Public safety patrol method and system based on laser calibration visual obstacle avoidance

Abstract

The invention provides a public safety patrol system and method based on laser calibration visual obstacle avoidance. According to the public safety patrol system and method based on laser calibration visual obstacle avoidance, a non-coding, irregular and dynamically changeable random laser dot matrix is projected to a front environment; acquiring a scene image containing laser dot matrix light spots; inverting the particulate matter concentration and visibility score of the current environment based on a Mie scattering model; fusing the light spot deformation depth and the light spot visual depth to obtain space depth information of the obstacle; fitting the three-dimensional contour of the obstacle according to the space depth information of the obstacle, identifying the material category of the obstacle through a classifier, and generating a collision-free early warning profile and a real-time path based on the distance, contour and material information of the integrated obstacle and the position of the robot; according to the invention, obstacle perception can still be stably realized in scenes of no light, weak light, smog, dust and the like; the method has the advantages of low cost, high robustness, high response speed, high material identification capability and the like, and can be widely applied to public scenes such as mine patrol and underground pipe gallery patrol.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to a public safety patrol method and system based on laser-calibrated visual obstacle avoidance. Background Technology

[0002] In public safety scenarios such as mines, oil and gas pipeline corridors, fire smoke environments, and dark tunnels, patrol equipment typically relies on cameras, infrared depth cameras, or lidar for obstacle detection, but all of these methods have significant shortcomings: 1) Visual sensors suffer from a severe lack of feature points in dark or low-light conditions, relying on environmental texture features for matching and triangulation. In dark, low-light, or low-texture environments (such as smooth walls or monochromatic surfaces), feature points are severely insufficient or even completely absent, leading to a sharp drop in positioning accuracy, trajectory drift, or even complete loss of tracking in visual odometry (VO) or simultaneous localization and mapping (SLAM) systems. 2) Infrared structured light is scattered in dusty and smoky environments, causing severe failure of the depth map. In media such as dust, smog, and water mist, the light spot undergoes severe scattering and absorption, leading to loss and distortion of the coding pattern, large-area holes or noise in the depth map, and a sharp decline in perception reliability. Traditional methods attempt to remove the scattering effect through image processing, but their effectiveness is limited in severely degraded environments. 3) LiDAR is based on the Time-of-Flight (ToF) principle. It can directly acquire high-precision point cloud distance information and is not sensitive to changes in illumination. However, its point cloud is sparse and lacks the surface texture, color and fine geometric semantic information of the target. It has difficulty in identifying obstacles such as glass and thin lines. Moreover, in dense smoke and dust, the laser beam will also be scattered and attenuated, affecting the effective detection range and point cloud quality.

[0003] Therefore, there is an urgent need for an active vision enhancement and obstacle avoidance patrol method and system that can maintain stable and reliable perception capabilities in extreme environments such as no light, low texture, and high scattering. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a public safety patrol method and system based on laser calibration and visual obstacle avoidance. By actively projecting random laser dot arrays, utilizing light spot deformation and scattering characteristics, and fusing multi-source depth information, robust obstacle recognition, material discrimination, and real-time obstacle avoidance can be achieved in extreme environments.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a public safety patrol system based on laser-calibrated visual obstacle avoidance, comprising: The laser calibration module is used to project a non-coded, irregular, and dynamically changing random laser dot pattern into the environment ahead. The visual perception module acquires scene images containing laser dot matrix light spots; The spot recognition and scattering correction module extracts spot features from scene images based on spot detection algorithms, and uses the Mie scattering model to inversely derive the current environment's particulate matter concentration and visibility score. The depth estimation module calculates the deformation depth and visual depth of the light spot based on the light spot features, and adaptively fuses the deformation depth and visual depth of the light spot with the visibility score to obtain the spatial depth information of the obstacle. The obstacle recognition module fits the three-dimensional contour of the obstacle based on its spatial depth information and identifies the material type of the obstacle through a classifier. The collision-free path planning module generates a collision-free warning profile and real-time path based on the distance, contour, and material information of obstacles, as well as the robot's own position. The patrol control module performs obstacle avoidance patrol tasks based on the planned collision-free path, and dynamically adjusts the laser dot density and deep fusion weights for the next cycle according to the obstacle avoidance results and the continuity of environmental changes.

