A positioning and navigation method based on multi-task cooperative optimization

Abstract

The application discloses a positioning and navigation method based on multi-task cooperative optimization, comprising the following steps: S1, collecting and preprocessing multi-source heterogeneous data of a target carrier; S2, decomposing a positioning and navigation task, constructing a position, a speed and a scene modeling subtask, and generating a subtask feature set; S3, performing feature decoupling and information sharing, establishing cross-dependent relationships and gradient conduction paths, constructing a multi-task objective function and calculating a loss and a cooperative error; S4, performing dynamic weight distribution and parameter updating based on the loss and the cooperative error, and generating a multi-task cooperative parameter set; S5, applying the cooperative parameters to path planning, generating a navigation path candidate sequence, and screening a feasible path; S6, generating a navigation control instruction according to the feasible path and sending the navigation control instruction to an execution unit, and completing the positioning and navigation task. The application realizes multi-task cooperative positioning and navigation under the driving of multi-source heterogeneous information, and has high precision, high robustness and strong adaptability.

Description

A localization and navigation method based on multi-task cooperative optimization Technical Field

[0001] This invention relates to the field of intelligent navigation and positioning technology, and in particular to a positioning and navigation method based on multi-task collaborative optimization. Background Technology

[0002] In existing positioning and navigation technologies, mainstream methods are mostly built on a single-task architecture, meaning that models and processes are designed independently for functions such as position estimation, velocity estimation, or environmental mapping. Position estimation methods, represented by the Global Navigation Satellite System (GNSS), rely on external signals and are susceptible to obstruction, multipath interference, and signal drift, resulting in significantly increased positioning errors in urban canyons, tunnels, or indoor environments. To overcome these shortcomings, researchers have introduced inertial measurement units (IMUs) and visual sensors, using multi-source data fusion to improve the robustness of navigation systems.

[0003] However, existing fusion navigation methods typically treat data fusion as a front-end processing step, relying on a separate, single-task optimization architecture in the back-end, lacking a collaborative mechanism between tasks. Position estimation, velocity estimation, and scene modeling are often completed by independent sub-processes, without sufficient sharing of intermediate feature information. This results in overall navigation accuracy depending on the performance of a single module, and insufficient robustness of the system in dynamic environments. Furthermore, most existing methods employ static loss weighting strategies, failing to automatically adjust task priorities based on scene changes and task conflicts, thus limiting the model's adaptability in complex scenarios.

[0004] Some studies have attempted to introduce deep learning frameworks to jointly model multiple navigation tasks, but the following problems still exist in general: First, the design of multi-task objective functions is unreasonable, failing to explicitly model the cross-dependencies between tasks, leading to gradient update conflicts and performance degradation; second, there is a lack of dynamic weighting mechanisms, making the task optimization process susceptible to the influence of gradient-dominant tasks, and failing to achieve true collaborative optimization; third, path planning often adopts heuristic search, making it difficult to integrate model prediction information for candidate path generation and evaluation, resulting in poor temporal consistency and spatial rationality of overall path generation.

[0005] Therefore, how to provide a positioning and navigation method based on multi-task collaborative optimization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a positioning and navigation method based on multi-task collaborative optimization. This invention integrates multi-source heterogeneous data to construct a collaborative optimization structure for three types of tasks: position estimation, velocity estimation, and scene modeling. It adopts a task feature decoupling and dynamic weighting mechanism to achieve joint modeling of multi-task objective functions, and improves the temporal consistency and spatial rationality of navigation paths through a feasible path generation strategy. It has the advantages of strong adaptability, high robustness, and excellent path accuracy.

[0007] A positioning and navigation method based on multi-task cooperative optimization according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source heterogeneous data of the target carrier and perform preprocessing;

[0009] S2. Based on the preprocessed multi-source heterogeneous data, the positioning and navigation task is decomposed to construct a position estimation subtask, a velocity estimation subtask, and a scene modeling subtask, and corresponding subtask feature sets are generated.

[0010] S3. Perform feature decoupling and information sharing operations on the feature set of subtasks, extract the cross-dependencies between subtasks, set gradient propagation paths according to the cross-dependencies, construct multi-task objective functions, and jointly calculate the loss value and collaborative error value of each subtask.

[0011] S4. Based on the loss value and the collaborative error value, a dynamic weight allocation mechanism is used to perform collaborative optimization on each subtask, and gradient backpropagation and parameter update operations are performed to generate a multi-task collaborative parameter set.

[0012] S5. Apply the multi-task collaborative parameter set to the position estimation subtask, velocity estimation subtask and scene modeling subtask, perform path planning calculation, generate navigation path candidate sequence, and perform temporal consistency judgment and spatial constraint verification to screen feasible path sequences.

