vanishing and gradient explosion in traditional RNNs in long time series. Suppose the embedded GPS feature sequence is ,in Indicates GPS feature vector at time, is the total length of the sequence. For the input at each moment, the state update formula of the LSTM network is: in, represents the forget gate, is the input gate, is the memory unit state, is the output gate, is the hidden state, is the Sigmoid activation function, Represents an element-wise multiplication operation. The weight matrix as well as are the weights of different gating mechanisms, is the bias term. Through the above state update formula, the LSTM network can effectively extract the time-dependent information in the GPS feature sequence and obtain the GPS feature representation after preliminary processing. Similarly, for the IMU feature sequence , processed using an LSTM network. IMU data usually contains timing information from a three-axis accelerometer and gyroscope, which can reflect changes in the user's motion state. Through the memory and forgetting mechanism of the LSTM network, the periodic motion patterns and mutation features in the IMU data are captured to obtain a refined representation of the motion state. For Wi-Fi and BLE feature sequences, they have different structural characteristics. The RSSI values ​​of Wi-Fi and BLE signals usually reflect the distance relationship between the device and the access point or beacon, so they have strong topological structural characteristics in space. In order to capture the spatial relationship in these data, a graph convolutional neural network (GCN) is used to process the Wi-Fi and BLE feature sequences. Suppose the Wi-Fi feature sequence is , where each vector Representative RSSI values ​​of different access points at different times. Construct RSSI values ​​into a graph structure ,in is a node set, representing each access point, is an edge set, which represents the connectivity between different access points. The propagation rule of GCN is: in, For the The node feature representation of the layer, is the adjacency matrix of the graph, which represents the connection relationship between access points. is the degree matrix, , that is, the degree of each node, For the The weight matrix of the layer, is the activation function (usually ReLU). GCN can effectively capture the spatial relationship between nodes and extract the spatial information of Wi-Fi features by weighted aggregation of neighbor node features. , build graph structure , represents the spatial relationship between different BLE beacons. Through the GCN network, the RSSI values ​​in the BLE feature sequence are weighted and aggregated to extract the BLE feature representation with spatial topological relationship. The GPS, IMU, Wi-Fi and BLE feature sequences are respectively input into the corresponding feature processing submodules. After LSTM and GCN processing, the preliminarily processed feature representation is obtained, which contains the time and space information of different data sources. The self-attention mechanism is applied to the preliminarily processed feature representation. The self-attention mechanism identifies key features and assigns higher weights by calculating the similarity matrix between different features in the same data source. Suppose the preliminarily processed feature representation is , where each is a feature vector. The self-attention mechanism calculates the self-attention weight matrix The formula is: ; in, and are the weight matrices for query and key respectively, is the dimension of the key vector. Attention weight It is used to measure the correlation between different features. With feature representation Perform weighted summation to obtain enhanced single-source feature representation: ; A cross-attention mechanism is applied to the enhanced single-source feature representation. The cross-attention mechanism realizes the interaction of cross-modal information by calculating the correlation matrix between features from different data sources. Assume that the feature representations from different data sources are and , then the correlation matrix The calculation formula is as follows: ; in, , Through the cross-attention mechanism, feature information from different modalities is transmitted to each other, complementing each other's contextual information, and obtaining a richer and more comprehensive fused feature representation. The fused feature representation is input into the fully connected layer, mapped into a unified latent space, and the latent space feature vector is obtained. The fully connected layer includes an input layer, a hidden layer, and an output layer. Let the fused feature vector be , then the mapping process of the fully connected layer is as follows: ; ; in, and are the weight matrices of the hidden layer and the output layer respectively, and is the corresponding bias term, ReLU is the activation function of the hidden layer, is the output of the hidden layer, is the result of the output layer, i.e., the latent space feature vector. Through this mapping process, the latent space feature vector extracts key low-dimensional information from the high-dimensional fusion features, which is convenient for subsequent regression and classification tasks. In order to obtain the estimation result of the initial position, the regression head network is applied to the latent space feature vector. Let the latent space feature vector be , the initial position coordinates are , then the output formula of the regression head network is: ; in, is the weight matrix of the regression head network, is the bias term. The regression head network can map the latent space feature vector to a specific coordinate space to obtain the initial position estimation result , which represents the location information of the smart bracelet at the current moment. At the same time, in order to estimate the uncertainty of the position, the classification head network is applied to the latent space feature vector. Let the classification head network output the discretized probability distribution , then its output formula is: ; in, is the weight matrix of the classification head network, is the bias term. The softmax function maps the output of the classification head network to a probability distribution, which represents the confidence distribution of the current position prediction. Through this probability distribution, the accuracy of the current position estimation is evaluated to obtain the uncertainty information of the position.

