Lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference

Abstract

The invention relates to a dynamic interference-oriented lightweight outdoor multi-sensor fusion SLAM optimization method, and belongs to the technical field of SLAM, and the method comprises the following steps: S1, introducing a dynamic convolution DynamicConv module and an RSCD segmentation head into YOLOv11, and constructing a DyRS-YOLO; s2, detecting and filtering a dynamic target in real time by using Dys-YOLO, generating semantic information, and coupling the semantic information into an SLAM visual trajectory; s3, fusing the IMU and the GPS through AESKF to obtain a global trajectory, an adaptive adjustment process and a measurement noise covariance matrix; s4, the global track and the visual track are fused through coordinate alignment and scale transformation, and unification of visual data and global position information is achieved; and S5, integrating the steps S1 to S4, constructing a complete multi-sensor SLAM system, and fusing vision, IMU and GPS data to realize global positioning and mapping.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-sensor fusion SLAM (Simultaneous Local Area Mapping) and relates to a lightweight outdoor multi-sensor fusion SLAM (Simultaneous Local Area Mapping) optimization method oriented to dynamic interference. Background Art

[0002] Multi-sensor fusion SLAM (Simultaneous Localization and Mapping) is a key technology for enabling precise positioning and real-time mapping of autonomous robots in complex, dynamic environments. With the advancement of autonomous driving and intelligent inspection technologies, an increasing number of inspection robots are being deployed in outdoor urban environments. These scenes often contain numerous dynamic disturbances, such as pedestrians, vehicles, and mobile machinery. They also present challenges such as sparse textures, varying lighting conditions, and obstructed vision. These complexities severely impact the positioning accuracy and long-term consistency of traditional vision-based SLAM systems in dynamic environments.

[0003] Most existing SLAM systems assume a relatively static environment, making it difficult to effectively cope with frequent dynamic disturbances. For example, the classic ORB-SLAM3 relies primarily on feature matching and back-end optimization to construct high-precision trajectories. However, in dynamic environments, feature point mismatches caused by moving targets can lead to severe positioning errors and trajectory drift. Although some improved methods incorporate semantic information, such as DS-SLAM, which detects and removes dynamic target areas through semantic segmentation, they still suffer from high computational overhead, poor real-time performance, and severe false detection and missed detection, limiting their application in actual outdoor urban scenarios.

[0004] To address these issues, researchers have proposed a variety of SLAM optimization methods based on multi-sensor fusion in recent years, attempting to improve the system's positioning accuracy and robustness by comprehensively utilizing vision, IMU, GPS, and semantic information. For example, VINS-Fusion significantly improves trajectory consistency by fusing tightly coupled IMU and vision data; LvioFusion further introduces GPS constraints to reduce long-term drift; and SuMa++ uses semantic point cloud fusion technology to enhance the system's robustness in sparse environments. However, these methods often rely on high-cost lidar or depth cameras and have high requirements for computing resources, making them unsuitable for deployment on edge devices or lightweight robotic platforms. Summary of the Invention

[0005] In view of this, the object of the present invention is to provide a lightweight SLAM optimization method that significantly improves positioning accuracy and robustness in dynamic outdoor environments by fusing semantic information, IMU and GPS data.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference includes the following steps:

[0008] S1: The DynamicConv module and RSCD segmentation head are introduced into YOLOv11 to build a lightweight semantic segmentation network DyRS-YOLO. The DynamicConv module is used to adaptively adjust the convolution kernel according to input features, and the RSCD segmentation head reduces parameter overhead through convolution sharing and cross-layer reuse strategies.

[0009] S2: Use DyRS-YOLO to detect and filter dynamic targets in real time, generate semantic information, and couple the semantic information into the SLAM visual trajectory;

[0010] S3: Fusion of IMU and GPS via AESKF to obtain global trajectory, adaptive adjustment process and measurement noise covariance matrix;

[0011] S4: The global trajectory and the visual trajectory are fused through coordinate alignment and scale transformation to achieve the unification of visual data and global position information;

[0012] S5: Integrate steps S1-S4 to build a complete multi-sensor SLAM system, integrating vision, IMU and GPS data to achieve global positioning and mapping.