[0013] S6. Generate corresponding navigation control commands based on the feasible path sequence and send them to the execution unit of the target vehicle to complete the positioning and navigation task.

[0014] Optionally, the multi-source heterogeneous data includes GPS signals, inertial measurement unit data, and visual image sequences.

[0015] Optionally, the preprocessing includes time interpolation of GPS signals, denoising and resampling of inertial measurement unit data, and distortion correction and inter-frame registration of visual image sequences.

[0016] Optionally, S3 specifically includes:

[0017] S31. Divide the subtask feature set into a shared feature set. With exclusive feature set The position estimation subtask, velocity estimation subtask, and scene modeling subtask correspond to respectively Perform residual comparison operations to generate a cross-dependency matrix between subtasks. The formula for calculating the cross-dependency matrix is ​​as follows:

[0018] ;

[0019] in, Subtasks In the sample The unique feature value on, For the sample size, Subtasks sub-tasks The degree of difference in characteristics between them;

[0020] S32. Based on the cross-dependency matrix Construct gradient propagation graphs between subtasks The gradient propagation graph controls the backpropagation path by setting the weights of dependent edges, and the path weight matrix... Defined as:

[0021] ;

[0022] in, As dependent on the attenuation coefficient, Subtasks Pair Task The intensity of the reverse influence, This represents the natural exponential function with base e;

[0023] S33, Based on shared feature sets , exclusive feature set and path weight matrix Construct a multi-task objective function The calculation formula is:

[0024] ;

[0025] in, Subtasks The base loss value, Weighting coefficients for subtasks This is the cooperative error adjustment coefficient. Subtasks sub-tasks The cooperative error term.

[0026] Optionally, S32 specifically includes:

[0027] S321. Based on the cross-dependency matrix The feature difference values ​​between each group of subtasks are used to construct a task node set, which corresponds to the position estimation subtask, the velocity estimation subtask, and the scene modeling subtask.

[0028] S322. Traverse any pairwise combination relationship in the task node set, set the connection direction of the edge according to the feature difference value in the cross dependency matrix, and map the feature difference value to the connection strength value.

[0029] S323. Set a threshold range for connection strength. Connections exceeding the upper limit are defined as weakly correlated. Weakly correlated edges are defined in the gradient propagation graph. In the gradient propagation graph, edges with low weights are used to represent connections below the lower bound. Connections below the lower bound are defined as strongly correlated. Strongly correlated edges are represented in the gradient propagation graph. The middle edge is represented by a high-weight edge;

[0030] S324. Combine the task node set with the weighted connection relationship to form a directed acyclic graph structure, where the direction and weight of each edge are determined by the feature difference values ​​between subtasks;

[0031] S325. Based on the directed acyclic graph structure, determine the backpropagation path of the gradient of each subtask. The propagation weight of each edge in the backpropagation path is used to control the degree of participation of error information in the transmission between tasks, and finally form a path weight matrix, which is used for the construction of multi-task objective function and collaborative optimization calculation.

[0032] Optionally, S4 specifically includes:

[0033] S41. Based on the multi-task objective function Calculate the base loss value for each subtask. and cooperative error term The total error index for each subtask is normalized, and the standardized loss is calculated as follows:

[0034] ;

[0035] in, These are the weighting coefficients for the cooperative error. Subtasks The standardized loss value, Subtasks sub-tasks The cooperative error term;

[0036] S42, Based on standardized loss value and historical gradient statistics Construct a dynamic weight allocation function to calculate the subtask weights for the current round. The weight calculation formula is as follows:

[0037] ;

[0038] in, Historical gradient suppression factor, Subtasks The gradient magnitude in the middle and upper iterations is used to suppress local oscillations and the risk of overfitting;

[0039] S43. Weighting subtasks Substitute into the multi-task objective function In the middle, construct the total loss function and to Perform a minimization solution, and the calculation formula is as follows:

[0040] ;

[0041] S44. Based on the total loss function Perform gradient backpropagation to calculate the partial derivative gradients of all parameters. The parameters are updated using a gradient accumulation strategy, and the parameter update formula is as follows:

[0042] ;

[0043] in, For learning rate, These are the gradient coefficients. These represent the sets of multi-task collaboration parameters for the previous round and the current round, respectively. These represent the partial derivative gradients of all parameters in the previous and current rounds, respectively;

[0044] S45. Update the multi-task collaboration parameter set. Write it to the task scheduling queue for use by the position estimation, velocity estimation and scene modeling subtasks, and for path planning and navigation instruction generation operations.