[0031] In a specific embodiment, the process of executing step 400 may specifically include the following steps: A factor graph model is established based on the initial position estimation result, and the position estimation in the continuous time period is represented as nodes in the graph to obtain the initial pose graph; the factor graph model includes: position nodes, observation factors and motion model factors; An absolute position factor is established for the GPS observation in the initial pose graph, a relative pose factor is constructed for the IMU data, and a pseudorange factor is established for the Wi-Fi and BLE observations to obtain a multi-source observation factor graph; the multi-source observation factor graph includes: GPS factor, IMU factor, Wi-Fi factor, and BLE factor; A nonlinear factor recovery algorithm is applied to the multi-source observation factor graph to extract key information from the dense prior factors generated by marginalization to obtain a sparse pose graph. The covariance matrix of each factor in the sparse pose graph is dynamically adjusted based on position uncertainty to obtain an adaptive pose graph. Incremental smoothing and mapping algorithms are applied to the adaptive pose graph to construct a sparse information matrix. At the same time, a Levenberg-Marquardt optimizer is constructed, and the damping factor and number of iterations are set to obtain an iterative optimization strategy. An iterative optimization strategy is executed on the adaptive pose graph, and multiple rounds of iterative calculations are performed to obtain the optimized node positions and edge covariances. Based on the optimized node positions and edge covariances, the position coordinates and the complete covariance matrix at the latest moment are extracted to obtain the optimized position information.

[0032] Specifically, the position estimation results in the continuous time period are represented as nodes in the graph to obtain the initial pose graph. In the factor graph model, the nodes are used to represent the position information of the smart bracelet at each time step, while the observation factor and motion model factor are used to represent the data constraints of different sensors and the motion relationship between adjacent time steps. Assume that in a time period there is time steps, and the initial position of each time step is estimated to be , then each All of them are used as a position node in the factor graph model. For these position nodes, appropriate observation factors and motion model factors are established. For the GPS observations in the initial pose graph, an absolute position factor is established. GPS data usually provides absolute geographic location. The absolute position factor directly links the GPS observation data with the corresponding time step position node. Suppose the GPS observation value at a certain time step is , and its measurement equation is: ; in, The location node A function that maps measurements into GPS coordinates, is the measurement noise, which is usually assumed to have zero mean and Gaussian distribution, expressed as ,in is the covariance matrix of GPS measurement noise. For IMU data, since IMU provides relative motion information between adjacent time steps, relative pose factors are established. Assume that a certain time step The IMU observation value is , and its measurement equation is: ; in, Is to place adjacent nodes and Mapped to the function of IMU measurement space, indicating that the smart bracelet is from Move to The relative motion state at any moment, is the IMU measurement noise, assuming it obeys For Wi-Fi and BLE observations, since they usually provide relative position or distance information, pseudo-range factors are used to represent them. The Wi-Fi or BLE observation value is , and its measurement equation is: ; in, The location node and Wi-Fi access point or BLE beacon locations Mapped as a function of the RSS1 measurement, expressed as a logarithmic relationship of distance: ; in and are the path loss model parameters, is the RSS1 measurement noise, assuming . By associating the above observation factors with the corresponding time step position nodes, a multi-source observation factor graph is constructed, which contains GPS factors, IMU factors, Wi-Fi factors and BLE factors. These factors provide data constraints from different sources for each position node in the factor graph. After obtaining the multi-source observation factor graph, a nonlinear factor recovery algorithm is applied to extract key information from the dense prior factors generated in the marginalization process to obtain a sparse pose graph. Marginalization refers to eliminating some unnecessary variables from the factor graph, thereby reducing the computational complexity. Let the dense prior factor be , whose covariance matrix is , the marginalized sparse factor covariance matrix is , sparsification is achieved by solving the following optimization problem: ; in, is the Frobenius norm, which is used to measure the density of a dense matrix. and sparse matrices The difference between It is a matrix norm, used to promote the sparsity of the matrix, is a regularization parameter used to balance the weights between the two objectives. By optimizing the problem, a sparse pose graph is obtained, in which only the most important factors and information are retained to reduce computational complexity. After obtaining the sparse pose graph, the covariance matrix of each factor in the sparse pose graph is dynamically adjusted based on position uncertainty. Suppose the sparse factor covariance matrix is , the position uncertainty is , , the covariance matrix of each factor is dynamically adjusted as follows: ; in, is the adjusted covariance matrix, is the covariance matrix of position uncertainty. This adjustment can adaptively optimize the factor graph according to the confidence of nodes at different positions, making the overall graph model more stable and robust. Incremental smoothing and mapping algorithms are applied to the adaptive pose graph to construct a sparse information matrix. Let the sparse information matrix be , which is defined as the inverse of the covariance matrix of all nodes and edges in the factor graph, that is: ; At the same time, a Levenberg-Marquardt optimizer is constructed for nonlinear least squares optimization. The Levenberg-Marquardt algorithm combines the advantages of gradient descent and Newton's method and has good convergence in nonlinear optimization. Its update formula is: ; in, For the The solution of the iteration, is the Jacobian matrix, which represents the partial derivative of the error term with respect to the position node, is the residual vector, which represents the difference between the observed value and the model predicted value, is the damping factor, is the unit matrix. By adjusting the damping factor And set the number of iterations to effectively avoid the divergence problem and obtain the iterative optimization strategy. Execute the iterative optimization strategy on the adaptive pose graph and perform multiple rounds of iterative calculations until the node position and edge covariance converge. Get the optimized node position and the marginal covariance matrix Based on the optimized node position and edge covariance matrix, the precise position coordinates and the complete covariance matrix at the latest moment are extracted to obtain the optimized position information.