[0013] Furthermore, the dynamic convolution DynamicConv module in step S1 is implemented as follows:

[0014] Extract feature context information through global average pooling;

[0015] A lightweight MLP is used to calculate the weight coefficient of each convolution kernel and normalized using the Softmax function;

[0016] According to the calculated weight coefficients, multiple expert convolution kernels are dynamically combined to form an adaptive convolution kernel;

[0017] Adaptive convolution kernels are used in the feature extraction layer to process the input features.

[0018] Furthermore, the RSCD segmentation head in step S1 includes: integrating the independent convolution branches of regression, classification and mask prediction into a shared convolution layer to realize parameter reuse; sharing the convolution kernel between feature layers of different scales and adopting a cross-layer reuse strategy so that adjacent feature layers reuse the same convolution kernel; adding a scale adaptive module to enhance the cross-layer feature fusion capability.

[0019] Furthermore, step S2 specifically includes the following steps:

[0020] DyRS-YOLO is used to detect and filter dynamic targets in real time. The DynamicConv module is used to extract features. The RSCD segmentation head is used to achieve lightweight detection. This generates dynamic masks during the detection phase and filters dynamic targets in real time.

[0021] Integrate semantic information in the visual SLAM front-end, couple the dynamic mask generated by DyRS-YOLO into the tracking thread of the SLAM system, and filter out feature points in the dynamic area based on the mask during the key point extraction stage;

[0022] Dynamic objects are divided according to semantic information, and the dynamic target area is identified using the dynamic mask generated by DyRS-YOLO. After extracting key points, the feature points in the dynamic target area are removed.

[0023] Furthermore, step S3 specifically includes the following steps:

[0024] Collect IMU and GPS data, where IMU is used to provide high-frequency inertial measurement information and GPS is used to provide low-frequency absolute position information;

[0025] Based on the error state Kalman filter (ESKF) framework, an adaptive error state Kalman filter (AESKF) is introduced to dynamically adjust the covariance matrix of process noise and measurement noise.

[0026] The collected IMU and GPS data are input into AESKF for data fusion, the innovation residual is monitored in real time, and the filter parameters are adaptively adjusted according to environmental changes;

[0027] Generate global trajectories and achieve tight coupling fusion of high-frequency IMU data and low-frequency GPS data.

[0028] Furthermore, step S4 specifically includes the following steps:

[0029] Select keyframes in the visual trajectory and match their timestamps with corresponding points in the global trajectory;

[0030] Compute the rotation matrix and scale factor between the visual track and the global track;

[0031] Correct the visual trajectory through rigid body transformation so that it is in the same coordinate system as the global trajectory;

[0032] Based on the aligned trajectories, error and outlier optimization is performed.

[0033] The beneficial effects of the present invention are as follows: the present invention integrates semantic segmentation, IMU-GPS fusion and global trajectory optimization modules, significantly reducing computational complexity while improving the positioning accuracy and long-term stability of the system in dynamic scenarios. The method uses a lightweight DyRS-YOLO model for real-time dynamic object detection and fuses IMU and GPS data through an adaptive error state Kalman filter (AESKF) to achieve high-frequency and accurate global pose estimation. In addition, through coordinate alignment and scale transformation, the scale drift and posture error in the visual trajectory are further corrected to construct a complete multi-sensor SLAM system, which significantly improves the positioning accuracy and system robustness of the outdoor urban inspection robot.

[0034] First, the present invention designs a DynamicConv module to achieve adaptive adjustment of convolution kernels, effectively reducing redundant calculations and improving feature extraction efficiency; secondly, the present invention proposes an RSCD segmentation head, which significantly reduces parameter overhead through convolution sharing and cross-layer reuse strategies, while improving the inference speed of the model; on this basis, the present invention constructs a DyRS-YOLO lightweight semantic segmentation network, combines DynamicConv and RSCD, and achieves efficient dynamic object detection, further reducing computational complexity and improving positioning robustness in dynamic scenes. Subsequently, the present invention fuses IMU and GPS data through an adaptive error state Kalman filter (AESKF), significantly improving the consistency and long-term stability of trajectory fusion, while solving the problem of decreased accuracy of traditional filters in complex dynamic environments; in addition, in the trajectory post-processing stage, the present invention proposes a global alignment method based on rigid body transformation, which corrects the errors in scale and coordinate system between the visual trajectory and the global trajectory, ensuring the global consistency of pose estimation. Finally, the present invention builds a complete multi-sensor SLAM system, which integrates semantic, visual, IMU and GPS data to achieve long-term stable global positioning and mapping, providing autonomous inspection robots in complex dynamic environments with higher positioning accuracy and environmental perception capabilities.