[0045] Optionally, the partial derivative gradient The calculation process is as follows:

[0046] According to the chain rule, the total loss function Regarding arbitrary parameters Taking the partial derivative, we construct the following expression:

[0047] ;

[0048] in, Subtasks The predicted output, express Regarding parameters The derivative of Represents the cooperative error term between subtasks Regarding parameters The derivative;

[0049] The backpropagation algorithm is used to backtrack layer by layer starting from the output layer, calculating the intermediate variables of each layer with respect to the parameters in turn. The partial derivative values ​​are calculated, and the gradient results of the partial derivatives corresponding to all parameters are stored in the gradient vector. It is used for parameter update calculations.

[0050] Optionally, S5 specifically includes:

[0051] S51, Set up multi-task collaboration parameters These methods are applied to the position estimation subtask, velocity estimation subtask, and scene modeling subtask, respectively, and forward computation is performed on the feature sets of each subtask to obtain the current position prediction vector. Current velocity prediction vector With the current scene structure vector ;

[0052] S52, Predicting vectors based on current location Velocity prediction vector With scene structure vector A state extension strategy is used to construct a location state set. And use the cost function to evaluate the state at each location. Score and sort:

[0053] ;

[0054] in, Indicates position status Path cost, The obstacle conflict factor represents the positional state and scene structure. These are weighting coefficients;

[0055] S53. Sort by path cost from the set of position states. Before being selected Position status as a candidate sequence for navigation path and to Execution timing consistency check, detection Do the position states at adjacent time points satisfy the acceleration continuity condition?

[0056] ;

[0057] It also performs spatial constraint verification to check whether the positional state in three-dimensional space meets the scene navigability conditions. ,in For continuity and accessibility thresholds, and ;

[0058] S54. The position status, determined by timing consistency and spatial constraint verification, is used to form a feasible path sequence for navigation control command generation and invocation.

[0059] Optionally, the forward computation performs a layer-by-layer mapping operation on the feature set of subtasks based on the multi-task collaborative parameter set. In each layer, intermediate representation vectors are calculated sequentially according to the pre-defined linear transformation and nonlinear activation rules, and the current position prediction vectors of the position estimation subtasks are generated in the output layer respectively. The current velocity prediction vector of the velocity estimation subtask The current scene structure vector of the scene modeling subtask .

[0060] Optionally, the state extension strategy specifically includes:

[0061] Using the current position prediction vector as the initial state, construct a state sequence according to the set time step to generate multiple candidate points for prediction states;

[0062] Within each time step, a state transition window is constructed by combining the current velocity prediction vector with the acceleration range constraint, and possible velocity change directions are enumerated and the corresponding displacement is calculated.

[0063] Based on the predicted scene structure vector, an environmental feasibility test is performed on each predicted state candidate point, and state points that fall into obstacle areas or impassable areas are eliminated.

[0064] Each retained state candidate point is uniquely identified and encoded, and its cumulative path cost with the current position is recorded for path cost function calculation and sorting.

[0065] All candidate states that meet the constraints are organized into a set of position states in chronological order, which is used to construct a candidate sequence of navigation paths.

[0066] The beneficial effects of this invention are:

[0067] First, by constructing sub-tasks for position estimation, velocity estimation, and scene modeling, and introducing feature decoupling and information sharing mechanisms, cross-dependency modeling and parameter collaborative optimization among the tasks are achieved, avoiding information redundancy and accuracy bottlenecks caused by isolated task modeling in traditional methods.

[0068] Secondly, a dynamic weight allocation mechanism is adopted to dynamically adjust the optimization direction based on the loss value and collaborative error of each sub-task, enabling the navigation system to automatically allocate computing resources according to changes in the actual environment, thereby improving its adaptability and stability in complex dynamic scenarios.

[0069] Finally, multiple path candidate sequences are generated through forward computation, and path selection is performed by combining temporal consistency judgment and spatial constraint verification. This improves the continuity and drivability of path generation, ensures safety and accuracy during navigation, and achieves higher accuracy and robustness in multi-source fusion positioning tasks. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 is a flowchart of a positioning and navigation method based on multi-task cooperative optimization proposed in this invention;

[0072] Figure 2 is a flowchart of the multi-task objective function and gradient co-optimization of a positioning and navigation method based on multi-task cooperative optimization proposed in this invention;

[0073] Figure 3 is a flowchart of the navigation path candidate generation and screening process of a positioning and navigation method based on multi-task collaborative optimization proposed in this invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0075] Referring to Figures 1-3, a localization and navigation method based on multi-task cooperative optimization includes the following steps:

[0076] S1. Collect multi-source heterogeneous data of the target carrier and perform preprocessing;

[0077] S2. Based on the preprocessed multi-source heterogeneous data, the positioning and navigation task is decomposed to construct a position estimation subtask, a velocity estimation subtask, and a scene modeling subtask, and corresponding subtask feature sets are generated.

