Method and system for determining position of mobile robot

Abstract

The invention relates to the technical field of robots, and discloses a method and a system for determining the position of a mobile robot, which are used for correcting indoor robot position judgment model parameters based on a wireless signal intensity change mode so as to compensate signal interference errors caused by wall blocking. Correcting the current coordinate of the robot according to the corrected model parameters by fusing the historical track and the real-time sensor information; the corrected current coordinates are combined with known map information and instantly updated map details to regulate and control path plan variables, and an uncertainty estimation matrix is updated through a result output by multi-sensor fusion to improve robustness positioning in a high dynamic environment; through the scheme of the embodiment of the invention, the problem of how to optimize the pose parameters of the mobile robot according to the environment sensing data acquired by the mobile robot so as to solve the problem of positioning error accumulation can be solved, and the robustness of the system is further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of robotics, and specifically to a method and system for determining the position of a mobile robot. Background Art

[0002] The determination of the position of a mobile robot aims to achieve high-precision and robust position determination by comprehensively applying various technical means such as environmental perception, historical trajectory analysis, real-time sensor information, map updating, and wireless signals.

[0003] The core of this method lies in continuously optimizing the pose parameters of the mobile robot according to the environmental perception data obtained by the mobile robot, thereby effectively solving the problem of cumulative positioning errors. At the same time, it combines historical trajectories and real-time sensor information to correct the current position of the robot, reducing positioning drift that may occur in complex terrains. To improve the accuracy of bypassing dynamic obstacles, this method adjusts the path planning variables in a timely manner according to the known map information and the immediately updated map details to adapt to the changing environmental requirements.

[0004] In addition, in an indoor environment, the system can correct the position determination model based on the change pattern of the wireless signal strength to avoid misjudgment caused by signal interference due to wall blocking. Finally, to enhance the robustness in a high-dynamic environment, the uncertainty estimation matrix is updated by fusing data from multiple sensors to improve the position determination result. In summary, these strategies work together to ensure that the mobile robot can maintain accurate self-positioning capabilities in various complex usage scenarios. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides the following technical solutions:

[0006] A method and system for determining the position of a mobile robot, including:

[0007] Correcting the parameters of the indoor robot position determination model based on the change pattern of the wireless signal strength to compensate for the signal interference error caused by wall blocking, and correcting the current coordinates of the robot by fusing the historical trajectory and real-time sensor information according to the corrected model parameters;

[0008] Using the corrected current coordinates, combining the known map information and the immediately updated map details to adjust the path planning variables, and updating the uncertainty estimation matrix through the result output by multi-sensor fusion to improve the robust positioning in a high-dynamic environment;

[0009] The correction of the parameters of the indoor robot position determination model based on the change pattern of the wireless signal strength further includes: obtaining the received signal strength indication (RSSI) r_i at each reference node i in the indoor environment;

[0010] Calculate the weighted average RSSI, \(R_{avg}=\sum(r_i*w_i)\), where \(w_i\) is the weight and satisfies \(\sum w_i = 1\); adjust the weight \(w_i\) by comparing it with the known standard signal attenuation function \(f(d)=AB*\log_{10}(d)\), where \(A\) and \(B\) correspond to constants and the function result of the transmission source distance factor \(d\) respectively; judge and correct the position parameter according to the formula error \(\Delta R_i = |R_i - f(d_i)|\), where \(d_i\) is the estimated distance.

[0011] Preferably, before selecting the weight, it further includes detecting each wall material and thickness \(T_j\), and associating it with the empirical database to determine the corresponding correction coefficient \(K_j\);

[0012] Use the KNN algorithm to perform cluster classification on the current point to obtain the nearest \(n\) samples \(\{S_k\}\) and their corresponding environmental labels \(E_k\);

[0013] Construct a mapping rule \(M(E, d)\) for determining how to dynamically allocate the weight \(w_i\) under given conditions;

[0014] For the nodes within the wall influence area, apply \(M(E_k, d_i)=\frac{C}{D_i^2 + C}\), where \(C\) is the compensation value and \(D_i\) represents the distance from point \(i\) to the nearest wall surface.

[0015] Preferably, it further includes:

[0016] Adopt the Kalman filter algorithm to process the offset caused by non-line-of-sight propagation;

[0017] Record the prediction deviation vectors \(V = \{\Delta p_t\}\) in the previous \(N\) iterations, which are used as prior information in the optimization process;

[0018] Use the gradient descent method to minimize the loss \(L(w)=(P_{obs}-P_{est})^T\sum^{-1}(P_{obs}-P_{est})\), where \(\sum\) represents the covariance matrix;

[0019] Update the formula according to the learning rate \(\eta\): Make the weight gradually tend to a reasonable value.

[0020] Preferably, define a threshold \(Th\) to distinguish between valid measurement values \(e\) and invalid values, i.e., \(e < RSSI_{ave}-Th\);

[0021] Combine the above threshold to screen out a high-quality data point set \(Q=\{(x,y)_k\}\) to form a sub-network \(SN\);

[0022] Divide \(Q\) into master nodes \(MN\) and slave nodes \(SN\), and then set the synchronization timestamp \(ST\);

[0023] For any pair of MN_m and SN_n, when |rssi_{m}rssi_{ave}(SN)|≥τ√Var(RSSI_SN), the relative position between the two is recalculated to ensure accuracy.

[0024] Preferably, it also includes enabling an alternative beacon Beacon when encountering a large-area metal occlusion resulting in a severe shadow effect;

[0025] The Beacon emits a special pulse sequence S, the duration and phase of which are designed to be easily recognizable without affecting the original communication system;

[0026] Collect the information P returned by each Beacon, as well as the actual position differences δx and δy;

[0027] Only when the absolute difference satisfies the condition (|δx|+|δy|)<ε (a preset small value) is the correction trusted to filter out the influence of abnormal disturbances.