[0033] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Based on the optimized location information, a global topology map is constructed, and the connection relationship between different areas is represented by the graph structure to obtain a macro-level environment representation; the global topology map includes: nodes, edges and weights; Extract local environmental features from the optimized location information, construct an occupancy grid map, accurately describe the local environmental geometric features, and obtain a micro-level environmental representation; the occupancy grid map includes: grid units, occupancy probability, and update rules; The macro-layer environmental representation and the micro-layer environmental representation are integrated to construct a multi-layer environmental map to obtain a complete environmental representation; the multi-layer environmental map includes: topological layer, measurement layer and semantic layer; Apply the A algorithm to the global topology map in the multi-level environment map to perform a rough path search to obtain the global navigation direction. Based on the global navigation direction, apply the dynamic window method to the local occupancy grid map, consider the user's motion characteristics and environmental obstacles, and obtain the local fine path. The local fine path is smoothed, and a continuous curve is generated using a cubic spline interpolation algorithm to obtain a smooth navigation path. Based on the smooth navigation path, hierarchical navigation instructions are generated, including direction instructions at the macro level and fine guidance at the micro level, to obtain multi-level navigation instructions; the multi-level navigation instructions include: text instructions, voice prompts, and vibration feedback; Multi-level navigation instructions are optimized and combined, and the instruction density is dynamically adjusted according to user preferences and environmental complexity to obtain navigation paths and navigation instructions.