[0078] S3. Perform feature decoupling and information sharing operations on the feature set of subtasks, extract the cross-dependencies between subtasks, set gradient propagation paths according to the cross-dependencies, construct multi-task objective functions, and jointly calculate the loss value and collaborative error value of each subtask.

[0079] S4. Based on the loss value and the collaborative error value, a dynamic weight allocation mechanism is used to perform collaborative optimization on each subtask, and gradient backpropagation and parameter update operations are performed to generate a multi-task collaborative parameter set.

[0080] S5. Apply the multi-task collaborative parameter set to the position estimation subtask, velocity estimation subtask and scene modeling subtask, perform path planning calculation, generate navigation path candidate sequence, and perform temporal consistency judgment and spatial constraint verification to screen feasible path sequences.

[0081] S6. Generate corresponding navigation control commands based on the feasible path sequence and send them to the execution unit of the target vehicle to complete the positioning and navigation task.

[0082] This invention constructs a multi-task collaborative optimization process driven by multi-source heterogeneous data, realizing the linkage processing of position estimation, velocity estimation and scene modeling tasks, effectively improving navigation accuracy and stability in complex environments.

[0083] In this embodiment, the multi-source heterogeneous data includes GPS signals, inertial measurement unit data, and visual image sequences. The GPS signals include timestamps, longitude, latitude, and altitude information. The inertial measurement unit data includes the raw triaxial measurements of the accelerometer and gyroscope. The visual image sequences are continuous image frames acquired at a fixed frame rate.

[0084] This invention achieves complementary fusion of multi-source data by introducing three types of heterogeneous information: GPS signals, inertial measurement unit data, and visual image sequences, thereby enhancing the navigation system's adaptability to environmental changes.

[0085] In this embodiment, the preprocessing includes: performing linear interpolation and drift removal on the GPS signal according to a unified time reference to construct a position data sequence with equal time intervals; performing bandpass filtering, resampling, and triaxial alignment on the inertial measurement unit data, and mapping the acceleration and angular velocity data to the global coordinate system; performing camera intrinsic parameter correction and radial distortion correction on the visual image sequence, using the pyramid Lucas-Kanade optical flow algorithm to complete inter-frame motion estimation, and constructing initial values ​​of relative pose between image frames based on feature point matching.

[0086] This invention improves the consistency and effectiveness of data input by performing differentiated preprocessing on various types of data, including temporal interpolation, denoising, resampling, and image registration, thus providing a stable feature foundation for multi-task modeling.

[0087] In this embodiment, S3 specifically includes:

[0088] S31. Divide the subtask feature set into a shared feature set. With exclusive feature set The position estimation subtask, velocity estimation subtask, and scene modeling subtask correspond to respectively Perform residual comparison operations to generate a cross-dependency matrix between subtasks. The formula for calculating the cross-dependency matrix is ​​as follows:

[0089] ;

[0090] in, Subtasks In the sample The unique feature value on, For the sample size, Subtasks sub-tasks The degree of difference in characteristics between them;

[0091] S32. Based on the cross-dependency matrix Construct gradient propagation graphs between subtasks The gradient propagation graph controls the backpropagation path by setting the weights of dependent edges, and the path weight matrix... Defined as:

[0092] ;

[0093] in, As dependent on the attenuation coefficient, Subtasks Pair Task The intensity of the reverse influence, This represents the natural exponential function with base e;

[0094] S33, Based on shared feature sets , exclusive feature set and path weight matrix Construct a multi-task objective function The calculation formula is:

[0095] ;

[0096] in, Subtasks The base loss value, Weighting coefficients for subtasks This is the cooperative error adjustment coefficient. Subtasks sub-tasks The cooperative error term.

[0097] This invention achieves information collaboration and error sharing among tasks by dividing task features into shared and exclusive sets and constructing cross-dependency matrices and path weight matrices, thereby improving the overall network's performance in multi-task scenarios.

[0098] In this embodiment, S32 specifically includes:

[0099] S321. Based on the cross-dependency matrix The feature difference values ​​between each group of subtasks are used to construct a task node set, which corresponds to the position estimation subtask, the velocity estimation subtask, and the scene modeling subtask.