[0028] Preferably, a position accuracy improvement framework based on fuzzy C-means clustering (FCM) is established;

[0029] Randomly select a group of centers μ_c from a large number of sensor readings and assign appropriate membership degrees μ_ij;

[0030] Iteratively calculate the value of J=\sum\limits_{i=1}^{C}\sum\limits_{j=1}^{N}{\mu_{ij}}^{β}|v_ip_j|^2;

[0031] If the change in J is less than the set tolerance σ after a certain round, it is considered convergent; otherwise, adjust the membership degree and repeat the above process until the convergence condition is met.

[0032] Preferably, it includes introducing the Doppler frequency shift d_f to assist in identifying the motion state, especially the velocity jumps that occur at the start or end stages;

[0033] Whenever v>=Vmax*α or d_f<=D_min*(1±δ), where α is a sensitivity coefficient, activate the supplementary verification step to lock the true position;

[0034] This operation can effectively eliminate the position estimation deviation caused by sudden acceleration and improve the overall positioning accuracy;

[0035] The highly reliable data generated under the above conditions is preferentially used for feedback correction to enhance the real-time response ability of the system.

[0036] Preferably, it includes adding the following mechanism:

[0037] Use sliding window strategy to manage the latest m groups of historical positioning data Z;

[0038] Let h(x) be the cumulative error estimate after smoothing, h(z_(tm),...,z_t)=γ*z+(1γ)*z_previous;

[0039] Each time a new reading is received, the content of Z is updated and an exponential decay average (EMA) is performed to reduce forward volatility. Once a trend reversal is observed - n consecutive generations of growth, a reset of the starting attitude is immediately triggered to prevent the accumulated distortion from continuing to expand.

[0040] Preferably, in a complex terrain environment, the visual feature F is extracted by a pre-trained deep convolutional network DCN;

[0041] Use adaptive linear discriminant analysis ALDA to map to the low-dimensional manifold U;

[0042] Evaluate the Euclidean distance between any two points on U d_{AB}=sqrt(Σ(u_B[j]u_A[j])^2);

[0043] If d_{curr,prev}<Ψ(predetermined limit), obviously impossible state transitions are eliminated to simplify the computational load.

[0044] Preferably, it includes the introduction of crowd behavior pattern recognition for highly complex dynamic scenes such as crowded areas;

[0045] The social force model SocialFM is used to simulate the dynamic interaction rules of the group and adjust the local density D_loc accordingly;

[0046] For individual k, its target direction vector_v and speed vector_s are subject to the distribution density of nearby pedestrians;

[0047] The additional monitoring procedure is triggered only when D_loc>ρ (set upper limit) and |vs_ref|≤ν (stability indicator) is established, avoiding excessive attention to the quiet interval and wasting computing resources.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0049] 1. Modify the parameters of the indoor robot position determination model based on the wireless signal strength change pattern to compensate for the signal interference error caused by wall blockage; correct the current coordinates of the robot by fusing the historical trajectory and real-time sensor information according to the modified model parameters; adjust the path planning variables by combining the corrected current coordinates with the known map information and the instantaneously updated map details; update the uncertainty estimation matrix through the result output by multi-sensor fusion to improve the robust positioning in a high-dynamic environment. Through the solution of the embodiment of the present disclosure, the problem of how to optimize the pose parameters of a mobile robot according to the environmental perception data obtained by the robot to solve the problem of cumulative positioning error can be solved.

[0050] 2. By adopting a learning and adjustment strategy based on environmental perception data, it aims to reduce the adverse impact of errors on the long-term stability of the system. In this stage, the robot collects environmental perception data and performs local mapping, using the information obtained by high-precision sensors such as vision and lidar as input to construct an adaptive correction algorithm. This process periodically analyzes the difference between the current state and the expected model, and feeds these differences back to the core control program for updating the model prediction weights and correcting the future behavior of the robot similar to this. In particular, this method uses the reinforcement learning framework to continuously iterate to find the optimal adjustment scheme to ensure that the position error remains at a low level during long-term operation. At the same time, it can automatically adapt to different scene changes and further enhance the system robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic flow chart of the specific process of the present invention;

[0052] Figure 2 It is a schematic flow chart of the indoor environment node process of the present invention;

[0053] Figure 3 It is a schematic flow chart of the process for detecting the thickness of wall materials of the present invention;

[0054] Figure 4 It is a schematic flow chart of the non-line-of-sight propagation algorithm of the present invention;

[0055] Figure 5 It is a schematic flow chart of the threshold algorithm of the present invention;

[0056] Figure 6 It is a schematic flow chart of the metal occlusion algorithm of the present invention;

[0057] Figure 7 It is a schematic flow chart of the position accuracy of the present invention;

[0058] Figure 8 It is a schematic flow chart of the speed jump of the present invention;

[0059] Figure 9 It is a schematic flow chart of the sliding window strategy management of the present invention;

[0060] Figure 10 Schematic diagram of the deep convolutional network process of the present invention;

[0061] Figure 11 Schematic diagram of the system recognition process of the present invention. Specific implementation manners

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0063] Please refer to Figures 1-11 , a method and system for determining the position of a mobile robot. This method mainly covers correcting the parameters of the position determination model based on the wireless signal strength change pattern to compensate for the signal interference error caused by wall blocking, fusing the historical trajectory and real-time sensor information to correct the current coordinates of the robot, using the corrected current coordinates to adjust the path planning variables in combination with the known map information and the instantaneously updated map details, and updating the uncertainty estimation matrix through the output result of multi-sensor fusion to improve the robust positioning in a high-dynamic environment.

[0064] The first step of this method is to correct the parameters of the indoor robot position determination model based on the wireless signal strength change pattern, so as to accurately compensate for the error caused by signal interference caused by walls or other physical obstacles. Specifically, a series of radio signal samples are collected and a database is formed at the initial stage of the robot startup or when it enters a new area. Each sample point records important parameters such as the signal power, signal-to-noise ratio, receiving frequency at that time, and the corresponding geographical marking information. Then, machine learning techniques are used to analyze the data sets collected in different environments, and an association mapping relationship is established between the input features and the position distribution in the physical space. After multiple iterations and optimizations, the system can identify the specific pattern features caused by various interference sources and integrate them into the predefined position evaluation model to adjust the weight parameters to reduce the deviation effect. For example, a service robot delivering goods inside a shopping mall may encounter a wall blocking the Wi-Fi access point, resulting in a large measurement fluctuation. At this time, the system will know through the above learning mechanism that a more strict smoothing filter should be used or other auxiliary reference objects should be added for compensation in a similar situation.