[0034] Specifically, the optimized location information is used as the basis of the global topology map, and the graph structure is used to represent the connection relationship between different areas. The global topology map is an abstract representation of the environment, in which nodes represent different key locations or areas, edges represent the traversable paths between these areas, and weights reflect the cost of the path, such as distance, time or difficulty. Let the global topology map be represented as ,in Represents a collection of nodes. represents the edge set, Represents a weight set. Assume that there are multiple rooms, corridors, and entrances and exits in a building. The center point of each room, the intersection of the corridor, and the entrance and exit are all nodes. . Two nodes and The edge between represents the traversable paths between them, and the weight It means from the node To Node The toll cost is expressed by the following formula: ; in, Represents the physical distance between two nodes. Indicates the travel time. Indicates the difficulty of the path (such as slope, obstacle density, etc.), and is a weight parameter used to balance different factors. In this way, the areas within the entire building and their connections are abstracted into a global topological map, and the macro-level environmental representation can effectively depict the structure and accessibility between different areas. The local environmental features are extracted from the optimized location information, and an occupancy grid map is constructed to describe the geometric characteristics of the local environment. The occupancy grid map divides the space into multiple grid cells, each of which represents the passability of a certain area or whether it is occupied by an obstacle. Let the occupancy grid map be represented as a matrix , where each grid cell With occupation probability , which represents the probability that the unit is occupied by an obstacle. The occupancy probability is estimated by sensor data. Assuming that the sensor observation data is , then the update rule of occupation probability can be expressed as: ; in, Given observation data The probability that a grid cell is occupied when is the prior probability (usually initialized to 0.5), is the overall probability of the observed data. By the Bayesian update rule, each new observation When it arrives, the probability of occupying the grid cell As sensor data is continuously collected, the occupancy grid map dynamically reflects the obstacle distribution and traversable areas in the current environment. The global topology map represented by the macro-layer environment is integrated with the occupancy grid map represented by the micro-layer environment to construct a multi-level environment map. The multi-level environment map contains three levels of representation: topological layer, measurement layer, and semantic layer. The topological layer reflects the structure and connection relationship between different areas, the measurement layer reflects specific geometric information, such as distance and location, and the semantic layer contains higher-level environmental information, such as room type, functional area, etc. Suppose the multi-level environment map is , where T represents the topological layer, represents the measurement layer, Represents the semantic layer. The information in the topology layer comes from the global topology map. , the information in the metric layer comes from the occupancy grid map The information in the semantic layer comes from predefined environmental semantic tags (such as room function, building structure, etc.). Based on the multi-level environmental map, multi-level path planning is performed. Skip the path search and get the global navigation direction. The algorithm is a heuristic search algorithm that can find the shortest path from the starting point to the end point in the topological graph. Its cost function It can be expressed as: ; in, From the starting point to the node The known cost, Represents a slave node The heuristic estimated cost to the end point is usually expressed as Euclidean distance or Manhattan distance. The A* algorithm gives priority to nodes with the smallest cost function f(n) during the search process to quickly find the global optimal path. The dynamic window method is applied in the process to obtain a local refined path by considering the user's motion characteristics and the distribution of obstacles in the environment. The dynamic window method searches for a series of speed combinations in the speed space. , and evaluate their path safety and target proximity in the future time window. Its objective function is: ; Among them, heading Indicates speed combination The deviation from the target direction in the current time window, distance Indicates the minimum distance to obstacles on the path. Indicates the current speed. is the weight parameter. By maximizing the objective function , find the optimal speed combination and generate a safe and efficient local fine path. After obtaining the local fine path, in order to make the path smoother and meet the actual motion requirements, the local path is smoothed. The cubic spline interpolation algorithm is used to generate a continuous curve. Let the discrete path points on the local path be , the goal of cubic spline interpolation is to construct a set of cubic polynomial functions to smoothly connect these path points. Each cubic polynomial The form is: ; in, are the coefficients to be solved, and these polynomials are required to be continuous at each path point and have continuous first-order and second-order derivatives. By solving these coefficients, a smooth cubic spline curve is obtained, which represents a continuous navigation path. Based on the smooth navigation path, hierarchical navigation instructions are generated. Hierarchical navigation instructions include direction instructions at the macro level and detailed guidance at the micro level. These instructions are presented to users in various ways such as text instructions, voice prompts, and vibration feedback, so that users can get clear navigation guidance in different environments. In order to improve the navigation experience, the multi-level navigation instructions are optimized and combined, and the instruction density is dynamically adjusted according to the user's preferences and the complexity of the environment. For example, in an environment familiar to the user, the navigation instructions are relatively brief, while in an unfamiliar or complex environment, the navigation instructions need to be more detailed and frequent. Dynamic adjustment reduces the user's cognitive burden and provides more accurate navigation information when needed.

[0035] In a specific embodiment, executing the smart bracelet position tracking and navigation method also includes the following steps: Based on Bluetooth mesh technology, a communication network between multiple smart bracelets is established, and the communication protocols for location information exchange, motion status sharing, and navigation intent broadcasting are defined to obtain a distributed multi-smart bracelet communication architecture. Encode the position, motion state and navigation intention of each smart bracelet, build a state encoder, convert high-dimensional raw data into low-dimensional feature vectors, and obtain the state representation of the smart bracelet; Based on the topological structure in the communication architecture, a graph attention network is constructed to perform weighted aggregation on the state information between different smart bracelets to obtain group state features. The group state features are then input into the state decoder to reconstruct the optimized position estimation and consistent motion state of each smart bracelet to obtain the collaborative positioning result. Apply group consistency constraints to the collaborative positioning results to minimize the relative position and speed differences between adjacent smart bracelets to obtain a globally consistent group state. Based on the globally consistent group state, a multi-agent reinforcement learning environment is constructed. Each smart bracelet is taken as an agent, and the state space, action space and reward function are defined to obtain a group navigation learning framework. A distributed proximal strategy optimization algorithm is applied to the group navigation learning framework to train the navigation strategy network of each smart bracelet to obtain the individual navigation strategy. Based on the principle of federated learning, the navigation experience of multiple smart bracelets is aggregated while protecting user privacy, and the global navigation strategy model is updated to obtain the initial group navigation strategy. Using the initial group navigation strategy, based on the user's historical trajectory and current navigation intention, potential group rally points are identified, and a recommended list of group rally points is generated to obtain a social navigation function. The social navigation function is then personalized and adjusted. According to user preferences and dynamic changes in the environment, the navigation strategy is adaptively updated to obtain the target group navigation strategy.