[0100] S322. Traverse any pairwise combination relationship in the task node set, set the connection direction of the edge according to the feature difference value in the cross dependency matrix, and map the feature difference value to the connection strength value.

[0101] S323. Set a threshold range for connection strength. Connections exceeding the upper limit are defined as weakly correlated. Weakly correlated edges are defined in the gradient propagation graph. In the gradient propagation graph, edges with low weights are used to represent connections below the lower bound. Connections below the lower bound are defined as strongly correlated. Strongly correlated edges are represented in the gradient propagation graph. The middle edge is represented by a high-weight edge;

[0102] S324. Combine the task node set with the weighted connection relationship to form a directed acyclic graph structure, where the direction and weight of each edge are determined by the feature difference values ​​between subtasks;

[0103] S325. Based on the directed acyclic graph structure, determine the backpropagation path of the gradient of each subtask. The propagation weight of each edge in the backpropagation path is used to control the degree of participation of error information in the transmission between tasks, and finally form a path weight matrix, which is used for the construction of multi-task objective function and collaborative optimization calculation.

[0104] This invention constructs a directed acyclic gradient propagation graph by quantizing the task relationships in the cross-dependency matrix, which precisely controls the propagation path of error information between tasks and effectively alleviates the gradient conflict problem.

[0105] In this embodiment, S4 specifically includes:

[0106] S41. Based on the multi-task objective function Calculate the base loss value for each subtask. and cooperative error term The total error index for each subtask is normalized, and the standardized loss is calculated as follows:

[0107] ;

[0108] in, These are the weighting coefficients for the cooperative error. Subtasks The standardized loss value, Subtasks sub-tasks The cooperative error term;

[0109] S42, Based on standardized loss value and historical gradient statistics Construct a dynamic weight allocation function to calculate the subtask weights for the current round. The weight calculation formula is as follows:

[0110] ;

[0111] in, Historical gradient suppression factor, Subtasks The gradient magnitude in the middle and upper iterations is used to suppress local oscillations and the risk of overfitting;

[0112] S43. Weighting subtasks Substitute into the multi-task objective function In the middle, construct the total loss function and to Perform a minimization solution, and the calculation formula is as follows:

[0113] ;

[0114] S44. Based on the total loss function Perform gradient backpropagation to calculate the partial derivative gradients of all parameters. The parameters are updated using a gradient accumulation strategy, and the parameter update formula is as follows:

[0115] ;

[0116] in, For learning rate, These are the gradient coefficients. These represent the sets of multi-task collaboration parameters for the previous round and the current round, respectively. These represent the partial derivative gradients of all parameters in the previous and current rounds, respectively;

[0117] S45. Update the multi-task collaboration parameter set. Write it to the task scheduling queue for use by the position estimation, velocity estimation and scene modeling subtasks, and for path planning and navigation instruction generation operations.

[0118] This invention constructs a dynamic weighting scheme for subtasks based on dynamic loss standardization and gradient magnitude adjustment mechanism, and combines joint gradient direction execution parameter updates to achieve adaptive scheduling in the multi-task collaborative optimization process.

[0119] In this embodiment, the partial derivative gradient The calculation process is as follows:

[0120] According to the chain rule, the total loss function Regarding arbitrary parameters Taking the partial derivative, we construct the following expression:

[0121] ;

[0122] in, Subtasks The predicted output, express Regarding parameters The derivative of Represents the cooperative error term between subtasks Regarding parameters The derivative;

[0123] The backpropagation algorithm is used to backtrack layer by layer starting from the output layer, calculating the intermediate variables of each layer with respect to the parameters in turn. The partial derivative values ​​are calculated, and the gradient results of the partial derivatives corresponding to all parameters are stored in the gradient vector. It is used for parameter update calculations.

[0124] This invention uses a chain-like rule and a structured gradient flow mechanism, combined with subtask weights and error influence factors, to accurately calculate the partial derivative gradients of each parameter, ensuring the accuracy and convergence stability of the parameter update process.

[0125] In this embodiment, S5 specifically includes:

[0126] S51, Set up multi-task collaboration parameters These methods are applied to the position estimation subtask, velocity estimation subtask, and scene modeling subtask, respectively, and forward computation is performed on the feature sets of each subtask to obtain the current position prediction vector. Current velocity prediction vector With the current scene structure vector ;

[0127] S52, Predicting vectors based on current location Velocity prediction vector With scene structure vector A state extension strategy is used to construct a location state set. And use the cost function to evaluate the state at each location. Score and sort:

[0128] ;

[0129] in, Indicates position status Path cost, The obstacle conflict factor represents the positional state and scene structure. These are weighting coefficients;

[0130] S53. Sort by path cost from the set of position states. Before being selected Position status as a candidate sequence for navigation path and to Execution timing consistency check, detection Do the position states at adjacent time points satisfy the acceleration continuity condition?

