A SLAM navigation method and system for a mobile intelligent cabinet

Abstract

The present application belongs to the technical field of mobile robot navigation, and particularly relates to a SLAM navigation method and system for a mobile intelligent cabinet, which comprises the following steps: acquiring a laser point cloud sequence and a grayscale image of the mobile intelligent cabinet and performing data cleaning; obtaining a geometric feature index based on the spatial jump distribution of the laser point cloud sequence; obtaining a visual texture index based on the local dispersion of the image gradient; calculating laser dynamic weight and visual dynamic weight by using the geometric feature index and the visual texture index, weighting and fusing the pose change quantity calculated by the single-line laser radar odometry and the visual odometry to obtain a fused pose quantity; and updating the global state based on the fused pose quantity and driving autonomous navigation. The present application can adjust the sensor weight in real time according to the environmental characteristics, solves the problem of positioning divergence in the long corridor of a shopping mall or a high-reflectivity environment, and improves the robustness of navigation.

Description

Technical Field

[0001] This invention relates to the field of mobile robot navigation technology. More specifically, this invention relates to a SLAM navigation method and system for mobile intelligent cabinets. Background Technology

[0002] With the rapid development of service robot technology, mobile smart lockers, as an emerging retail terminal, are widely used in complex indoor scenarios such as shopping malls and office buildings. They can achieve autonomous navigation and on-demand sales services. High-precision autonomous positioning and navigation capabilities are the foundation for their stable operation and directly affect service efficiency and equipment safety. To achieve autonomous movement in unknown environments, the industry generally uses Simultaneous Localization and Mapping (SLAM) technology to build environmental maps in real time and estimate robot pose. Among these technologies, multi-sensor fusion solutions combining LiDAR (Light Detection and Ranging) and visual sensors have become the mainstream technical approach due to their complementary perception characteristics.

[0003] However, the operating environment of commercial complexes exhibits typical non-cooperative characteristics, posing a significant challenge to traditional navigation algorithms. Specifically, this manifests in the prevalence of long, straight corridors and the extensive use of glass curtain walls or high-gloss floor tiles. The former causes the laser point cloud sequence acquired by LiDAR to degenerate geometrically into parallel lines lacking longitudinal constraints, while the latter generates optical reflection, refraction, or transmission phenomena, thus interfering with feature extraction by visual sensors. Existing multi-sensor fusion frameworks, such as algorithms based on the Extended Kalman Filter (EKF), typically employ fixed covariance matrices or static weight allocation strategies during the fusion process. These frameworks assume that sensor observation noise is constant or only related to the sensor's own physical properties, failing to perceive and respond to real-time changes in environmental geometry and texture richness.

[0004] When mobile smart cabinets move from feature-rich areas into degraded scenarios such as long corridors or high-reflectivity areas, if traditional algorithms still blindly trust sensor data that has actually failed due to environmental factors, it will inevitably lead to divergence in positioning trajectory, distortion of mapping, or even collision accidents. Summary of the Invention

[0005] To address the technical problems of decreased positioning accuracy and mapping distortion in long corridors or high-reflectivity scenarios, the present invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a SLAM navigation method for mobile smart cabinets, comprising: acquiring a laser point cloud sequence and a grayscale image of the mobile smart cabinet and performing data cleaning; dividing the laser point cloud sequence into multiple sectors, counting the number of depth transition points in each sector, and obtaining a geometric feature index based on the number of depth transition points; dividing the grayscale image into multiple grid regions, filtering out effective texture grids, and obtaining a visual texture index based on the number of effective texture grids and the average gradient magnitude of the entire image; obtaining laser dynamic weights and visual dynamic weights based on the geometric feature index and the visual texture index, and obtaining a fused pose quantity based on the laser dynamic weights and the visual dynamic weights to realize navigation of the mobile smart cabinet.

[0007] This invention, by acquiring the geometric feature index of the laser point cloud sequence and the visual texture index of the grayscale image, can assess in real time the friendliness of the current environment to different sensors, and adjust the laser dynamic weight and visual dynamic weight accordingly. Thus, when the mobile smart cabinet faces a scenario where a single sensor is easily interfered with, such as a long straight corridor or glass curtain wall, it can automatically focus on the sensor with more reliable data, reduce positioning drift and improve the environmental adaptability of the navigation system.