[0036] Specifically, an efficient distributed communication architecture is constructed to realize information exchange, status sharing and navigation intention broadcasting among multiple smart bracelets. In a Bluetooth mesh network, each smart bracelet acts as a node and establishes communication links with other surrounding bracelet nodes. The characteristic of a Bluetooth mesh network is that each node can not only send and receive information, but also act as a relay node to forward and diffuse information in the network, forming an efficient distributed network structure. In this communication architecture, communication protocols for location information exchange, motion status sharing and navigation intention broadcasting are defined. The location information exchange protocol is used to transmit the current location information of each bracelet between smart bracelets in real time, ensuring that all nodes can obtain the latest spatial location information. Assume that the bracelet The location information is , each bracelet regularly broadcasts its location information in the network , and receive the location information of other bracelets to form a global location information matrix ,in Indicates the number of wristband nodes in the network. The motion state sharing protocol is used to transmit the motion state information of the wristband. The motion state of the wristband includes information such as speed and acceleration. The motion state is and acceleration , each bracelet regularly broadcasts its movement status and , forming a motion state matrix and the acceleration matrix The navigation intention broadcast protocol is used to transmit the user's navigation target and route between wristbands. The navigation intent is ,in represents the angle of the target direction, Indicates the target distance. Each wristband broadcasts its navigation intention periodically , forming a navigation intention matrix Through the collaborative work of these three protocols, it is possible to achieve real-time and efficient exchange of status information and navigation intentions between multiple smart bracelets, and build a distributed multi-smart bracelet communication architecture. After the communication architecture is established, the position, motion state and navigation intention of each smart bracelet are encoded to build a state encoder. The purpose of the state encoder is to convert high-dimensional raw data into low-dimensional feature vectors to reduce communication overhead and data processing complexity. The original state data is , the encoded state vector is , then the mapping relationship of the state encoder is: ; in, is the weight matrix of the state encoder, is the bias term. Through the state encoder, the original high-dimensional state data Mapped to a low-dimensional feature vector , and obtain the state representation of the smart bracelet. Based on the topological structure in the communication architecture, a graph attention network is constructed to perform weighted aggregation on the state information between different smart bracelets to obtain group state features. The graph attention network can assign different attention weights between each node and its neighboring nodes to capture the complex relationship between nodes. Suppose the adjacency matrix of the smart bracelet network is , bracelet and bracelet The attention weight between , then the calculation formula is: ; in, is the weight matrix of node features, is the parameter vector used to calculate the attention coefficient, || represents the vector concatenation operation, Representation Node The set of neighbor nodes. Through the attention weight , for nodes Status information on the node Different importance weights are assigned in the feature aggregation process to obtain more accurate group state features. Through weighted aggregation of the graph attention network, the group state feature vector of the entire smart bracelet network is obtained. ,Right now: ; in, is an activation function (such as ReLU) used to introduce nonlinear characteristics. After obtaining the group state features, they are input into the state decoder to reconstruct the state of each smart bracelet to obtain the optimized position estimation and consistent motion state to form a collaborative positioning result. Suppose the decoded bracelet The optimized position is , the motion state is , the mapping relationship of the state decoder is: ; in, is the weight matrix of the state decoder, is the bias term. Through the reconstruction of the state decoder, the group state features are mapped to the specific position and motion state of each smart bracelet, and the optimized collaborative positioning result is obtained. After obtaining the collaborative positioning result, the group consistency constraint is applied to it. The goal of the group consistency constraint is to minimize the relative position and speed differences between adjacent smart bracelets, so that all smart bracelets in the group maintain a consistent motion state. Assume that the adjacent smart bracelets and The relative position error between , the relative speed error is , then the objective function of the group consistency constraint is: ; in, represents the set of adjacent node pairs in the smart bracelet network. By minimizing the objective function , so that all smart bracelets in the group maintain consistency in position and speed, and obtain a globally consistent group state. Based on the globally consistent group state, a multi-agent reinforcement learning environment is constructed. Each smart bracelet is regarded as an agent, and the state space, action space and reward function are defined for it. The state space of , the action space is , the reward function is , represents the relationship between the current position of the bracelet and the navigation target. By defining a multi-agent reinforcement learning framework, the navigation strategy of each smart bracelet is trained in the group. In the multi-agent reinforcement learning framework, the distributed proximal policy optimization algorithm is applied to train the navigation strategy network of each smart bracelet. The distributed proximal policy optimization algorithm is a policy gradient method that can dynamically adjust the step size of policy update during the training process to maintain the stability and convergence of the strategy. Through distributed training, each smart bracelet learns the navigation strategy independently, and through the principle of federated learning, the navigation experience of multiple smart bracelets is aggregated under the premise of protecting user privacy, and the global navigation strategy model is updated to obtain the initial group navigation strategy. Using the initial group navigation strategy, based on the user's historical trajectory and current navigation intention, potential group assembly points, such as the user's gathering place and navigation target, are identified, and a recommended list of group assembly points is generated. Through the social navigation function, the smart bracelet can help users find other group members more conveniently in a complex environment, and through personalized adjustment, the navigation strategy is adaptively updated according to the user's preferences and dynamic changes in the environment, providing users with accurate and personalized navigation services, and finally obtaining the target group navigation strategy.