[0131] ;

[0132] It also performs spatial constraint verification to check whether the positional state in three-dimensional space meets the scene navigability conditions. ,in For continuity and accessibility thresholds, and ;

[0133] S54. The position status, determined by timing consistency and spatial constraint verification, is used to form a feasible path sequence for navigation control command generation and invocation.

[0134] This invention applies collaborative optimization parameters to the navigation path generation process, combining position, speed and scene information to construct a cost function and path candidate sequence, effectively improving the accuracy of path planning and task executability.

[0135] In this embodiment, the forward computation performs a layer-by-layer mapping operation on the feature set of subtasks based on the multi-task collaborative parameter set. In each layer, intermediate representation vectors are calculated sequentially according to the pre-defined linear transformation and nonlinear activation rules, and the current position prediction vectors of the position estimation subtasks are generated in the output layer. The current velocity prediction vector of the velocity estimation subtask The current scene structure vector of the scene modeling subtask .

[0136] This invention achieves efficient forward mapping from subtask features to navigation predictions through a multi-layered linear and nonlinear computational structure, thereby enhancing the reliability of path reasoning.

[0137] In this embodiment, the state extension strategy specifically includes:

[0138] Using the current position prediction vector as the initial state, construct a state sequence according to the set time step to generate multiple candidate points for prediction states;

[0139] Within each time step, a state transition window is constructed by combining the current velocity prediction vector with the acceleration range constraint, and possible velocity change directions are enumerated and the corresponding displacement is calculated.

[0140] Based on the predicted scene structure vector, an environmental feasibility test is performed on each predicted state candidate point, and state points that fall into obstacle areas or impassable areas are eliminated.

[0141] Each retained state candidate point is uniquely identified and encoded, and its cumulative path cost with the current position is recorded for path cost function calculation and sorting.

[0142] All candidate states that meet the constraints are organized into a set of position states in chronological order, which is used to construct a candidate sequence of navigation paths.

[0143] The state extension strategy proposed in this invention combines dynamic speed adjustment, scene constraint elimination, and path cost accumulation to realize the construction and screening of path states within the feasible region, ensuring the continuity and security of the generated path.

[0144] Example 1:

[0145] To verify the feasibility of this invention in practice, it was applied to a ground-based autonomous mobile robot equipped with a multi-mode navigation device in a complex scenario. The robot was performing continuous movement tasks in a densely built-up outdoor area. The environment contained uncertainties such as tall buildings blocking the view, weak GPS signals, complex road structures, and dynamic pedestrian interference. This was a typical scenario type with multi-source information interference and high-precision navigation difficulty.

[0146] In this scenario, traditional single navigation methods, such as GPS+IMU fusion-based positioning, experience frequent drift and trajectory jumps in areas with severe building obstruction. Visual SLAM-based solutions suffer from feature extraction failures and relocation difficulties in low-texture walls and repetitive structure areas, leading to path interruptions and execution failures. The positioning and navigation method proposed in this invention, based on multi-task collaborative optimization, relies on the joint input of GPS signals, raw IMU measurement data, and visual image sequences to construct a multi-task feature structure simultaneously addressing position estimation, velocity estimation, and scene modeling. During model training and inference, gradient propagation relationships and dynamic weighting mechanisms between tasks are introduced, enabling each sub-task to complement and optimize each other when information is compromised, thereby maintaining the stability and robustness of the overall navigation accuracy.

[0147] The robot underwent a 120-minute continuous operation test in a complex scenario, completing path planning and execution for 31 different routes. Compared to traditional methods, this invention significantly reduces positional errors in heavily obstructed areas. Statistics show that in areas with high GPS obstruction, the average positioning error of traditional methods reaches 3.82 meters, while the error of the implementation method of this invention is controlled within 0.96 meters, a reduction of 74.9%. Regarding path generation, this invention ensures the temporal continuity and spatial feasibility of paths through a state extension strategy and a collaborative path selection mechanism. The proportion of candidate paths satisfying both continuity and spatial traversability reaches 92.1%, compared to only 68.4% for traditional methods. Furthermore, in dynamically dense pedestrian intersections, the path pass rate generated by this invention is 85.3%, significantly higher than the 63.7% of the comparative methods, effectively avoiding local path jitter and redundant planning problems.