[0065] Next, after having the corrected model, we began to re-establish the accurate and stable current position coordinates by integrating the processed path data with the latest sensor output values within a period of time according to this optimized method. This involves the application of a time series filtering technique that can effectively eliminate noise and make reasonable predictions for target values containing noise components. Specifically, to ensure that the cumulative error points generated during each state transition are correctly corrected without causing a wider spread, the entire process regularly samples the information feedback from devices such as inertial sensors (e.g., accelerometers, gyroscopes) and vision systems and integrates all the original readings into a unified state equation for comprehensive evaluation and inference. For example, in an application scenario of a logistics center, an AGV (Automated Guided Vehicle) needs to continuously shuttle between multiple shelves. Since the site is relatively open, it is difficult to find a fixed external reference system. Therefore, under such conditions, it relies on its own SLAM (Simultaneous Localization and Mapping) algorithm to obtain more reliable navigation support after optimizing the position estimation ability, avoiding accidental collisions.

[0066] With the determination of the current coordinates, the next task is to revise and implement the path strategy by using the newly determined positioning conclusion, referring to the existing static terrain file, and adding newly detected elements. This means continuously synchronously perceiving the surrounding situation while maintaining a clear plan for the travel route in the next period of time. Specifically, when a dynamic entity appears within the pre-set safe distance, it will trigger the emergency avoidance subsystem to quickly calculate a new optimal alternative path and notify the relevant control unit to adjust actions such as speed, direction, etc., ensuring that the task is successfully completed without being affected too much by external mutation factors. For example, in the case of an automated mechanical sweeper for airport runway cleaning, in addition to following the fixed work area restrictions, it also has to deal with possible human factor disturbances such as changes in construction enclosures and temporary parking area demarcations. At this time, with the help of high-precision map construction capabilities and flexible road change measures, it can ensure high operating efficiency and is not prone to errors.

[0067] The last step is to dynamically adjust the uncertainty measurement scale according to the feedback from the multi-source sensor group, thereby enhancing the adaptability level in complex working conditions. The sources of uncertainty are diverse, including random interference at the perception level and the fuzzy nature of physical rules, etc. Considering these issues, a self-organizing Bayesian update network is proposed. It can mine valuable information from a large amount of multi-source heterogeneous detection data, establish a cognitive model of the interdependent structure between the probabilities of each independent node, so that the final position conclusion is more credible and reliable and has certain fault tolerance characteristics. In practical applications, for example, the smart home patrol and monitoring assistant can quickly respond to special events such as sudden fire alarms with the support of powerful intelligent decision-making. Even when the firelight obscures the field of view, it can still accurately locate the exact geographical location of the anomaly through the mutual verification and cross-comparison of multiple functions such as sonar ranging and temperature sensing, facilitating the rescue and evacuation work and winning the initiative to reduce the potential loss risk. In short, the above several core operation steps are closely connected to jointly form an efficient and perfect design idea for the moving object attitude judgment framework.

[0068] The present invention further proposes a method for correcting the parameters of an indoor robot positioning model based on the dynamic change pattern of wireless signal strength, and the steps are as follows:

[0069] Obtain the received signal strength indication (RSSI) r_i at each reference node i in the indoor environment. This process includes sampling the signal strength at a place where several known position reference points are pre-arranged. These reference points form a map covering the entire space for subsequent calculation and comparison. Each reference node represents a different geographical location, and the measured received signal strength indication directly reflects the strength of the wireless signal at the location.

[0070] Then calculate the weighted average RSSI R_avg = Σ(r_i * w_i), where w_i represents the weight corresponding to the received signal strength, and ensure that the sum of all w_i is 1. The range of the w_i parameter should be adjusted within [0, 1] so that the weighted mean can effectively represent the relative signal level of the robot in the actual environment. The determination of the optimal w_i needs to be achieved according to specific environmental conditions and target accuracy requirements. The formula is used to generate a comprehensive measurement method to more accurately depict the wireless signal strength distribution state in the overall environment.

[0071] By comparing with the standard signal attenuation function f(d) = AB * log10(d), where both A and B represent constant terms under specific environmental conditions, which affect the rate of signal strength decline. d represents a measurement of the distance from the emission source. This formula is intended to simulate and quantify the signal attenuation values expected at different distances. By matching the calculated ideal attenuation value sequence and the actual RSSI data collected from the site, the previously set wis are dynamically corrected, making the finally calculated position information more accurate and reliable. This operation is a key link to ensure that the robot maintains a high navigation accuracy in a complex and changing environment.

[0072] Calculate the position error ΔR_i according to |R_i - f(d_i)| and use this error to adjust and correct the estimated distance d_i. This means that if there is an obvious deviation between the current prediction and the ideal result, d_i should be appropriately increased or decreased to make the newly estimated distance value more in line with the actual situation, thus continuously approaching the correct position parameters. For example, in a specific embodiment, assume that the theoretical RSSI generated when initially considering the robot to be at a specified position in a certain room is x; while in reality, another slightly different specific reading y is observed, and the difference between the two constitutes the above-mentioned ΔR_i. With the help of it, the original hypothesis can be verified and necessary modifications can be made to improve the positioning accuracy.

[0073] Through such a process, the position estimation performance of the mobile device can be optimized. Especially in a complex and changeable environment, this method helps to continuously improve the navigation reliability of the robot.

[0074] Next, further describe the present invention before selecting the weights.

[0075] In the first step of executing this process, it is to detect the wall materials and thickness T_j, and associate these detection results with the empirical database to determine the corresponding correction coefficient K_j. This step is to identify the structural characteristics in the environment and use the pre-accumulated data to adjust the basic parameters for subsequent calculations. For example, when the sensor identifies a wall with a thickness of 30 cm and a material of concrete, the corresponding K_j found in the empirical database may be 0.75. This means that the influence brought by such a structure will be corrected with an influence ratio of 75%.