[0006] Preferably, the laser calibration module uses a programmable laser array PLA or a digital micromirror device (DMD) in conjunction with a laser source; and controls the spatial distribution of laser points through a pseudo-random number generator (PRNG); generates a new set of matrix coordinates for each cycle, and dynamically adjusts the matrix density based on ambient light sensor and scattering feedback.

[0007] Preferably, the spot recognition and scattering correction module uses bilateral filtering or nonlocal mean filtering to reduce noise in the scene image containing the laser dot matrix obtained from the visual perception module, preserves the edge structure of the spot, and uses adaptive threshold segmentation or spot detection algorithm based on morphological reconstruction to extract the laser spot region in the scene image. Each laser spot region is marked and its basic features such as pixel coordinates, area, brightness, and shape are extracted.

[0008] Preferably, the light spot characteristics include light spot ellipticity characteristics, light spot energy distribution characteristics, and light spot edge diffusion characteristics.

[0009] Preferably, the depth estimation channel A of the depth estimation module calculates the deformation depth of the laser spot based on the deformation characteristics of the laser spot in the image and in combination with a pre-calibrated lookup table (LUT).

[0010] Preferably, the depth estimation channel B of the depth estimation module calculates the position of the light spot in three-dimensional space using a binocular parallax or multi-frame motion recovery structure algorithm, thereby obtaining a visual depth estimate of the obstacle.

[0011] Preferably, the obstacle recognition module includes a multi-scale point cluster analysis unit and a material discrimination unit; the multi-scale point cluster analysis unit uses the multi-scale DBSCAN algorithm to cluster the light spots, and for each cluster... The 3D contour is fitted using convex hull or Alpha Shape algorithms. The material discrimination unit extracts the sharpness, energy distribution, and diffuse scattering features of the laser spot based on the reflection pattern of the laser spot in the image, and identifies the material category of the obstacle surface through a pre-trained material classification model or a discrimination tree based on physical rules.

[0012] Preferably, the laser calibration module and the visual perception module constitute a rapid obstacle avoidance prediction channel. The laser calibration module samples at a frequency of 1kHz and combines the historical trajectory of the obstacle with the current speed to predict the future obstacle avoidance. The location of obstacles within a given time frame is determined, and when the predicted location enters the warning zone, the visual perception module is triggered to perform a local high-frame-rate scan for confirmation.

[0013] Preferably, the laser dot matrix is ​​used to construct an active feature point set, which is then fused with the natural feature points of the visual SLAM system. This adds a laser feature layer to the ORB-SLAM3 framework, enabling parallel extraction and matching of dual-source features.

[0014] Secondly, the present invention provides a public safety patrol method based on laser-calibrated visual obstacle avoidance, comprising: S1) Based on the initial judgment of the front by the ambient light sensor, project a random laser dot array of initial density forward and simultaneously acquire a scene image containing the laser dot array spots. S2) Extract the scattering features of the light spot from the scene image, and based on the Mie scattering model, inversely derive the current environment's particulate matter concentration and visibility score; S3) Calculate the laser spot deformation depth and spot visual depth using the depth estimation channel respectively, and generate spatial depth information of obstacles by dynamic fusion with adaptive weights; S4) Based on the spatial depth information of the obstacles, perform obstacle contour extraction, material identification, and collision-free path planning; S5) Perform obstacle avoidance patrol tasks based on the planned collision-free path, and dynamically adjust the laser dot density and deep fusion weight for the next cycle according to the obstacle avoidance results and the continuity of environmental changes.