[0008] Preferably, the step of counting the number of depth jump points in each sector includes: calculating the distance difference between adjacent points in the laser point cloud sequence; determining that a depth jump point exists in response to the distance difference between adjacent points exceeding a preset threshold; and counting the number of depth jump points contained in each sector.

[0009] This invention captures discontinuous geometric features in the environment by calculating the distance difference between adjacent points in a laser point cloud sequence and counting the number of depth jump points. This serves as the basis for evaluating the quality of single-line lidar observation data, ensuring that the mobile smart cabinet can promptly identify the lack of geometric information in feature-sparse areas and reducing longitudinal positioning errors caused by blindly trusting laser data.

[0010] Preferably, the geometric feature index satisfies the expression: ;in, Indicates geometric characteristic index; Indicates the total number of sectors in the partition; Indicates the sector index number; Indicates the first The number of depth transition points detected within a sector; Represents the characteristic saturation constant; Represents the natural constant.

[0011] Preferably, the step of filtering out effective texture meshes includes: calculating the gradient magnitude of pixels within each mesh region; and determining the mesh as an effective texture mesh in response to the average gradient magnitude within the mesh being greater than a base threshold and the variance of the gradient direction being greater than a direction threshold.

[0012] This invention filters effective texture meshes by combining the variance of average gradient magnitude and gradient direction. It can distinguish low-texture areas such as white walls or single lines from high-texture areas containing rich details in grayscale images, thereby providing an accurate texture distribution basis for evaluating the reliability of visual odometry and reducing the risk of visual positioning failure of mobile smart cabinets in weak texture environments.

[0013] Preferably, the visual texture index satisfies the expression: ;in, Indicates the visual texture index; This indicates the number of texture meshes that were determined to be valid. Indicates the total number of grid cells; This represents the average gradient magnitude of all valid texture meshes in the entire image; Indicates the reference gradient value; Indicates the reflection suppression coefficient; Represents the absolute value symbol.

[0014] This invention introduces a reflection suppression coefficient and a reference gradient value to obtain a visual texture index. It can reduce the visual texture index when the average gradient amplitude of the whole image is abnormally increased due to strong reflection, thereby suppressing the influence of visual data affected by illumination on the final positioning result and reducing the visual odometry drift caused by light spots on the mobile smart cabinet near the high-gloss floor or glass curtain wall of the shopping mall.

[0015] Preferably, the laser dynamic weights satisfy the expression: ;in, Indicates the dynamic weight of the laser; Indicates geometric characteristic index; Indicates the visual texture index; It represents a tiny quantity that prevents the denominator from being zero.

[0016] This invention achieves a complementary weight allocation mechanism by calculating the geometric feature index, the visual texture index, and a small amount to prevent the denominator from being zero, and obtaining the laser dynamic weight. This mechanism enables the weight to be automatically tilted when one sensor feature is missing while another sensor feature is rich, ensuring that the mobile smart cabinet maintains a relatively smooth and accurate pose estimation during the switching of different environmental features.

[0017] Preferably, the fused pose quantity satisfies the expression: ;in, Indicates the fused pose quantity; Indicates the dynamic weight of the laser; This represents the change in pose calculated using a single-line lidar odometry. This represents the change in pose calculated using visual odometry.

[0018] Preferably, the data cleaning includes: performing distance filtering on the laser point cloud sequence to remove points that are outside the range or too close; and performing noise reduction processing on the grayscale image to remove random noise.

[0019] Preferably, the characteristic saturation constant is set to 5.

[0020] Secondly, the present invention provides a SLAM navigation system for mobile smart cabinets, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned SLAM navigation method for mobile smart cabinets is implemented.

[0021] By adopting the above technical solution, a computer program for the SLAM navigation method for mobile smart cabinets is generated and stored in a memory for loading and execution by a processor. This allows for the creation of terminal devices based on the memory and processor, facilitating their use.

[0022] The beneficial effects of this invention are as follows:

[0023] This invention comprehensively utilizes the geometric structure information of laser point cloud sequences and the texture detail information of grayscale images. Through real-time evaluation and complementary fusion of dual feature indices, it solves the problem of decreased data reliability of a single sensor in specific degradation scenarios and reduces the probability of location loss of mobile smart cabinets in complex environments such as shopping malls.

[0024] This invention is specifically designed for long, straight corridors and highly reflective materials commonly found in commercial complexes. It utilizes depth transition points to identify geometric degradation and gradient features to identify visual interference. This allows the mobile smart cabinet to dynamically adjust its reliance on LiDAR and visual cameras as it moves through different feature areas, improving the robustness of the navigation system in non-cooperative environments.