[0076] The second action is to use the KNN (K-Nearest Neighbors) algorithm to perform cluster classification on the current position, obtaining the n nearest neighbor samples {S_k} to this point along with the environmental label E_k of the surrounding space at each position. This helps extract the knowledge that best matches the current actual situation from existing information sources for analyzing and predicting future trends or pattern changes. For example, in one embodiment, assuming the robot is now near a corner inside a building, the system will find the top 10 instances under approximate conditions based on the distribution characteristics of other surrounding position points such as corridors, and mark the specific room type or special area nature to which they belong.

[0077] The third step is to establish a method called the mapping rule M(E, d) to determine how to reasonably arrange the allocation ratio of the weights w_i under specific preconditions. It is actually a dynamic response function, aiming to give a suitable selection strategy by comprehensively considering both aspects of E and d. For example, for the case of a corridor that is open and may have corner obstructions, the mapping formula can tailor a flexible and efficient weighting mechanism according to the actually detected distance difference.

[0078] The last step involves, for the data nodes in the range of the wall shadow, using a specific equation form M(E_k, d_i) = C / D_i 2 + C to adjust the originally default-set weight ratio numerical relationship problem. Among them, the symbol C represents the compensation base; and D_i represents the true linear geometric distance between an arbitrarily selected point and the wall. The purpose of this formulaic processing method is to ensure that the factors more affected by interference closer to the wall barrier can have a more prominent emphasis opportunity. The original intention of setting the formula like this is to consider that the surface adjacent to the obstacle often has an obvious blocking effect on the wireless signal. By introducing the square term, the influence of the near-distance effect can be effectively amplified while ensuring the natural continuity of the overall distribution, thereby improving the accuracy of the robot's position measurement.

[0079] Next, the non-line-of-sight (NLOS) propagation correction process of the present invention is described. Mobile robots often encounter obstacles when working in complex environments, resulting in measured position offsets. In response to this situation, the Kalman filter algorithm is applied to correct the data errors caused by NLOS. This process can filter out abnormal readings in real time and estimate the current situation based on the previous moment's state, making the position determination more accurate and reliable.

[0080] Next, in the optimization iteration, the prediction deviation vectors V = {Δp_t} in the previous N iterations are tracked and recorded. The element Δp_t represents the gap between the predicted value and the actual observation in each iteration, and the historical information stored in this set becomes the basis for the system's adaptive adjustment. For example, when a robot explores an unknown environment and continuously learns its geographical features, the previously accumulated experience can assist in improving the accuracy of subsequent trajectory planning.

[0081] Then, according to the update rule \[w\leftarrow w+\eta\nabla_w L(w)\], the parameter w is gradually changed until the optimal setting value is found. Here, the learning rate η determines the step size of each adjustment, generally belonging to the interval (0, 1), and the optimal value can be flexibly determined according to the specific application scenario; It represents the rate of change of the weight with respect to the total loss, that is, the gradient direction. Setting the formula in this way can be regarded as a process of finding a parameter combination path that minimizes the sum of squared errors, so as to be closest to the theoretically correct solution. Specifically, in a robot path-finding training process, with the reasonable selection of η and continuous repeated fine-tuning of the strength of each node connection, the ability to construct a stable and reliable walking route from the starting point to the target is ultimately improved.

[0082] Next, a more specific definition of the present invention is described. A threshold Th is defined to distinguish between valid measurement values e and invalid values, i.e., e < RSSI_ave - Th. The setting of the threshold Th is to filter out low-quality data to reduce the negative impact on subsequent position determination. Among them, RSSI represents Received Signal Strength Indicator, which is used to measure the signal strength. The average value of RSSI, RSSI_ave, is the average result of multiple RSSI values, and the RSSI measurement value usually ranges from -100 to 0 dBm. The value range of Th can be different from 10 to 20 dBm, and the default optimal value is 15 dBm. When the measured RSSI is less than (RSSI_ave - Th), these data points are regarded as invalid, otherwise they are regarded as valid measurement values.

[0083] In one embodiment, assume that three nodes record RSSI data points in a certain period, and after calculation, RSSI_ave = -60 dBm is obtained. If Th = 15 dBm, only the data with RSSI ≥ (-75 dBm) is retained as high-quality data. For example, the RSSI of node A is -58 dBm, B = -73 dBm, C = -79 dBm. At this time, only the measurements of nodes A and B are considered to be of high quality.

[0084] Then, the high-quality RSSI value set \(\{(x, y)_k\}\) selected according to the above threshold constitutes a high-quality data point set \(Q\), and a sub-network \(SN\) is constructed from these data. \(x\) and \(y\) represent the spatial coordinates of the robot, and the subscript \(k\) indicates different measurement samples. The generation of the sub-network is crucial for improving the system stability and positioning accuracy.

[0085] Divide \(Q\) into two parts: the master node \(MN\) and the slave node \(SN\), and then set up a synchronization timestamp \(ST\) so that all operations are carried out according to the unified time mark, ensuring the clock synchronization and communication coordination between nodes and reducing errors. The synchronization mechanism improves the success rate and timeliness of communication between nodes. In one instance, the set time is at the second level, and the accuracy reaches the sub-second level to meet the demanding application requirements. This arrangement also simplifies the problem-solving strategies in the implementation process of many timing control algorithms. For example, the mutual verification between \(MN_m\) and \(SN_n\) or the instruction transmission task depends on the accurate time basis support to be successfully completed.

[0086] The evaluation of the connection situation for any pair of \(MN_m\) and \(SN_n\) is based on a formula \(|rssi_{m}-rssi_{ave}(SN)| \geq \tau * \sqrt{Var(RSSI_SN)}\) to re-adjust the relative position to ensure accuracy, that is, if the gap between the actually observed node \(m\) and the predicted mean is too large and exceeds \(\tau *\) variance \(Var(RSSI_SN)\), corrective work needs to be done on the distance or azimuth angle here. Among them, the variance is used to reflect the discrete situation and stability of RSSI; the parameter \(\tau\) can be flexibly changed according to the application scenario and is selected within about 2 - 3, and it is recommended to default to 2.5. This condition helps to eliminate the estimation error phenomenon caused by the multipath effect due to the non-direct propagation path.