[0035] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0037] Figure 1Schematic diagram of the MoE mechanism; (a) is the normal layer; (b) is the MoE layer, dynamically selecting expert calculations;

[0038] Figure 2 This is the RSCD segmentation header structure diagram;

[0039] Figure 3 This is the semantic segmentation effect diagram;

[0040] Figure 4 This is the effect diagram of coupling semantic information into the SLAM system;

[0041] Figure 5 This is a schematic diagram of coordinate alignment;

[0042] Figure 6 This is the complete SLAM system structure. DETAILED DESCRIPTION

[0043] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0044] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0045] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0046] Example 1:

[0047] The present invention provides a lightweight outdoor multi-sensor fusion SLAM optimization method oriented to dynamic interference, comprising the following steps:

[0048] S1: DynamicConv is introduced into YOLOv11n to adaptively adjust the convolution kernel according to the input features, reduce redundant calculations, and improve feature extraction efficiency. A new RSCD segmentation head is proposed to reduce parameter overhead and improve inference speed through convolution sharing and cross-layer reuse strategies. A new lightweight semantic segmentation network DyRS-YOLO is constructed based on this, which combines DynamicConv and RSCD to further reduce computational complexity and improve the efficiency of dynamic object detection.

[0049] Specifically, such as Figure 1 As shown in (a) and (b), define a DynamicConv module that includes global average pooling, MLP weight calculation, and convolution kernel combination; replace the C3k2 module in YOLOv11n with the C3k2_DynamicConv module; modify the network configuration file to ensure that the number of channels in the DynamicConv module matches the number of channels in the backbone network; adjust the number of convolution layers in C3k2_DynamicConv, and call the adaptive convolution kernel calculation module during the network forward propagation process.

[0050] like Figure 2 As shown in the figure, an RSCD segmentation head is constructed, and the adaptive convolution kernel calculation module is called during the forward propagation of the network. By reusing cross-layer convolution, convolution kernels of feature layers of different scales can be shared. The RSCD segmentation head is integrated into DyRS-YOLO, and the RSCD segmentation head replaces the original segmentation head module of YOLOv11n. A shared convolution layer is added to the detection head, and the network configuration file is modified to ensure that the number of channels and feature layer structure are consistent with the backbone network.

[0051] Use Ubuntu 20.04 as the operating system and install necessary dependency libraries such as Python 3.11, PyTorch, OpenCV, and ROS. Build the DyRS-YOLO network, including the backbone network, C3k2_DynamicConv module, and RSCD segmentation head. Use the COCO dataset in the pre-training phase, and select data from the three categories of person, car, and bicycle for model initialization.

[0052] S2: Couple semantic information into visual trajectories to enhance localization robustness in dynamic environments.

[0053] Specifically, such as Figure 3As shown in the figure, the pre-trained DyRS-YOLO model is loaded into the inference environment, and the model weights pre-trained on the COCO dataset are used. The test images in the KITTI dataset are used as input to perform dynamic target detection and segmentation prediction. The prediction results are post-processed to extract the segmentation masks of the three categories of person, car, and bicycle for the subsequent dynamic feature point removal in the visual trajectory optimization stage. In the segmentation results, the dynamic area is divided according to the target category to generate a complete dynamic mask map. At the front end of the SLAM system, the feature points of the area containing dynamic objects are filtered to reduce the interference of dynamic targets on the pose estimation. Figure 4 As shown in Figure 3, the segmentation mask is integrated with the feature extraction module in ORB-SLAM3, and feature points belonging to dynamic areas are eliminated in the tracking thread to ensure that only feature points in static environments are retained.