[0025] This invention employs a sector partitioning and grid region partitioning processing strategy to transform massive amounts of raw data into lightweight feature indices. While ensuring that environmental features are fully extracted, it reduces computational complexity, enabling mobile smart cabinets to achieve real-time multi-sensor fusion positioning and navigation with limited computing resources. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a SLAM navigation method for mobile smart cabinets according to the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the division of laser point cloud sectors and the detection of depth jump points;

[0028] Figure 3 This is a schematic diagram illustrating the statistical points of depth jumps in each sector. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] This invention discloses a SLAM navigation method for mobile smart cabinets, referring to... Figure 1 This includes steps S1-S4:

[0032] S1. Acquire the laser point cloud sequence and grayscale image of the mobile smart cabinet and perform data cleaning.

[0033] It should be noted that, in order to enable the mobile smart cabinet to navigate and locate autonomously in complex environments such as shopping malls or office buildings, this invention uses sensors mounted on the smart cabinet to perceive the geometric structure and texture information of the surrounding environment in real time.

[0034] Specifically, the single-line LiDAR installed in the smart cabinet is activated to collect distance data on the horizontal plane at a set frequency (e.g., 10Hz), forming a laser point cloud sequence composed of several ranging points; at the same time, the RGB-D depth camera is activated to acquire a grayscale image of the environment in front of the mobile smart cabinet; the laser point cloud sequence is filtered by distance to remove invalid points that are beyond the range (e.g., greater than 20 meters) or too close (e.g., less than 0.1 meters); the grayscale image is denoised to remove random noise caused by low illumination, resulting in clear input data.

[0035] S2. Divide the laser point cloud sequence into multiple sectors, count the number of depth transition points in each sector, and obtain the geometric feature index based on the number of depth transition points.

[0036] It should be noted that the stability of single-line lidar positioning is highly dependent on the richness of environmental geometric features. In long straight corridors, the laser point cloud sequence is continuous and smooth, resulting in a lack of longitudinal constraints. In feature-rich areas, the laser point cloud sequence will produce depth jumps at the edges of objects. This invention uses the number of features that a single-line lidar can capture in the current environment as a basis for trusting the laser data.

[0037] Specifically, a frame of laser point cloud sequence is divided equally according to the scanning angle into... Each sector, experience points ; Calculate the distance difference between adjacent points in the laser point cloud sequence. When the distance difference between adjacent points exceeds a preset threshold (empirical value 0.3 meters), determine that there is a geometric feature at that location and record it as a depth transition point; count the number of depth transition points contained in each sector.

[0038] Furthermore, based on the distribution of depth transition points, a geometric feature index is obtained, which satisfies the expression:

[0039]

[0040] In the formula, It is a geometric characteristic index; This represents the total number of sectors. It is the sector index number; It is the first The number of depth transition points detected within a sector; It is the characteristic saturation constant; It is a natural constant.

[0041] In the formula, The relative density of features within a single sector was evaluated; It is a saturation mapping function, which is used when there is no depth transition point within the sector. When it is 0, The term is 0; as the number of depth transition points increases, The number of items increases rapidly; when the number of depth transition points is very large, The term tends towards 1, simulating the feature requirements of the localization algorithm, meaning that a sufficient number of features is enough to determine the constraints without needing an infinite number. By summing and averaging over all sectors, this expression comprehensively evaluates the total number of features and the uniformity of the angular distribution. If only some sectors in a long corridor have data, then... Extremely low values ​​reflect the unreliability of laser data; if there are abrupt change points on all four sides, then... A higher value indicates that the single-line lidar positioning is reliable.

[0042] Regarding the characteristic saturation constant In this embodiment, the empirical value is 5. This constant is used to control the sensitivity to the number of features. If the setting is too small, a few depth transition points will cause the geometric feature exponent to saturate, making the algorithm unable to distinguish between regions with general features and regions with extremely rich features; if... Setting the value too high will require a large number of depth transition points to achieve a high level of trust, which may result in the laser dynamic weight remaining low for a long period of time.

[0043] For example, Figure 2 This is a schematic diagram of laser point cloud sector division and depth transition point detection. In the diagram, blue dots represent ordinary laser point clouds collected by single-line lidar, red pentagrams represent detected depth transition points, black dashed lines represent multiple sectors, and black dots represent the location of mobile smart cabinets. It shows how the system divides environmental data into different regions and identifies key points with rich geometric features.