[0087] Next, in a position determination method of the present invention, in the case of encountering a large-area metal occlusion resulting in a severe shadow effect, an alternative beacon \(Beacon\) will be enabled to ensure the positioning accuracy. This step means that when it is detected that there is a large-area metal substance in the environment, which may cause significant attenuation or distortion of the conventional positioning signal, immediately switch to a set of pre-arranged and calibrated alternative beacons for work. For example, in a warehouse environment, when a robot passes through an area where a large number of metal products are stacked, it may encounter the problem of signal attenuation.

[0088] Beacon then sends out a special pulse sequence S in a predetermined manner for identification. The characteristics of these pulses are that their duration and phase are specially designed so that they can be easily identified without interfering with other ordinary communication activities. This particularity can effectively improve the stability and accuracy of data transmission while maintaining the compatibility of the entire system. Specifically, a pulse can be as short as tens of nanoseconds, and the frequency is carefully selected to meet the goal of easy detection without affecting the operation of the existing system; the phase selection in this step should try to avoid other channels or cycles that may be used to prevent conflicts or confusion.

[0089] Further collect the information P received from each Beacon and the calculated position offsets δx and δy. This means decoding the information of each received pulse to determine its identity and transmission time and other related data, and comparing and analyzing these data with the expected ideal reference position to evaluate the deviation of the current coordinates from the standard value. For example, if centimeters are used as the unit of distance measurement, the range of δx and δy in a specific application scenario should vary from negative one meter to positive one meter.

[0090] Finally, only the position correction value obtained under the condition of (|δx|+|δy|)<ε is accepted as the effective correction value to eliminate the influence of uncertainty factors caused by abnormal fluctuations. In this expression, ε represents a pre-set and relatively small value used to determine what is a slight misalignment deviation within the tolerable range. The default recommendation is to take 0.1 to 0.2 meters as a reasonable threshold range, because this level of accuracy is sufficient to ensure that the task execution needs of most mobile robots are not significantly affected. The purpose of this setting is to avoid misleading the system's understanding of self-positioning due to non-realistic large drift results caused by instantaneous errors or accidental events, thereby ensuring long-term reliability and efficient operation performance.

[0091] Next, the method further includes the following steps:

[0092] The first step is to select a set of centers from a large number of sensor readings and assign initial memberships. In this process, C cluster centers μ_c are randomly selected from the large batch of data obtained by the sensor, and appropriate memberships μ_ij (i belongs to 1 to C, j represents each sample) are assigned between each sensor reading and each cluster center. The membership indicates the degree of probability of each reading for a certain class, ranging from 0 to 1, and the sum of the memberships of all classes for each reading is equal to 1.

[0093] In one embodiment, it is assumed that there are a total of N = 1000 sensor data. The number of clusters C is selected to be 3, which means that three cluster center positions μ_c need to be randomly determined, and the membership values of each sample belonging to these three classes (such as [0.7, 0.2, 0.1]) are set, indicating that this reading is very likely to be a member of the first cluster but there is also a slight possibility of being an element of the other two classes.

[0094] The second step is to calculate the formula \(J=\sum_{i = 1}^{C}\sum_{j = 1}^{N}{\mu_{ij}}^β|x_i - v_j|^2\). This formula is intended to calculate the objective function. Among them, \(μ_ij\) is the membership degree of the data point j belonging to the category i determined above. \(v_j\) represents the position information of the sample point p_j, which can be composed of information from a laser rangefinder or a wheel speedometer in this scenario; \(x_i\) is the position vector of each cluster center. The power parameter β is usually set to be greater than or equal to 1 and not greater than 2, and the optimal value is 2. The overall meaning of this formula is to quantify the error size of the entire system to evaluate the performance and accuracy of the model.

[0095] In the third step, if the change value of the sum of errors J calculated and the result of the previous round is less than the predetermined decimal σ, it indicates that the algorithm has reached a stable stage. The threshold σ here is a small value set to avoid overfitting, such as 0.0001.

[0096] The last step is to take measures to readjust the membership degree for the situation where J fails to drop to the required accuracy level and repeat the above calculation process until the conditions are met. For example, specifically, after a round of calculation, if J still fluctuates greatly and does not tend to be stable, at this time, the membership degree μ_ij needs to be corrected backward according to the latest obtained error, and then the iterative calculation is performed again to ensure more accurate and reliable positioning. In this way, by continuously optimizing μ_ij to improve the value of J to approximate the true form of the actual physical world coordinate distribution.

[0097] Through these operations, a relatively reliable mobile robot position prediction framework can be finally obtained, and the accuracy of the measurement can be effectively improved by using this iterative mechanism based on fuzzy C-means.

[0098] Next, the following mechanism added in the present invention is described. First, the sliding window strategy is used to manage the latest m sets of historical positioning data Z. This means that the system will keep the records of the latest m positioning data. m is usually a small positive integer that can reflect trend changes (such as 5 or 10). Each time a new reading is received, the earliest data will be eliminated and the latest data will be added to ensure that the data set always remains up-to-date. This can make the position determination method have better real-time performance and dynamic response ability. For example, in an embodiment, it is assumed that the machine sends a new positioning information once per second in the environment, then the system saves the latest five positioning information for processing and analysis.

[0099] Next, the smoothed cumulative error estimate h(x) = γ * z + (1 - γ) * z_previous is set to evaluate the cumulative error. In the formula, z is the newly received reading currently, and z_previous represents the average or estimated value processed previously; the parameter γ represents the exponential weight (also called the decay factor), and the range of this parameter is between 0 and 1. Usually, it is set to be close to about 0.1 as the best, aiming to give higher importance to the recent historical positioning data and gradually reduce the influence of the long-term positioning data. The exponentially weighted moving average (EMA) is a technique for smoothing by assigning different degrees of importance to different observed values, and it plays a role in smoothing during the process of updating the position of the mobile robot.

[0100] Then, once the system receives new positioning data, it immediately performs EMA on the historical positioning data. This action helps to suppress the influence of long-term noise and the large jumps in positions that are not desired. Specifically, if a certain robot suddenly has an abnormally large position deviation due to obstacle reflection or other unreliable factors, then this process can reduce the impact of this short-term error on the entire system.