[0037] The above describes the smart wristband position tracking and navigation method in the embodiment of the present application. The following describes the smart wristband position tracking and navigation device 10 in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the smart wristband position tracking and navigation device 10 includes: The acquisition module 11 is used to collect multi-source positioning data through the multi-sensor system built into the smart bracelet, and pre-process it to obtain pre-processed multi-source data; An extraction module 12 is used to perform feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence; A processing module 13 is used to input the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; A construction module 14 is used to construct an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimize the adaptive pose graph to obtain optimized position information; The planning module 15 is used to generate a multi-level environment map according to the optimized location information, and perform dynamic path planning to obtain a navigation path and navigation instructions.

[0038] Through the synergy of the above components, by integrating multiple sensor data such as GPS, IMU, Wi-Fi and BLE, the positioning accuracy and robustness are improved, and reliable location information can be continuously provided in different indoor and outdoor environments. The sensor sampling frequency is dynamically adjusted according to the user's motion state, which effectively reduces energy consumption while ensuring positioning accuracy and prolongs the use time of the smart bracelet. Targeted feature extraction methods, such as principal component analysis of Wi-Fi data and compressed sensing algorithm for BLE data, greatly reduce the data dimension and reduce the computational complexity of subsequent processing. A multi-source information fusion model including feature processing layer, self-attention mechanism and cross-attention mechanism is designed, which can effectively capture the spatiotemporal correlation between different data sources and improve positioning accuracy. Through nonlinear factor recovery and dynamic covariance adjustment, an adaptive pose graph is constructed, which improves optimization efficiency and positioning accuracy, while reducing storage requirements. Combined with the global topology map and the local occupancy grid map, the macro and micro environmental representation is realized, providing comprehensive environmental information for subsequent path planning. The combination of A* algorithm and dynamic window method not only ensures global optimality, but also adapts to the local dynamic environment, improving the intelligence and flexibility of navigation. By generating hierarchical navigation instructions and optimizing instruction combinations, the interactive characteristics of smart bracelets are fully considered to provide an intuitive and easy-to-understand navigation experience.

[0039] The present application also provides an electronic device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the smart bracelet position tracking and navigation method in the above-mentioned embodiments.

[0040] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the smart bracelet position tracking and navigation method.

[0041] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0042] If the integrated unit is implemented in the form of 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 the present application is essentially 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. The computer software product is stored in a storage medium, including several instructions for an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0043] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A smart bracelet position tracking and navigation method, characterized in that: The method comprises: The multi-source positioning data collected by the multi-sensor system built into the smart bracelet is pre-processed to obtain pre-processed multi-source data; Performing feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence; Inputting the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; Constructing an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimizing the adaptive pose graph to obtain optimized position information; A multi-level environment map is generated according to the optimized position information, and dynamic path planning is performed to obtain a navigation path and navigation instructions.

2. The smart bracelet position tracking and navigation method according to claim 1, characterized in that: The multi-source positioning data collected by the multi-sensor system built into the smart bracelet is preprocessed to obtain preprocessed multi-source data, including: The satellite signal strength, number of visible satellites and pseudo-range information collected by the GPS receiver are recorded to obtain the original GPS data; the accelerometer and gyroscope three-axis data collected by the inertial measurement unit are collected to obtain the original IMU data; the wireless communication module scans the service set identifier and received signal strength indication of the surrounding Wi-Fi access points to obtain the original Wi-Fi data; the Bluetooth low energy module detects the surrounding BLE beacons and records their received signal strength indication to obtain the original BLE data; The sampling frequency of each sensor is dynamically adjusted based on the user's motion state. When the user is detected to be stationary, the sampling frequency of GPS and IMU is reduced, and the Wi-Fi and BLE scanning intervals are increased. When the user is in motion, the sampling frequency of GPS and IMU is increased, and the Wi-Fi and BLE scanning intervals are reduced to obtain adaptive sampling data. Performing outlier removal processing on the original GPS data, the original IMU data, the original Wi-Fi data, and the original BLE data to obtain preliminarily cleaned multi-source data; A filtering and smoothing algorithm is performed on the preliminarily cleaned multi-source data to obtain filtered multi-source data, and the filtered multi-source data are unified to the same time reference to obtain pre-processed multi-source data.