[0148] Regarding system operational stability, the present invention achieves only one navigation task interruption under continuous operation, compared to six interruptions in the traditional method, resulting in a task interruption rate reduction of over 80%. Furthermore, in terms of velocity estimation accuracy, comparison of trajectory velocity profiles shows that the present invention maintains a mean square error of velocity prediction within 0.27 m / s², while the traditional method has a mean error of 0.63 m / s², representing a jitter frequency reduction of nearly 57.1%. In addition, the average navigation command response delay is 0.38 seconds, lower than the 0.95 seconds of the traditional system, indicating that the present invention possesses faster path feedback capabilities at the command generation layer.

[0149] To more intuitively demonstrate the implementation effects of this invention, the following comparative data table is provided:

[0150] Table 1. Comparison of navigation performance between the method of this invention and traditional methods in complex environments.

[0151]

[0152] As can be seen from the above embodiments and measured data, the present invention is significantly superior to existing navigation technologies in terms of multi-source data fusion, collaborative optimization between highly dependent tasks, path continuity maintenance and navigation feedback response, and has good application feasibility and engineering promotion value.

[0153] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A positioning and navigation method based on multi-task cooperative optimization, characterized in that, The process includes the following steps: S1, collecting multi-source heterogeneous data of the target carrier and preprocessing it; S2, decomposing the positioning and navigation task based on the preprocessed multi-source heterogeneous data, constructing a position estimation subtask, a velocity estimation subtask, and a scene modeling subtask, and generating corresponding subtask feature sets. S3. Perform feature decoupling and information sharing operations on the feature set of subtasks, extract the cross-dependencies between subtasks, set gradient propagation paths according to the cross-dependencies, construct multi-task objective functions, and jointly calculate the loss value and collaborative error value of each subtask. S4. Based on the loss value and the collaborative error value, a dynamic weight allocation mechanism is used to perform collaborative optimization on each subtask, and gradient backpropagation and parameter update operations are performed to generate a multi-task collaborative parameter set. S5. Apply the multi-task collaborative parameter set to the position estimation subtask, velocity estimation subtask, and scene modeling subtask, perform path planning calculations, generate navigation path candidate sequences, and perform temporal consistency judgment and spatial constraint verification to select feasible path sequences; S6. Generate corresponding navigation control commands based on feasible path sequences and send them to the execution unit of the target vehicle to complete the positioning and navigation task.

2. The positioning and navigation method based on multi-task cooperative optimization according to claim 1, characterized in that, The multi-source heterogeneous data includes GPS signals, inertial measurement unit data, and visual image sequences.

3. The positioning and navigation method based on multi-task cooperative optimization according to claim 1, characterized in that, The preprocessing includes time interpolation of GPS signals, denoising and resampling of inertial measurement unit data, and distortion correction and inter-frame registration of visual image sequences.

4. The positioning and navigation method based on multi-task collaborative optimization according to claim 1, characterized in that, S3 specifically includes: S31, dividing the subtask feature set into a shared feature set and a dedicated feature set, the dedicated feature set corresponding to the position estimation subtask, velocity estimation subtask, and scene modeling subtask respectively, extracting feature differences between subtasks based on residual comparison operation, and constructing a cross-dependency matrix representing the feature dependency relationship between each subtask; S32, constructing a gradient propagation graph between subtasks based on the cross-dependency matrix, the gradient propagation graph setting the edge weights of the backpropagation path according to the subtask dependency strength, forming a path weight matrix that controls the degree of influence between subtasks; S33, constructing a multi-task objective function based on the shared feature set, dedicated feature set, and path weight matrix, jointly calculating the basic loss value of each subtask and the collaborative error term between subtasks, and realizing the weighted fusion calculation of multi-task loss by introducing subtask weighting coefficients and error adjustment coefficients.

5. A positioning and navigation method based on multi-task cooperative optimization according to claim 4, characterized in that, S32 specifically includes: S321, constructing a task node set based on the feature difference values ​​between each group of subtasks in the cross-dependency matrix, wherein the task node set corresponds to the position estimation subtask, velocity estimation subtask, and scene modeling subtask; S322, traversing any pairwise combinations in the task node set, setting the connection direction of the edges based on their feature difference values ​​in the cross-dependency matrix, and mapping the feature difference values ​​to connection strength values; S323, setting a connection strength threshold range, defining connections exceeding the upper limit as weakly correlated, and weakly correlated edges are assigned low weights in the gradient propagation graph. Multiple edges are represented as follows: connections below the lower limit are defined as strongly correlated, and strongly correlated edges are represented by high-weight edges in the gradient propagation graph; S324, the task node set and weighted connections are combined to form a directed acyclic graph structure, where the direction and weight of each edge are determined by the feature difference values ​​between subtasks; S325, the backpropagation path of the gradient of each subtask is determined based on the directed acyclic graph structure, and the propagation weight of each edge in the backpropagation path is used to control the degree of participation of error information in the propagation between tasks, and finally a path weight matrix is ​​formed, which is used for the construction of multi-task objective functions and collaborative optimization calculation.