[0101] Finally, when it is observed that there are n consecutive increases in errors, an attitude adjustment instruction is immediately triggered to re-determine the initial pose, where n should be selected as an appropriate number based on specific circumstances, and this is used to avoid the further expansion of cumulative distortion. For example, in a specific application scenario, if it is detected that there are obvious increases in three consecutive periods, the correction measures are started to quickly resume normal operation and ensure that precise positioning continues to be reliable.

[0102] Through such a series of steps, potential risk problems are effectively controlled and the overall efficiency and accuracy level are improved by effectively managing the positioning data and making immediate feedback adjustments.

[0103] Next, a further limitation of the present invention will be described, specifically including an improved method for determining the position of a mobile robot in a complex terrain environment through the following steps. First, a pre-trained deep convolutional network (DCN) is used to extract the visual feature F in the environment; next, the extracted feature F is mapped to a low-dimensional manifold U using the adaptive linear discriminant analysis (ALDA); then, the Euclidean distance \(d_{AB}=\sqrt{\sum_j(u_B[j]-u_A[j])^2}\) between any two points on the low-dimensional manifold U is evaluated, and finally, if the Euclidean distance \(d_{curr,prev}<\Psi\) (predetermined threshold) between the current state and the previous state, then obviously impossible state transitions are excluded to simplify the computational load.

[0104] In this process, the purpose of using the pre-trained deep convolutional network DCN is to automatically obtain high-dimensional but discriminative representations, namely visual feature F, from the image data captured by the mobile robot. For example, in a mobile robot navigation task in a forest scene, the visual features may include information such as the distance between tree trunks and the unevenness of the ground. The deep convolutional network can be composed of a series of neural layers for processing different levels of details. The input can be a photo in RGB color format or data generated by a thermal imager, and its optimal configuration is customized and optimized according to the actual application situation. These extracted F will serve as the basis for the next step.

[0105] Next, the extracted visual features are mapped to a smaller-scale space that is easier to operate, process, and compare, called the low-dimensional manifold U. Here, the adaptive linear discriminant analysis ALDA is used, which helps to improve the classification accuracy and ensure that different types of features can be clearly separated in the low-dimensional manifold. In one embodiment, if the working range of the robot is for patrol and monitoring in an indoor environment, then this method can help better identify the specific differences at each position by distinguishing small differences such as changes in the door frame contour. The vector-form U obtained after this step of processing is more suitable for measuring the position change relationship, where j refers to the index subscript of the coordinate component on each low-dimensional manifold point, representing the number of features after mapping, and this value is not fixed for different robot application scenarios.

[0106] Subsequently, the relative displacement between two points on U is measured, which is the Euclidean distance, \(d_{AB}=\sqrt{\sum_j(u_B[j]-u_A[j])^2}\). This step is mainly used to quantitatively describe whether there are obvious changes in the same position feature points observed at different times or perspectives. The parameters AB refer to two observation points, and the subscripts B and A represent two different measurement instances, such as the position relationship of the image feature points obtained when the robot takes pictures at the same place twice successively; while Psi refers to the maximum error threshold within the set expectation, and the default setting should be based on empirical values and specific application scenarios. Generally, it is a small positive value to ensure accurate determination. When the value of d_curr - prev is lower than Psi, it indicates that the state similarity between the two moments is very high. For example, specifically, if it is desired to filter out the invalid path updates caused by accidental jitters in a small range during the continuous operation of the robot, then Ψ needs to be set at a moderately low level to reduce the subsequent unnecessary computational burden, thereby achieving an efficient and accurate position tracking ability.

[0107] The last part is to ensure that the mobile robot can accurately and effectively estimate its own position in a complex environment while minimizing the pressure on hardware computing requirements. Once it is found that the Euclidean distance difference between the two positioning calculations before and after is less than the preset threshold Psi, it is defaulted that the robot's travel trajectory has not changed or has changed by a negligible amount during this period. Thus, unnecessary further parsing processes are omitted to achieve the effects of saving energy resources and improving the reaction speed.

[0108] Next, a method for determining the position of a mobile robot according to the present invention is described. A crowd behavior pattern recognition mechanism is introduced in a complex scenario to improve the robot's positioning accuracy and path planning ability. By simulating the interaction of the crowd through the Social Force Model (SocialFM), the system can automatically adjust the strategy according to the change of the pedestrian density in the local area.

[0109] First, apply crowd behavior pattern recognition in a highly complex dynamic environment. This environment includes high - traffic places such as shopping malls and stations. Here, it refers to the intelligent analysis of the behavior patterns of the crowd, aiming to predict the action trends of each person to help the device pre - sense the upcoming crowd obstacles or crowded areas and ensure the safety and smoothness of navigation. For example, in a specific instance, if a large - scale concert is held in a city park, it can learn and identify the flow direction of the people on the scene in advance, which helps to formulate a more appropriate moving route.

[0110] Then, dynamic interaction rules between groups are established through SocialFM, and based on this, the parameter value of the local density D_loc of pedestrians in the area is modified. This process relies on an algorithm to evaluate the number of pedestrians and their density around a specific location within a unit time, and represents it in numerical form. For example, such a system is deployed in the waiting hall of a railway station. Usually, the afternoon period is the peak period, and the D_loc in the local area will be much higher than other times. The determination and adjustment of this value depend on the data obtained from real-time monitoring.

[0111] Furthermore, for individual members, that is, the forward direction vector vector_v and the speed vector vector_s of each pedestrian are restricted by the distribution density formed by the people around in the current environment, that is, the local density index mentioned above, so as to achieve personalized speed guidance and safety control. In a specific example, consider a person walking through a shopping mall. There is a group of friends walking together on the right side of this person. Then the space on the left side of this person is relatively large and it is faster and more comfortable to walk. At this time, according to the calculation results of the algorithm, it is recommended that this person choose the relatively empty path on the left to avoid walking into an overcrowded area.