[0044] S3. Divide the grayscale image into multiple grid regions, filter out the effective texture grids, and obtain the visual texture index based on the number of effective texture grids and the average gradient magnitude of the entire image.

[0045] It should be noted that in a shopping mall environment, white walls can lead to a loss of image gradients, while glass curtain walls or high-gloss floors can produce false textures or strong reflections. These factors can reduce the reliability of visual positioning. This invention evaluates the credibility of visual data by identifying effective texture areas in the image and suppressing areas with abnormal lighting.

[0046] Specifically, the preprocessed grayscale image is divided into several grid regions, for example, into... The grid is constructed; the gradient magnitude is calculated for each pixel in the grid; the effective texture grids are selected based on the average gradient magnitude of the grid being greater than the base threshold (e.g., 20) and the variance of the gradient direction being greater than the direction threshold (e.g., 0.2) to exclude interference from white walls and single lines.

[0047] Furthermore, based on the proportion of effective texture meshes and the gradient features of the entire image, a visual texture index is obtained, which satisfies the expression:

[0048]

[0049] In the formula, It is the visual texture index; It is the number of texture meshes that are determined to be valid; It is the total number of grid cells; It is the average gradient magnitude of all valid texture meshes in the entire image; It is the reference gradient value; It is the reflection suppression coefficient, and its unit is the reciprocal of the gradient magnitude; It is an absolute value.

[0050] In the formula, This visually reflects the proportion of the field of view containing effective texture content. If facing a white wall, The terms approach 0, thus reducing the visual dynamic weight; the denominator It is a suppression factor for strongly reflective or excessively dark environments. Under strong reflective or direct light, images will exhibit extremely bright spots, causing the average gradient amplitude to rise abnormally and far exceed the reference gradient value. This results in a larger denominator, leading to... The decrease in light intensity suppresses the impact of visual data, which is affected by strong light interference, on positioning.

[0051] Regarding the reference gradient value and reflection suppression coefficient In this embodiment The value is 40. The value is 0.05. The calibration should be performed under normal diffuse reflection conditions based on the camera's light-sensing characteristics. If the setting deviation is large, the visual dynamic weights under normal conditions will be incorrectly suppressed. Control the severity of penalties for abnormal lighting conditions, if Excessive intensity, even slight changes in lighting can cause drastic fluctuations in visual dynamic weights; if If it is too small, it cannot effectively suppress strong light interference.

[0052] For example, Figure 3 This is a statistical diagram of depth jump points in each sector, where the orange bars represent the number of depth jump points counted in each sector. By statistically analyzing the distribution of feature points in each sector, the system can calculate the geometric feature index, thereby assessing the reliability of laser positioning in the current environment and providing a basis for subsequent weight allocation.

[0053] S4. Obtain laser dynamic weights and visual dynamic weights based on geometric feature indices and visual texture indices, and obtain fused pose values ​​based on the laser dynamic weights and visual dynamic weights to realize mobile smart cabinet navigation.

[0054] It should be noted that after evaluating the environmental adaptability of single-line lidar and visual sensors respectively, this invention uses these two indices to dynamically generate fusion weights to achieve adaptive navigation and positioning, ensuring that the system always relies on more reliable sensor data in the current environment.

[0055] Specifically, the laser dynamic weights are obtained, and the laser dynamic weights satisfy the expression:

[0056]

[0057] In the formula, It is laser dynamic weighting; It is a geometric characteristic index; It is the visual texture index; It is a tiny quantity that prevents the denominator from being zero.

[0058] Furthermore, calculate the visual dynamic weights. The first attitude change was calculated using a single-line lidar odometry system. The second pose change was calculated using visual odometry. The two pose changes are weighted and fused based on dynamic weights to obtain the fused pose value.

[0059] Furthermore, the fused pose quantity satisfies the expression:

[0060]

[0061] In the formula, It is a fusion of pose parameters; It is laser dynamic weighting; It is the first pose change quantity; It is a visual dynamic weight; It is the second pose change.

[0062] In the formula, when the smart locker moves through the mall's corridors, it cannot detect longitudinal features. It gets lower, and the texture of the ground makes it... Maintain a certain level, at which point the expression will automatically decrease. It primarily relies on visual odometry to prevent longitudinal drift; when the smart cabinet faces a glass curtain wall or a dimly lit area, The drop is sudden due to reflective interference or texture loss, while the geometric features of surrounding objects captured by single-line lidar make it possible to... The expression is automatically adjusted to a higher value when the value is relatively high. It relies on laser data to maintain stable positioning.