[0015] The beneficial effects of this invention are as follows: 1. This invention provides active feature points through a non-coded random laser dot array, and combines scattering inversion and deformation analysis techniques to still stably achieve obstacle perception in scenarios where traditional sensing systems fail, such as no light, weak light, smoke, and dust. 2. This invention is based on laser spot deformation depth estimation, without relying on structured light coding or LiDAR point clouds, and transforms textureless areas into measurable deformation areas to achieve monocular three-dimensional perception. 3. This invention can not only accurately identify the location and three-dimensional contour of obstacles, but also distinguish multiple types of materials, providing more comprehensive environmental semantic information for patrolling high-risk scenarios; 4. This invention fuses laser feature points with natural feature points from visual SLAM, adding a laser feature layer to the ORB-SLAM3 framework to ensure that SLAM can run continuously for more than 10 minutes in textureless and low-light environments, with positioning drift controlled within 0.3m; 5. The laser fast response mechanism of this invention, combined with the adaptive fusion algorithm, reduces the obstacle avoidance reaction time to 45ms, achieves an obstacle avoidance success rate of 96% in medium and high speed scenarios, and has a complete abnormal response and safety strategy. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system framework of the present invention; Figure 2 This is a schematic flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the random laser dot projection of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1 like Figure 1 As shown, this embodiment provides a public safety patrol system based on laser calibration and visual obstacle avoidance, including: The laser calibration module is used to project a non-coded, irregular, and dynamically changing random laser dot pattern into the environment ahead. In this embodiment, the laser calibration module employs a programmable laser array (PLA) or a digital micromirror device (DMD) in conjunction with a laser source; and a pseudo-random number generator (PRNG) controls the spatial distribution of laser points; a new set of point coordinates is generated each cycle, and the point density is dynamically adjusted based on ambient light sensor and scattering feedback. The generated random laser point array can be found in [reference needed]. Figure 3 As shown.

[0018] The visual perception module acquires scene images containing laser dot matrix spots; In this embodiment, the visual perception module uses a dual-channel imaging camera, with the main channel being an RGB camera and the auxiliary channel being an infrared enhancement camera.

[0019] The spot recognition and scattering correction module extracts spot features from scene images based on spot detection algorithms, and uses the Mie scattering model to inversely derive the current environment's particulate matter concentration and visibility score. In this embodiment, the spot recognition and scattering correction module uses bilateral filtering or nonlocal mean filtering to reduce noise in the scene image containing the laser dot matrix obtained from the visual perception module, preserves the edge structure of the spot, and uses adaptive threshold segmentation or spot detection algorithm based on morphological reconstruction to extract the laser spot region in the scene image. Each laser spot region is marked and its basic features such as pixel coordinates, area, brightness, and shape are extracted.

[0020] In this embodiment, the light spot characteristics include light spot ellipticity characteristics, light spot energy distribution characteristics, and light spot edge diffusion characteristics, as detailed below: Let the region of a single laser spot extracted from the scene image be... Its pixel coordinate set is The corresponding brightness value ; The ellipticity characteristic of the light spot is calculated by fitting the second moment of the light spot, namely:

[0021] in, The center of gravity of the light spot brightness; Indicates the position of the detector The intensity of the scattered light received at that location; They represent direction and The order exponent of moments in a direction; express The order central moment is the intensity-weighted moment with the centroid of the light spot as the origin; based on Constructing the covariance matrix ,Right now:

[0022] Calculate the ellipticity characteristics of the light spot ,Right now:

[0023] in , Covariance matrix eigenvalues.

[0024] The aforementioned light spot energy distribution characteristics reflect the uniformity of brightness diffusion within the light spot, using normalized brightness distribution skewness. and kurtosis describe:

[0025]

[0026] in The standard deviation of brightness, Average brightness; This represents the total number of pixels within the light spot area; For the first The grayscale value of each pixel.

[0027] The aforementioned light spot edge diffusion characteristics reflect the degree of blurring of the light spot, and the edge diffusion coefficient is defined. The radius ratio when the brightness gradient decays to a certain proportion of the center gradient: ; in Brightness decay to radius at time, The equivalent diffusion radius of the light spot is given.

[0028] The equivalent diffusion radius of the light spot The calculation formula is:

[0029] in, Distance from the center of the light spot Average brightness at that location The brightness is the center of the light spot.