[0112] The last step is to stipulate the conditions for triggering the additional monitoring program. When and only when the local density D_loc is greater than the preset threshold upper limit ρ (this maximum allowable level can be customized according to the actual situation), and at the same time the reference modulus of speed |vs_ref| does not exceed the set speed limit ν, then the special inspection step is further executed to reduce the unnecessary monitoring consumption problem in the sparse flow interval and ensure high computing performance. Assume that the value of ν is 0.8m / s and ρ = 4 people / m 2 , which neither affects the activities of the crowd nor guarantees the stability of the system; the reason for doing this is to optimize resource allocation and save costs such as electricity as much as possible without affecting the overall function.

[0113] A method for determining the position of a mobile robot according to the present invention includes: optimizing and improving the position determination of the robot through a series of steps, so as to improve its positioning accuracy and path planning ability in a dynamic and complex environment. This method specifically covers the in-depth fusion processing of wireless signals, multi-sensor information and historical trajectories, and proposes a series of solution strategies for different problems to ensure that the robot has higher reliability and flexibility when performing tasks.

[0114] To address the problems caused by the accumulation of positioning errors, we adopted a learning and adjustment strategy based on environmental perception data, aiming to reduce the adverse effects of errors on the long-term stability of the system. In this stage, the robot collects environmental perception data and performs local mapping, using the information obtained from high-precision sensors such as vision and lidar as input to construct an adaptive correction algorithm. This process periodically analyzes the differences between the current state and the expected model and feeds these differences back to the core control program to update the model prediction weights and correct the robot's similar future behaviors. In particular, this method uses a reinforcement learning framework to continuously iterate and find the optimal adjustment scheme to ensure a low level of position error during long-term operation. At the same time, it can automatically adapt to different scenario changes (such as indoor-outdoor conversion), further enhancing the system robustness.

[0115] Based on the coordinate correction mechanism of historical trajectory and real-time sensor fusion, to address the problem of positioning drift in complex terrains, this method adopts an advanced historical trajectory matching algorithm, combined with the signals of vision or ultrasonic sensors collected immediately for accurate coordinate verification. This process not only considers the specific coordinate position where the robot is currently located, but also pays more attention to the historical record of the movement path it has experienced in the entire environment. When encountering difficult-to-define boundary conditions (such as a sudden widening at the end of a corridor), it can find similar patterns by comparing its past movement trajectories to make a judgment, thereby obtaining the most accurate current position estimate. More importantly, this method effectively alleviates the risk of false detection that may be caused by a single data source; through multiple data verifications, the accuracy of the results can be greatly improved, especially in an environment with many obstacles and large interference, showing obvious performance advantages. This combination improves the adaptability and reliability of the robot under various unknown conditions, ensuring that it will not cause significant yaw due to short-term measurement errors.

[0116] To improve the speed and accuracy in avoiding dynamic obstacles through refined adjustment of map-based path variables, we also designed a map update mechanism that relies on a static existing map database plus newly detected geographical features continuously monitored, so as to guide the design and implementation of the next action plan. The known information here includes the previously drawn building floor plans or the relevant area structure data accumulated and saved during previous tasks; the newly added data refers to the continuously recorded changes in the surroundings during the navigation process, such as newly added items, temporarily stacked furniture, or occlusion points formed by other moving elements. The two work together. On the one hand, it ensures that the starting point of the plan has sufficient cognitive basis to avoid obstacles in fixed forms; on the other hand, it can quickly respond to emergencies and make corresponding adjustment suggestions. The entire system relies on the support of advanced artificial intelligence recognition technology. After detecting new unknown objects or significant changes in regional characteristics in the coming direction, it immediately takes countermeasures to ensure that the machine can freely shuttle in the complex and changeable actual operation space, always maintaining high operating efficiency without being limited by previous settings.

[0117] The correction value determination model eliminates the problem of signal attenuation caused by walls. In response to the challenge of signal distortion caused by wall blockage, a set of empirical rule sets dedicated to precise positioning in indoor environments is established as a reference template by carefully observing and simulating the reflection attenuation value distribution characteristics of walls composed of different materials. According to the change trend of the received signal strength indication (RSSI), the internal working parameter configuration of the positioning system is dynamically adjusted to reach the optimal matching state. Simply put, a lot of preparatory work is done before actual deployment to study clearly the specific effects of various types of barriers on the transmission quality of various wireless communication frequency bands such as WiFi / GPS, forming a methodology toolbox that can distinguish the clear boundary between open spaces and shielded objects for ready reference. Then, whenever the robot detects corresponding fluctuations during activities in the controlled area, a pre-trained neural network classifier is immediately used for real-time calculation, and a set of parameter values most suitable for the specific scenario conditions is selected and assigned to the position parsing subroutine for use. This method effectively compensates for the inevitable data deviation caused by physical barriers, making the final calculated result more in line with the truth and closer to the actual situation.

[0118] Optimization of multi-source information fusion for uncertainty estimation in a dynamic environment. Finally, to ensure a continuous and good grasp of one's own orientation relative to the reference frame in a highly dynamic background with high-frequency changes and unforeseeable consequences, it is necessary to correctly grasp the degree of uncertainty - that is, to construct a complete global perspective by pooling scattered and independent information sources provided by various sensing components, and on this basis, form a dynamically increasing knowledge structure to describe the potential relationships between things and the probability density distribution. The specific implementation method is to repeatedly evaluate the credibility levels of messages from each level using the technical idea of the Bayesian inference formula in modern mathematical statistics theory, quantify them to form a weighing factor, and apply it to each new measurement link to complete the closed-loop control loop; at the same time, it is also necessary to consider the time difference compensation effect caused by the time stamp difference to ensure that all elements participating in the interactive calculation are in a synchronous state. With such an efficient and stable update mechanism, high-quality self-positioning accuracy output can be maintained at any time and place, significantly reducing the loss range caused by being impacted by unexpected factors.

[0119] The core advantage of the above complete embodiment content lies in transforming the traditional one-way passive acceptance of external instructions into an intelligent autonomous perception and adjustment mechanism, greatly enhancing the overall performance of the mobile platform in maintaining a good operating state under non-calibrated conditions, and having great innovative and practical value.