6. The positioning and navigation method based on multi-task cooperative optimization according to claim 1, characterized in that, S4 specifically includes: S41. Based on the multi-task objective function, calculate the basic loss value and collaborative error term for the position estimation sub-task, velocity estimation sub-task, and scene modeling sub-task respectively, and then weight and combine them into the total error index of each sub-task. Then, normalize the error index of all sub-tasks to generate a standardized loss value; S42. Combine the standardized loss value and the gradient magnitude of each sub-task in the previous iteration to calculate the optimization weight of each sub-task in the current iteration. The gradient magnitude is used to suppress the unstable influence of oscillating gradients on parameter updates, and finally form a sub-task weight set; S43. Apply the sub-task weight set to the multi-task objective function. The function performs a weighted fusion of the loss values ​​of each subtask and the collaborative error term to construct a total loss function, and performs gradient backpropagation based on the total loss function to calculate the joint gradient direction; S44, according to the joint gradient direction and the current learning rate, calculates the partial derivative gradient of each parameter, and updates the parameters by introducing a gradient accumulation strategy. The update process uses a weighted average of the gradients of the current round and the previous round to avoid repeated oscillations and generates a new set of multi-task collaborative parameters; S45, writes the updated set of multi-task collaborative parameters into the task scheduling queue for use by the position estimation, velocity estimation and scene modeling subtasks for path planning and navigation command generation operations.

7. A positioning and navigation method based on multi-task cooperative optimization according to claim 6, characterized in that, The partial derivative gradient is constructed by creating a total loss function and extracting the difference between the predicted output and the true label for each subtask. Using a chain-like calculation rule, the error information is propagated forward from the output layer. In each layer, based on the calculation relationship between the current layer output and the parameters of the previous layer, combined with the influence of the subtask weights and the cooperative error term, the partial derivative value corresponding to each parameter is calculated item by item. The partial derivative results of all parameters are organized into a gradient vector according to a preset structure for parameter update calculation.

8. The positioning and navigation method based on multi-task cooperative optimization according to claim 1, characterized in that, S5 specifically includes: S51, applying the multi-task collaborative parameter set to the position estimation subtask, velocity estimation subtask, and scene modeling subtask respectively, and obtaining the current position prediction vector, velocity prediction vector, and scene structure vector respectively by performing forward computation on the feature set of the subtasks; S52, constructing a position state set using a state extension strategy based on the current position error, velocity deviation, and scene obstacle conflict degree of the position state, constructing a path cost function based on the current position error, velocity deviation, and scene obstacle conflict degree of the position state, and scoring and sorting the cost of each position state; S53, selecting position states with lower path costs to form a navigation path candidate sequence, performing a temporal consistency judgment on the navigation path candidate sequence to check whether the acceleration changes between each position state in the path meet the continuity condition, and performing a spatial constraint check to verify whether the position state is within a passable area in the scene; S54, summarizing the position states that simultaneously meet the temporal consistency judgment and spatial constraint check into a feasible path sequence set for use in generating navigation control commands.

9. A positioning and navigation method based on multi-task cooperative optimization according to claim 8, characterized in that, The forward computation performs a layer-by-layer mapping operation on the feature set of subtasks based on the multi-task collaborative parameter set. In each layer, the intermediate representation vector is calculated sequentially according to the pre-set linear transformation and nonlinear activation rules. The current position prediction vector of the position estimation subtask, the current velocity prediction vector of the velocity estimation subtask, and the current scene structure vector of the scene modeling subtask are generated in the output layer, respectively.

10. A positioning and navigation method based on multi-task cooperative optimization according to claim 8, characterized in that, The state expansion strategy specifically includes: using the current position prediction vector as the initial state, constructing a state sequence according to a set time step, and generating multiple predicted state candidate points; within each time step, constructing a state transition window by combining the current velocity prediction vector and acceleration range constraints, enumerating possible velocity change directions and calculating the corresponding displacement; performing an environmental feasibility check on each predicted state candidate point based on the predicted scene structure vector, and eliminating state points that fall into obstacle areas or impassable areas; assigning a unique identifier to each retained state candidate point and recording its cumulative path cost with the current position for path cost function calculation and sorting; and organizing all state candidate points that meet the constraints into a position state set in chronological order for constructing a navigation path candidate sequence.

Images