[0054] S3: The global trajectory is obtained by fusing IMU and GPS through AESKF, and the process and measurement noise covariance matrix are adaptively adjusted to improve trajectory consistency and long-term stability.

[0055] Specifically, IMU and GPS data are collected in real time through a self-developed mobile robot platform. The IMU sampling frequency is 100Hz and the GPS update frequency is 10Hz. This ensures that the IMU and GPS data timestamps are aligned and stored as time series data in a unified format. An adaptive error state Kalman filter (AESKF) is introduced in the filtering module to dynamically adjust the process noise and measurement noise covariance matrix. State variables, including position, velocity, and attitude, are set in the initialization phase, and the IMU and GPS observation models are defined. Based on the absolute position information provided by the GPS, low-frequency correction is performed on the IMU prediction results to reduce the IMU integration error and achieve tight coupling fusion of IMU and GPS data.

[0056] S4: The global trajectory and visual trajectory are fused through coordinate alignment and scale transformation to solve the problems of scale drift and coordinate misalignment, and achieve the unification of visual data and global position information.

[0057] Specifically, such as Figure 5 As shown in the figure, keyframes are selected from the visual trajectory generated by ORB-SLAM3, and the pose and timestamp of each keyframe are extracted. The keyframe timestamps in the visual trajectory are matched with the GPS timestamps in the global trajectory to ensure temporal alignment between the two. The rotation matrix and scale factor are calculated between the visual trajectory and the global trajectory to resolve the issues of coordinate system rotation and scale inconsistency. A rigid body transformation is applied to convert the visual trajectory to the global coordinate system to achieve position and pose consistency. Based on the aligned trajectory, the matching results are further optimized to eliminate outliers and ensure the stability and accuracy of trajectory fusion.

[0058] S5: Through the above steps, a complete multi-sensor SLAM system is constructed, which integrates vision, IMU and GPS data to achieve long-term stable global positioning and mapping. The complete structure is as follows: Figure 6 shown.

[0059] Example 2:

[0060] In this embodiment, the DynamicConv module in step S1 is specifically implemented as follows:

[0061] First, global average pooling is performed on the input features to extract overall contextual information. Global average pooling effectively reduces the size of the feature matrix while preserving global statistical information, providing a foundation for subsequent weight calculations. This approach not only reduces computational effort but also enhances the compactness of feature representation, making feature extraction more efficient.

[0062] Next, the pooled global context information is fed into a lightweight multi-layer perceptron (MLP) to calculate the weight coefficients for each convolution kernel. To ensure that the weight coefficients are comparable across different convolution kernels and adhere to the properties of the probability distribution, they are finally normalized using the Softmax function. This not only improves the interpretability of the weights but also ensures greater robustness in the selection of adaptive convolutions.

[0063] According to the calculated weight coefficients, multiple expert convolution kernels are dynamically combined to form an adaptive convolution kernel:

[0064]

[0065] The formula defines the basic process of dynamic convolution, which is to combine multiple expert convolution kernels W i Dynamically combine to form the adaptive convolution kernel W′.

[0066] Dynamic convolution kernels are used in the feature extraction layer to reduce redundant calculations and improve the model's feature expression capabilities in complex scenarios. The adaptive convolution kernel W' is applied to the input feature X to obtain the final output feature Y. The calculation process is:

[0067] Y=X*W′

[0068] This mechanism is applied in the C3k2 module of YOLOv11n to significantly improve feature extraction efficiency and reduce computational costs.

[0069] Example 3:

[0070] In this embodiment, step S3 specifically includes the following steps:

[0071] IMU and GPS data are collected, with the IMU sampling frequency at 100Hz and the GPS update frequency at 10Hz. The IMU provides high-frequency inertial measurement information, including triaxial acceleration and angular velocity data, while the GPS provides low-frequency absolute position information, primarily used to correct IMU accumulated errors and ensure global trajectory consistency.