[0030] The diffusion coefficient is calculated by fitting the brightness distribution curve at the edge of the light spot. : .

[0031] In this embodiment, the Mie scattering model infers the current particulate matter concentration in the environment by analyzing the diffusion radius and intensity attenuation of the light spot. Specifically: ; In the formula, Indicates the position of the detector The intensity of the scattered light received at that location; Represents the scattering phase function; The brightness of the light spot center; This represents the current concentration of particulate matter in the environment. For distance; Indicates the laser incident angle; This represents the laser wavelength directional scattering function.

[0032] The visibility score and particulate matter concentration mentioned above They are inversely proportional, as shown in the formula below: ; in, This represents the maximum visibility score under ideal, scatter-free conditions. Environmental degradation coefficient, This indicates the concentration of particulate matter.

[0033] The depth estimation module calculates the deformation depth and visual depth of the light spot based on the light spot features, and adaptively fuses the deformation depth and visual depth of the light spot with the visibility score to obtain the spatial depth information of the obstacle. In this embodiment, the depth estimation channel A of the depth estimation module calculates the deformation depth of the laser spot based on the deformation characteristics of the laser spot in the image and in combination with the pre-calibrated lookup table (LUT).

[0034] In this embodiment, the depth estimation channel B of the depth estimation module calculates the position of the light spot in three-dimensional space through binocular parallax or multi-frame motion recovery structure algorithm, thereby obtaining the visual depth estimate of the obstacle.

[0035] In this embodiment, the visual depth calculation of the light spot based on binocular parallax includes three stages: stereo calibration and epipolar correction, light spot feature matching, and depth calculation. During the initialization phase, the intrinsic and extrinsic parameters of the binocular camera are calibrated, and stereo correction is performed on the acquired image pairs to ensure that the corresponding light spots are located on the same horizontal scan line, thus simplifying the stereo matching problem into a one-dimensional search. The spot feature matching uses over-normalized cross-correlation or absolute difference matching criteria to find the most similar spot region and obtain its pixel coordinates. Calculate parallax : ; In the formula, Here are the horizontal pixel coordinates of the light spot in the left image; Here are the horizontal pixel coordinates of the light spot in the right image; Based on the principle of triangle similarity, the visual depth Z of the light spot is calculated as follows:

[0036] in, Indicates the camera's focal length; The baseline length of the binocular camera; This is the calculated disparity.

[0037] In this embodiment, the visual depth of the light spot is calculated using the structure-of-motion algorithm, as follows: In a continuous image sequence In this process, cross-frame tracking is performed on the detected laser spot to establish the correspondence between the same physical spot in different frames; let the spot... In frame and The normalized image coordinates in are respectively and Odometry data is obtained from camera frames via the camera's own odometry or visual odometry. and Relative pose transformation: rotation matrix Translation vector ; Using epipolar geometry constraints, the light spot spatial location satisfy:

[0038] in K This is the camera intrinsic parameter matrix; The light spot can be solved by linear triangulation or nonlinear optimization that minimizes the reprojection error. If the light spot is stably tracked across multiple consecutive frames, the visual depth estimate of the light spot can be obtained by jointly optimizing the camera pose and the 3D coordinates of the light spot across all frames using Bundle Adjustment. .

[0039] The obstacle recognition module fits the three-dimensional contour of the obstacle based on its spatial depth information and identifies the material type of the obstacle through a classifier. In this embodiment, the obstacle recognition module includes a multi-scale point cluster analysis unit and a material discrimination unit; the multi-scale point cluster analysis unit uses the multi-scale DBSCAN algorithm to cluster the light spots, and for each cluster... The 3D contour is fitted using convex hull or Alpha Shape algorithms. The material discrimination unit extracts the sharpness, energy distribution and diffuse scattering features of the laser spot based on the reflection pattern of the laser spot in the image. It identifies the material category of the obstacle surface through a pre-trained material classification model or a discrimination tree based on physical rules, including but not limited to: metal, rock, soft fabric, water surface and human body.