[0063] Furthermore, based on the fused pose quantity, the global pose of the previous moment is accumulated to obtain the current global pose of the mobile smart cabinet in the global coordinate system. According to the current global pose, the laser point cloud sequence of the current frame is projected onto the global grid map to update the occupancy status of the grid cells in the map, thereby marking the positions of obstacles in the environment in real time. The navigation controller uses the current global pose as the starting point and the preset target waypoint as the ending point to generate a motion path that avoids obstacles in the global grid map using a path planning algorithm (such as the A* algorithm). The motion controller decomposes the motion path into linear velocity commands and angular velocity commands, and sends them to the underlying driver to control the chassis motor of the mobile smart cabinet to rotate, thereby realizing autonomous navigation.

[0064] Thus, a SLAM navigation method for mobile smart cabinets has been realized.

[0065] This invention also discloses a SLAM navigation system for mobile smart cabinets, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a SLAM navigation method for mobile smart cabinets according to the present invention.

[0066] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A SLAM navigation method for mobile smart cabinets, characterized in that, include: Acquire the laser point cloud sequence and grayscale image of the mobile smart cabinet and perform data cleaning; The laser point cloud sequence is divided into multiple sectors, the number of depth transition points in each sector is counted, and the geometric feature index is obtained based on the number of depth transition points. The grayscale image is divided into multiple grid regions, and effective texture grids are selected. The visual texture index is obtained based on the number of effective texture grids and the average gradient magnitude of the whole image. Laser dynamic weights and visual dynamic weights are obtained based on geometric feature indices and visual texture indices, and fused pose values ​​are obtained based on the laser dynamic weights and visual dynamic weights to realize mobile smart cabinet navigation. The geometric characteristic index satisfies the expression: ;in, Indicates geometric characteristic index; Indicates the total number of sectors in the partition; Indicates the sector index number; Indicates the first The number of depth transition points detected within a sector; Represents the characteristic saturation constant; Represents the natural constant.

2. The SLAM navigation method for mobile smart cabinets according to claim 1, characterized in that, The count of depth transition points within each sector includes: Calculate the distance difference between adjacent points in the laser point cloud sequence; if the distance difference between adjacent points exceeds a preset threshold, determine that there is a depth transition point; count the number of depth transition points contained in each sector.

3. The SLAM navigation method for mobile smart cabinets according to claim 1, characterized in that, The process of filtering out valid texture meshes includes: Calculate the gradient magnitude of each pixel within the grid region; if the average gradient magnitude within the grid is greater than the base threshold and the variance of the gradient direction is greater than the direction threshold, the grid is determined to be a valid texture grid.

4. The SLAM navigation method for mobile smart cabinets according to claim 1, characterized in that, The visual texture index satisfies the expression: ； in, Indicates the visual texture index; This indicates the number of texture meshes that were determined to be valid. Indicates the total number of grid cells; This represents the average gradient magnitude of all valid texture meshes in the entire image; Indicates the reference gradient value; Indicates the reflection suppression coefficient; Represents the absolute value symbol.

5. The SLAM navigation method for mobile intelligent cabinets according to claim 1, characterized in that, The laser dynamic weights satisfy the expression: ； in, Indicates the dynamic weight of the laser; Indicates geometric characteristic index; Indicates the visual texture index; It represents a tiny quantity that prevents the denominator from being zero.

6. The SLAM navigation method for mobile smart cabinets according to claim 1, characterized in that, The fused pose quantity satisfies the expression: ； in, Indicates the fused pose quantity; Indicates the dynamic weight of the laser; This represents the change in pose calculated using a single-line lidar odometry. This represents the change in pose calculated using visual odometry.

7. The SLAM navigation method for mobile intelligent cabinets according to claim 1, characterized in that, The data cleaning includes: Distance filtering is applied to the laser point cloud sequence to remove points that are outside the measurement range or too close; noise reduction processing is performed on the grayscale image to remove random noise.

8. A SLAM navigation method for mobile smart cabinets according to claim 1, characterized in that, The characteristic saturation constant is set to 5.

9. A SLAM navigation system for mobile smart cabinets, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a SLAM navigation method for a mobile intelligent cabinet according to any one of claims 1-8.

Images