[0072] Based on the Error State Kalman Filter (ESKF) framework, the Adaptive Error State Kalman Filter (AESKF) is introduced to dynamically adjust the process noise and measurement noise covariance matrix to improve positioning accuracy in complex dynamic environments. During the prediction phase, the AESKF uses an adaptive factor to adjust the state covariance matrix P in real time to reflect environmental changes. Its calculation formula is:

[0073]

[0074] Among them, the adaptive factor α k Dynamic adjustment based on state estimation error:

[0075]

[0076] Where Δx k is the state estimation error, and the trace tr(P k ) related.

[0077] The collected IMU and GPS data are input into AESKF for data fusion. First, IMU data is used for high-frequency prediction to calculate the state estimate X k and covariance matrix

[0078]

[0079]

[0080] Then, during GPS updates, the state estimate is corrected based on the measurements:

[0081]

[0082]

[0083]

[0084] Among them, K k It is an adaptive Kalman gain matrix that can dynamically adjust the fusion weight of IMU and GPS to ensure that the filter can still effectively estimate the system state when the GPS signal fluctuates or is lost

[0085] Generate global trajectories and achieve tight coupling fusion of high-frequency IMU data and low-frequency GPS data, significantly improving the consistency and long-term stability of the global trajectory.

[0086] Example 3:

[0087] In this embodiment, step S4 specifically includes the following steps:

[0088] S41: Select the keyframes in the visual track and match the timestamps with the corresponding points in the global track. To ensure the alignment of the two on the time axis, linear interpolation is used to calculate the closest timestamp in the global track. The interpolation formula is as follows:

[0089]

[0090] Where P is the global trajectory pose at time t, t v is the timestamp of the visual trajectory, t p and t p+1 They are the closest to t in the global trajectory v The two timestamps before and after.

[0091] S42: Calculate the rotation matrix and scale factor between the visual track and the global track to solve the problem of inconsistent coordinate system rotation and scale. The specific calculation includes:

[0092] (1) Rotation matrix R gv It is calculated by the direction vector, and the formula is as follows:

[0093]

[0094]

[0095]

[0096] Where θ is the rotation angle, μ is the rotation axis, [μ] × is an antisymmetric matrix.

[0097] (2) The scaling factor k is obtained by minimizing the sum of squared errors:

[0098] argmin{∑||kR gv x v -x g || 2}

[0099] This optimization process is often performed using the singular value decomposition (SVD) method to reduce computational complexity and improve efficiency.

[0100] S43: Correct the visual trajectory through rigid body transformation so that it is in the same coordinate system as the global trajectory, achieving global consistency in position and posture. The form of rigid body transformation is:

[0101] x g =kR gv x v +t

[0102] Where t is the translation vector, representing the spatial displacement between the visual trajectory and the global trajectory.

[0103] S44: Based on the aligned trajectories, further optimize the errors and outliers to improve the accuracy and long-term stability of trajectory fusion. The specific steps include:

[0104] (1) Calculate the error between each pair of matching points:

[0105] s=||kR gv x v -x g ||

[0106] (2) Define a threshold range [0.98, 1.02] to filter out local outliers in the trajectory, and further correct the error points larger than this range.

[0107] Example 4:

[0108] An electronic device comprising a memory and a processor;

[0109] The memory is used to store computer programs;

[0110] The processor is configured to implement the method described in Example 1 when executing the computer program.

[0111] Example 5:

[0112] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in Example 1 is implemented.

[0113] Example 6:

[0114] A computer program product includes a computer program, which implements the method described in embodiment 1 when executed by a processor.

[0115] In the above embodiments, references to "this embodiment" in the specification indicate that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple occurrences of "this embodiment" do not necessarily refer to the same embodiment.

[0116] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the present invention are intended to encompass all such alternatives, modifications, and variations that fall within the broad scope of the appended claims.

[0117] Regarding the computer-readable storage medium in this embodiment, those skilled in the art will appreciate that all or part of the steps in the aforementioned method embodiments can be implemented using hardware associated with the computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps in the aforementioned method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0118] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication with each other. The memory is used to store computer programs, the communication interface is used for communication, and the processor and the transceiver are used to run computer programs so that the electronic terminal executes the various steps of the above method.

[0119] In this embodiment, the memory may include a random access memory (RAM), and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage.