[0040] In this embodiment, the multi-scale point cluster analysis unit uses the multi-scale DBSCAN algorithm to cluster light spots, as detailed below: Let the three-dimensional point set of the light spot obtained from the depth estimation module be... Each point is accompanied by a depth estimation confidence level. First, voxel mesh downsampling is performed:

[0041] in, Let the voxel side length be denoted as 'side length'. This represents the three-dimensional point set of the light spot obtained after voxel mesh downsampling; Represents a voxel grid filter; For each light spot Calculate its up to the th k Distance of neighbors As a local density estimate, The resulting curves are sorted, and a set of candidate neighborhood radii is selected using inflection point detection or the elbow rule. The corresponding minimum number of points parameter is adaptively adjusted based on the local density:

[0042] in For scale The density proportionality coefficient below, This represents the number of point clouds after downsampling. This represents the minimum number of neighborhood points required to identify a core point in the DBSCAN clustering algorithm across j scales.

[0043] For each scale parameter Perform DBSCAN clustering to obtain the cluster set at this scale. .

[0044] Preferably, the convex hull is a minimal convex polyhedron containing a set of points, suitable for obstacles with approximate convex shapes, and is calculated using a fast convex hull algorithm: Input: Cluster point set , Indicates the first Three-dimensional point coordinates; Represents clusters The number of midpoints.

[0045] Output: Set of convex hull vertices Dough sheet collection .

[0046] Convex Hull Volume It can be calculated using the following formula:

[0047] in, It is a tetrahedral facet of a convex hull; , These represent the coordinates of the four vertices of the tetrahedron.

[0048] In this embodiment, Alpha Shape can generate a concave contour, which better fits the shape of non-convex obstacles. Given parameters >0, It is defined as a point set. In the Delaunay triangulation, all radii smaller than The union of the simplexes corresponding to the empty spheres, the specific steps are as follows: Constructing point sets Delaunay triangulation (three-dimensional tetrahedral triangulation); Calculate the minimum circumsphere radius for each simplex. ; like If the circumscribed sphere does not contain any other points, then the simplex is preserved. The boundaries of all retained simplexes constitute the Alpha Shape surface mesh; parameter It can be adaptively determined by the following formula:

[0049] in, The average nearest neighbor distance of the cluster center point. β It is a regulating factor.

[0050] The fitted 3D contour was optimized using the following steps: Laplacian smoothing: Iteratively smoothing the positions of mesh vertices.

[0051] in, As vertices The set of adjacent vertices, where λ is the smoothing factor; Represents vertices In the The new position after the next iteration; Represents vertices An adjacent vertex; Represents vertices In the The position after the next iteration; The edge-folding algorithm is used to reduce the number of facets while keeping the shape error below a threshold. Under the premise of improving the efficiency of subsequent processing.

[0052] Based on the point cloud distribution characteristics of the cluster, an automatic contour fitting algorithm is selected to calculate the cluster's convexity index. :

[0053] in, For volume estimation based on Alpha Shape, if If the shape is positive, use a convex hull; otherwise, use an alpha shape.

[0054] In this embodiment, the material discrimination unit identifies the material type of the obstacle surface, and the steps are as follows: Constructing feature vectors :

[0055] In the formula, This represents the marginal average gradient magnitude. To reflect the consistency of edge gradients; This refers to the brightness deviation. For kurtosis; Energy concentration; Relative diffusion radius; The attenuation coefficient; ; A one-to-one strategy is used to train a multi-class SVM classifier, constructing a classifier for C material categories (metal, rock, soft fabric, water surface, human body). A binary classification SVM sub-model, decision function :

[0056] in For kernel function, and For the first Class and the Hyperplane parameters between classes; The final category will be decided by voting:

[0057] In the formula, Indicates the final predicted material category; Indicates an indicator function; Indicates category label; This indicates that in a certain one-to-one SVM sub-model For eigenvectors The winning category during classification; In this embodiment, a random forest classifier can be trained, which determines the category through voting among multiple decision trees. Each tree is trained on a random subset of features to enhance generalization ability, and the output category probability is:

[0058] in, For the number of trees, For the first The predicted category of each tree.