[0120] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirits of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for determining the position of a mobile robot, characterized in that: include: Based on the wireless signal strength change pattern, the indoor robot position determination model parameters are corrected to compensate for the signal interference error caused by wall obstruction. The robot's current coordinates are corrected based on the corrected model parameters, fusing historical trajectories with real-time sensor information. The corrected current coordinates are combined with known map information and real-time updated map details to adjust the path planning variables, and the uncertainty estimation matrix is ​​updated through the results of multi-sensor fusion output to improve the robust positioning in high dynamic environments. The method of correcting the indoor robot position determination model parameters based on the wireless signal strength change pattern further includes: obtaining a received signal strength indication (RSSI) r_i at each reference node i in the indoor environment; Calculate the weighted average RSSI R_avg = Σ(r_i*w_i), where w_i is the weight, satisfying Σw_i = 1; adjust the weight w_i by comparing with the known standard signal attenuation function f(d), f(d) = AB*log10(d), A and B correspond to the function results of the constant and the transmitter distance factor d respectively; judge and correct the position parameter according to the formula error ΔR_i = |R_i f(d_i)|, where d_i is the estimated distance.

2. A method for determining the position of a mobile robot according to claim 1, characterized in that: Before selecting the weight, it further includes detecting each wall material and thickness T_j, and associating it with the empirical database to determine the corresponding correction coefficient K_j; Use the KNN algorithm to cluster the current point and obtain the nearest n samples {S_k} and their corresponding environment labels E_k; Construct a mapping rule M(E,d) to determine how to dynamically allocate weights w_i under given conditions; For nodes within the wall influence area, apply M(E_k,d_i)=C / D_i^2+C, where C is the compensation value and D_i represents the distance from point i to the nearest wall.

3. A method for determining the position of a mobile robot according to claim 2, characterized in that: Also includes: The Kalman filter algorithm is used to deal with the offset caused by non-line-of-sight propagation; Record the prediction deviation vector V = {Δp_t} in the first N iterations, which is used as prior information in the optimization process; Use gradient descent to minimize the loss L(w) = (P_obsP_est)^TΣ^{1}(P_obsP_est), where Σ represents the covariance matrix; According to the learning rate η update formula: w←w+η▽_wL(w), the weights gradually tend to reasonable values.

4. A method for determining the position of a mobile robot according to claim 2, characterized in that: include: Define a threshold Th to distinguish valid measurement values ​​e from invalid values ​​ie,e <RSSI_aveTh; Combined with the above threshold, a high-quality data point set Q = {(x, y)_k} is selected to form a sub-network SN; Divide Q into a master node MN and a slave node SN, and then set a synchronization timestamp ST; For any pair of MN_m and SN_n, when |rssi_{m}rssi_{ave}(SN)|≥τ√Var(RSSI_SN), the relative position between the two is recalculated to ensure accuracy.

5. The method for determining the position of a mobile robot according to claim 3, characterized in that: It also includes enabling alternative beacons when encountering large areas of metal occlusion causing severe shadow effects; Beacon emits a special pulse sequence S, whose duration and phase are designed to facilitate identification without affecting the original communication system; Collect the information P returned by each Beacon, as well as the actual position differences δx and δy; The correction is trusted only when the absolute difference satisfies the condition (|δx|+|δy|)<ε (a small value set in advance), thereby filtering out the impact of abnormal disturbances.

6. The method for determining the position of a mobile robot according to claim 1, characterized in that: A location accuracy improvement framework based on fuzzy C-means clustering (FCM) is established; A set of centers μ_c are randomly selected from a large number of sensor readings and assigned appropriate memberships μ_ij; Iteratively calculate the value of J = \sum\limits_{i=1}^{C}\sum\limits_{j=1}^{N}{\mu_{ij}}^{β}|v_ip_j|^2; If the change of J after a round is less than the set tolerance σ, it is considered converged, otherwise the membership degree is adjusted and the above process is repeated until the convergence condition is met.

7. A method and system for determining the position of a mobile robot according to claim 1, characterized in that: This includes introducing Doppler frequency shift d_f to assist in identifying the motion state, especially the speed jump that occurs at the beginning or end stage; Whenever v>=Vmax*α or d_f<=D_min*(1±δ), where α is the sensitivity coefficient, activate the supplementary calibration step to lock the true position; This operation can effectively eliminate the position estimation deviation caused by sudden acceleration and improve the overall positioning accuracy; The high-confidence data generated under the above conditions is used preferentially for feedback correction to enhance the real-time response capability of the system.

8. A method and system for determining the position of a mobile robot according to claim 1, characterized in that: This includes adding the following mechanisms: Use sliding window strategy to manage the latest m groups of historical positioning data Z; Let h(x) be the cumulative error estimate after smoothing, h(z_(tm),...,z_t)=γ*z+(1γ)*z_previous; Each time a new reading is received, the content of Z is updated and an exponential decay average (EMA) is performed to reduce forward volatility. Once a trend reversal is observed - n consecutive generations of growth, a reset of the starting attitude is immediately triggered to prevent the accumulated distortion from continuing to expand.

9. A method and system for determining the position of a mobile robot according to claim 1, characterized in that: In complex terrain environments, visual features F are extracted through the pre-trained deep convolutional network DCN; Use adaptive linear discriminant analysis ALDA to map to the low-dimensional manifold U; Evaluate the Euclidean distance between any two points on U d_{AB}=sqrt(Σ(u_B[j]u_A[j])^2); If d_{curr,prev}<Ψ(predetermined limit), obviously impossible state transitions are eliminated to simplify the computational load.

10. A mobile robot position determination system according to claims 1-9, characterized in that: This includes the introduction of crowd behavior pattern recognition for highly complex dynamic scenarios such as crowded areas; The social force model SocialFM is used to simulate the dynamic interaction rules of the group and adjust the local density D_loc accordingly; For individual k, its target direction vector_v and speed vector_s are subject to the distribution density of nearby pedestrians; The additional monitoring procedure is triggered only when D_loc>ρ (set upper limit) and |vs_ref|≤ν (stability indicator) is established, avoiding excessive attention to the quiet interval and wasting computing resources.

Images