3. The smart bracelet position tracking and navigation method according to claim 2, characterized in that: The performing feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence includes: Extract satellite geometric distribution features from preprocessed GPS data, calculate horizontal precision factor and vertical precision factor, and obtain GPS feature vector; perform Fourier transform on preprocessed IMU data, extract frequency domain features, and obtain IMU feature vector; construct RSSI fingerprint vector from preprocessed Wi-Fi data, and use principal component analysis to reduce the dimension of RSSI fingerprint vector to obtain Wi-Fi feature vector after dimension reduction; apply compressed sensing algorithm to preprocessed BLE data, extract significant features in sparse RSSI fingerprint space, and obtain BLE feature vector; Concatenate the GPS feature vector, the IMU feature vector, the Wi-Fi feature vector, and the BLE feature vector to obtain a multi-source feature vector; Applying a time embedding algorithm to the multi-source feature vector, using sinusoidal position coding to convert absolute time into a periodic representation, to obtain a time embedding vector, and applying a space embedding algorithm to the multi-source feature vector, using grid coding to discretize prior position information into a grid index of a fixed resolution, to obtain a space embedding vector; The multi-source feature vector is concatenated with the time embedding vector and the space embedding vector, and divided into time windows of fixed length to obtain an embedded multi-source feature sequence.

4. The smart bracelet position tracking and navigation method according to claim 1, characterized in that: The embedded multi-source feature sequence is input into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty, including: The embedded multi-source feature sequences are respectively input into independent neural network submodules of the feature processing layer, the GPS and IMU feature sequences are processed using a long short-term memory network, and the Wi-Fi and BLE feature sequences are processed using a graph convolutional neural network to obtain a feature representation after preliminary processing; the feature processing layer includes: a GPS feature processing submodule, an IMU feature processing submodule, a Wi-Fi feature processing submodule, and a BLE feature processing submodule; Applying a self-attention mechanism to the feature representation after the preliminary processing, calculating a similarity matrix of features in the same data source, highlighting important features through weighted summation, and obtaining an enhanced single-source feature representation; Applying a cross-attention mechanism to the enhanced single-source feature representation, calculating a correlation matrix between features from different data sources, realizing cross-modal information interaction, and obtaining a fused feature representation; Input the fused feature representation into a fully connected layer, map it to a unified latent space, and obtain a latent space feature vector; the fully connected layer includes: an input layer, a hidden layer, and an output layer, wherein the hidden layer uses a ReLU activation function; A regression head network is applied to the latent space feature vector to output the initial position coordinates to obtain the initial position estimation result, and a classification head network is applied to the latent space feature vector to output a discretized probability distribution to obtain the position uncertainty.

5. The smart bracelet position tracking and navigation method according to claim 1, characterized in that: The step of constructing an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimizing the adaptive pose graph to obtain optimized position information includes: A factor graph model is established based on the initial position estimation result, and the position estimation in the continuous time period is represented as a node in the graph to obtain an initial pose graph; the factor graph model includes: position nodes, observation factors and motion model factors; Establishing an absolute position factor for the GPS observation in the initial pose graph, constructing a relative pose factor for the IMU data, and establishing a pseudorange factor for the Wi-Fi and BLE observations to obtain a multi-source observation factor graph; the multi-source observation factor graph includes: GPS factor, IMU factor, Wi-Fi factor and BLE factor; Applying a nonlinear factor recovery algorithm to the multi-source observation factor graph, extracting key information from the dense prior factors generated by marginalization to obtain a sparse pose graph, and dynamically adjusting the covariance matrix of each factor in the sparse pose graph based on the position uncertainty to obtain an adaptive pose graph; Applying an incremental smoothing and mapping algorithm to the adaptive pose graph to construct a sparse information matrix, and at the same time, constructing a Levenberg-Marquardt optimizer, setting a damping factor and the number of iterations, and obtaining an iterative optimization strategy; The iterative optimization strategy is executed on the adaptive pose graph, and multiple rounds of iterative calculations are performed to obtain optimized node positions and edge covariances, and based on the optimized node positions and edge covariances, the position coordinates and the complete covariance matrix at the latest moment are extracted to obtain optimized position information.