[0059] The collision-free path planning module generates a collision-free warning profile and real-time path based on the distance, contour, and material information of obstacles, as well as the robot's own position. The patrol control module performs obstacle avoidance patrol tasks based on the planned collision-free path, and dynamically adjusts the laser dot density and deep fusion weights for the next cycle according to the obstacle avoidance results and the continuity of environmental changes.

[0060] In this embodiment, the laser calibration module and the visual perception module constitute a rapid obstacle avoidance prediction channel. The laser calibration module samples at a frequency of 1kHz and combines the historical trajectory of the obstacle with the current speed to predict the future obstacle avoidance. The location of the obstacle within a given time period, i.e.:

[0061] In the formula, This represents the predicted position of the obstacle at the future time t+Δt; This indicates the measured position of the obstacle at the current time t; This represents the acceleration of the obstacle at the current time t; When the predicted location enters the warning area, the visual perception module is triggered to perform a local high frame rate scan for confirmation.

[0062] In this embodiment, the laser dot matrix is ​​used to construct an active feature point set, which is fused with the natural feature points of the visual SLAM system. A new laser feature layer is added to the ORB-SLAM3 framework to achieve parallel extraction and matching of dual-source features.

[0063] In this embodiment, the patrol control module has a built-in exception response mechanism: If no valid light spot is detected for three consecutive frames, the system automatically switches to pure visual SLAM mode and triggers an audible and visual alarm; if there is no response to the wheel speed command for 2 seconds and the IMU detects an angular acceleration of <0.05 rad / s², the system will automatically switch to pure visual SLAM mode and trigger an audible and visual alarm. 2 At that time, perform the following escape maneuver: reverse 0.3m → turn left 45° → move forward 0.5m.

[0064] Example 2 like Figure 2 As shown, this embodiment provides a public safety patrol method based on laser calibration visual obstacle avoidance, including: S1) Based on the initial judgment of the front by the ambient light sensor, project a random laser dot array of initial density forward and simultaneously acquire a scene image containing the laser dot array spots. S2) Extract the scattering features of the light spot from the scene image, and based on the Mie scattering model, inversely derive the current environment's particulate matter concentration and visibility score; S3) Calculate the laser spot deformation depth and spot visual depth using the depth estimation channel respectively, and generate spatial depth information of obstacles by dynamic fusion with adaptive weights; S4) Based on the spatial depth information of the obstacles, perform obstacle contour extraction, material identification, and collision-free path planning; S5) Perform obstacle avoidance patrol tasks based on the planned collision-free path, and dynamically adjust the laser dot density and deep fusion weight for the next cycle according to the obstacle avoidance results and the continuity of environmental changes.

[0065] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A public safety patrol system based on laser-calibrated visual obstacle avoidance, characterized in that, include: The laser calibration module is used to project a non-coded, irregular, and dynamically changing random laser dot pattern into the environment ahead. The visual perception module acquires scene images containing laser dot matrix spots; The spot recognition and scattering correction module extracts spot features from scene images based on spot detection algorithms, and uses the Mie scattering model to inversely derive the current environment's particulate matter concentration and visibility score. The depth estimation module calculates the deformation depth and visual depth of the light spot based on the light spot features, and adaptively fuses the deformation depth and visual depth of the light spot with the visibility score to obtain the spatial depth information of the obstacle. The obstacle recognition module fits the three-dimensional contour of the obstacle based on its spatial depth information and identifies the material type of the obstacle through a classifier. The collision-free path planning module generates a collision-free warning profile and real-time path based on the distance, contour, and material information of obstacles, as well as the robot's own position. The patrol control module performs obstacle avoidance patrol tasks based on the planned collision-free path, and dynamically adjusts the laser dot density and deep fusion weights for the next cycle according to the obstacle avoidance results and the continuity of environmental changes.