[0120] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0121] The present invention can be used in a wide variety of general-purpose or special-purpose computing system environments or configurations, such as personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the above.

[0122] The present invention may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference, characterized by: The following steps are involved: S1: The DynamicConv module and RSCD segmentation head are introduced into YOLOv11 to build a lightweight semantic segmentation network DyRS-YOLO. The DynamicConv module is used to adaptively adjust the convolution kernel according to input features, and the RSCD segmentation head reduces parameter overhead through convolution sharing and cross-layer reuse strategies. S2: Use DyRS-YOLO to detect and filter dynamic targets in real time, generate semantic information, and couple the semantic information into the SLAM visual trajectory; S3: Fusion of IMU and GPS via AESKF to obtain global trajectory, adaptive adjustment process and measurement noise covariance matrix; S4: The global trajectory and the visual trajectory are fused through coordinate alignment and scale transformation to achieve the unification of visual data and global position information; S5: Integrate steps S1-S4 to build a complete multi-sensor SLAM system, integrating vision, IMU and GPS data to achieve global positioning and mapping.

2. The lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to claim 1, wherein: The steps for implementing the dynamic convolution DynamicConv module in step S1 are as follows: Extract feature context information through global average pooling; A lightweight MLP is used to calculate the weight coefficient of each convolution kernel and normalized using the Softmax function; According to the calculated weight coefficients, multiple expert convolution kernels are dynamically combined to form an adaptive convolution kernel; Adaptive convolution kernels are used in the feature extraction layer to process the input features.

3. The lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to claim 1, wherein: The RSCD segmentation head described in step S1 includes: integrating the independent convolution branches of regression, classification and mask prediction into a shared convolution layer to achieve parameter reuse; sharing convolution kernels between feature layers of different scales and adopting a cross-layer reuse strategy so that adjacent feature layers reuse the same convolution kernel; adding a scale adaptive module to enhance the cross-layer feature fusion capability.

4. The lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to claim 1, wherein: Step S2 specifically includes the following steps: DyRS-YOLO is used to detect and filter dynamic targets in real time. The DynamicConv module is used to extract features. The RSCD segmentation head is used to achieve lightweight detection. This generates dynamic masks during the detection phase and filters dynamic targets in real time. Integrate semantic information in the visual SLAM front-end, couple the dynamic mask generated by DyRS-YOLO into the tracking thread of the SLAM system, and filter out feature points in the dynamic area based on the mask during the key point extraction stage; Dynamic objects are divided according to semantic information, and the dynamic target area is identified using the dynamic mask generated by DyRS-YOLO. After extracting key points, the feature points in the dynamic target area are removed.

5. The lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to claim 1, characterized in that: Step S3 specifically includes the following steps: Collect IMU and GPS data, where IMU is used to provide high-frequency inertial measurement information and GPS is used to provide low-frequency absolute position information; Based on the error state Kalman filter (ESKF) framework, an adaptive error state Kalman filter (AESKF) is introduced to dynamically adjust the covariance matrix of process noise and measurement noise. The collected IMU and GPS data are input into AESKF for data fusion, the innovation residual is monitored in real time, and the filter parameters are adaptively adjusted according to environmental changes; Generate global trajectories and achieve tight coupling fusion of high-frequency IMU data and low-frequency GPS data.

6. The lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to claim 1, characterized in that: The step S4 specifically includes the following steps: Select keyframes in the visual trajectory and match their timestamps with corresponding points in the global trajectory; Compute the rotation matrix and scale factor between the visual track and the global track; Correct the visual trajectory through rigid body transformation so that it is in the same coordinate system as the global trajectory; Based on the aligned trajectories, error and outlier optimization is performed.

7. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the dynamic interference-oriented lightweight outdoor multi-sensor fusion SLAM optimization method according to any one of claims 1 to 6 when executing the computer program.

8. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to any one of claims 1 to 6 is implemented.

9. A computer program product, characterized in that: The invention comprises a computer program, which, when executed by a processor, implements the lightweight outdoor multi-sensor fusion SLAM optimization method for dynamic interference according to any one of claims 1 to 6.

Images