6. The smart bracelet position tracking and navigation method according to claim 1, characterized in that: The generating of a multi-level environment map according to the optimized location information and performing dynamic path planning to obtain a navigation path and navigation instructions includes: Based on the optimized location information, a global topology map is constructed, and the connection relationship between different areas is represented by a graph structure to obtain a macro-level environment representation; the global topology map includes: nodes, edges and weights; Extracting local environmental features from the optimized location information, constructing an occupancy grid map, accurately describing the local environmental geometric features, and obtaining a micro-level environmental representation; the occupancy grid map includes: grid units, occupancy probabilities, and update rules; The macro-layer environment representation and the micro-layer environment representation are integrated to construct a multi-layer environment map to obtain a complete environment representation; the multi-layer environment map includes: a topological layer, a metric layer and a semantic layer; Applying the A algorithm to the global topology map in the multi-level environment map to perform a rough path search to obtain a global navigation direction, and based on the global navigation direction, applying a dynamic window method to the local occupancy grid map to obtain a local fine path by considering the user's motion characteristics and environmental obstacles; The local fine path is smoothed, and a continuous curve is generated using a cubic spline interpolation algorithm to obtain a smooth navigation path, and hierarchical navigation instructions are generated based on the smooth navigation path, including direction instructions at a macro level and fine guidance at a micro level, to obtain multi-level navigation instructions; the multi-level navigation instructions include: text instructions, voice prompts, and vibration feedback; The multi-level navigation instructions are optimized and combined, and the instruction density is dynamically adjusted according to user preference and environmental complexity to obtain a navigation path and navigation instructions.

7. The smart bracelet position tracking and navigation method according to claim 1, characterized in that: The smart bracelet position tracking and navigation method also includes: Based on Bluetooth mesh technology, a communication network between multiple smart bracelets is established, and the communication protocols for location information exchange, motion status sharing, and navigation intent broadcasting are defined to obtain a distributed multi-smart bracelet communication architecture. Encode the position, motion state and navigation intention of each smart bracelet, build a state encoder, convert high-dimensional raw data into low-dimensional feature vectors, and obtain the state representation of the smart bracelet; Based on the topological structure in the communication architecture, a graph attention network is constructed to perform weighted aggregation on the state information between different smart bracelets to obtain group state features, and the group state features are input into a state decoder to reconstruct the optimized position estimation and consistent motion state of each smart bracelet to obtain a collaborative positioning result; Applying group consistency constraints to the collaborative positioning results to minimize the relative position and speed differences between adjacent smart bracelets to obtain a globally consistent group state, and based on the globally consistent group state, constructing a multi-agent reinforcement learning environment, taking each smart bracelet as an agent, defining the state space, action space and reward function, and obtaining a group navigation learning framework; A distributed proximal strategy optimization algorithm is applied to the group navigation learning framework to train the navigation strategy network of each smart bracelet to obtain an individual navigation strategy. Based on the principle of federated learning, the navigation experience of multiple smart bracelets is aggregated under the premise of protecting user privacy, and the global navigation strategy model is updated to obtain an initial group navigation strategy. By using the initial group navigation strategy, based on the user's historical trajectory and current navigation intention, potential group rally points are identified, a recommended list of group rally points is generated, a social navigation function is obtained, and the social navigation function is personalized. According to user preferences and dynamic changes in the environment, the navigation strategy is adaptively updated to obtain the target group navigation strategy.

8. A smart wristband position tracking and navigation device, characterized in that: The smart wristband position tracking and navigation device comprises: The acquisition module is used to collect multi-source positioning data through the multi-sensor system built into the smart bracelet, and pre-process it to obtain pre-processed multi-source data; An extraction module is used to perform feature extraction and spatiotemporal embedding processing on the preprocessed multi-source data to obtain an embedded multi-source feature sequence; A processing module, used for inputting the embedded multi-source feature sequence into a multi-source information fusion model for processing to obtain an initial position estimation result and position uncertainty; A construction module, used to construct an adaptive pose graph based on the initial position estimation result and the position uncertainty, and optimize the adaptive pose graph to obtain optimized position information; The planning module is used to generate a multi-level environment map according to the optimized location information, and perform dynamic path planning to obtain a navigation path and navigation instructions.

9. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory so that the electronic device executes the smart bracelet position tracking and navigation method as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the smart bracelet position tracking and navigation method as described in any one of claims 1-7 is implemented.

Images