2. A public safety patrol system based on laser calibration and visual obstacle avoidance according to claim 1, characterized in that: The laser calibration module uses a programmable laser array PLA or a digital micromirror device (DMD) in conjunction with a laser source; and controls the spatial distribution of laser points through a pseudo-random number generator (PRNG); generates a new set of matrix coordinates for each cycle, and dynamically adjusts the matrix density based on ambient light sensor and scattering feedback.

3. A public safety patrol system based on laser calibration and visual obstacle avoidance according to claim 1, characterized in that: The light spot recognition and scattering correction module uses bilateral filtering or nonlocal mean filtering to reduce noise in the scene image containing the laser dot matrix obtained from the visual perception module, preserves the edge structure of the light spot, and uses adaptive threshold segmentation or morphological reconstruction-based light spot detection algorithm to extract the laser light spot region in the scene image. Each laser light spot region is marked and its basic features such as pixel coordinates, area, brightness, and shape are extracted.

4. A public safety patrol system based on laser calibration visual obstacle avoidance according to claim 3, characterized in that: The light spot characteristics include light spot ellipticity characteristics, light spot energy distribution characteristics, and light spot edge diffusion characteristics.

5. A public safety patrol system based on laser calibration and visual obstacle avoidance according to claim 1, characterized in that: The depth estimation channel A of the depth estimation module calculates the deformation depth of the laser spot based on the deformation characteristics of the laser spot in the image and in combination with a pre-calibrated lookup table (LUT).

6. A public safety patrol system based on laser calibration visual obstacle avoidance according to claim 5, characterized in that: The depth estimation channel B of the depth estimation module calculates the position of the light spot in three-dimensional space through a binocular parallax or multi-frame motion recovery structure algorithm, thereby obtaining the visual depth estimate of the obstacle.

7. A public safety patrol system based on laser calibration visual obstacle avoidance according to claim 1, characterized in that: The obstacle recognition module includes a multi-scale point cluster analysis unit and a material discrimination unit; the multi-scale point cluster analysis unit uses the multi-scale DBSCAN algorithm to cluster light spots, and for each cluster... The 3D contour is fitted using convex hull or AlphaShape algorithms. The material discrimination unit extracts the sharpness, energy distribution, and diffuse scattering features of the laser spot based on the reflection pattern of the laser spot in the image, and identifies the material category of the obstacle surface through a pre-trained material classification model or a discrimination tree based on physical rules.

8. A public safety patrol system based on laser calibration visual obstacle avoidance according to claim 1, characterized in that: The laser calibration module and the visual perception module constitute a rapid obstacle avoidance prediction channel. The laser calibration module samples at a frequency of 1kHz and combines the historical trajectory of the obstacle with the current speed to predict the future obstacle avoidance. The location of obstacles within a given time frame is determined, and when the predicted location enters the warning zone, the visual perception module is triggered to perform a local high-frame-rate scan for confirmation.

9. A public safety patrol system based on laser calibration visual obstacle avoidance according to claim 1, characterized in that: The laser dot matrix is ​​used to construct an active feature point set, which is then fused with the natural feature points of the visual SLAM system. A new laser feature layer is added to the ORB-SLAM3 framework to achieve parallel extraction and matching of dual-source features.

10. A public safety patrol method based on laser-calibrated visual obstacle avoidance, characterized in that, include: S1) Based on the initial judgment of the front by the ambient light sensor, project a random laser dot array of initial density forward and simultaneously acquire a scene image containing the laser dot array spots. S2) Extract the scattering features of the light spot from the scene image, and based on the Mie scattering model, inversely derive the current environment's particulate matter concentration and visibility score; S3) Calculate the laser spot deformation depth and spot visual depth using the depth estimation channel respectively, and generate spatial depth information of obstacles by dynamic fusion with adaptive weights; S4) Based on the distance, contour, and material information of the obstacles and the robot's own position, generate a collision-free warning profile and real-time path; S5) Perform obstacle avoidance patrol tasks based on the planned collision-free path, and dynamically adjust the laser dot density and deep fusion weight for the next cycle according to the obstacle avoidance results and the continuity of environmental changes.

Images