Unmanned vehicle local high-precision positioning mapping system and method based on reinforcement learning polarization normal deambiguity and depth information fusion

Abstract

The invention provides an unmanned vehicle local high-precision positioning mapping system and method based on reinforcement learning polarization normal deambiguity and depth information fusion, and the method comprises the steps: synchronously obtaining a polarization image and a depth image of a target scene in the driving process of an unmanned vehicle in a low-texture and low-illumination scene; constructing the polarization normal candidate, the depth normal, the neighborhood surface curvature and the depth gradient information into a state vector, inputting the state vector into a reinforcement learning agent based on Dueling DQN, and obtaining the depth gradient according to a reward function fusing the included angle error of the polarization normal and the depth normal, the neighborhood normal smoothness loss and the consistency error of the polarization normal and the depth gradient direction. Outputting an optimal polarization normal selection action to obtain an unambiguous surface normal; polarization estimation depths are generated, and confidence coefficients are calculated respectively; according to the depth confidence coefficient and the polarization confidence coefficient, pixel-level weighted fusion is carried out, a fused depth map is obtained, synchronous positioning and mapping are carried out on the fused depth map and the synchronous RGB image, and a point cloud map of the target scene and the moving track of the unmanned vehicle are output.

Description

Technical Field

[0001] This invention belongs to the field of unmanned vehicle localization and mapping technology, specifically involving an unmanned vehicle local high-precision localization and mapping system and method based on reinforcement learning polarization normal deambiguation and depth information fusion. Background Technology

[0002] In the current field of autonomous driving technology, vision-based simultaneous localization and mapping (SLAM) systems mainly rely on visual sensors such as monocular, binocular, or RGB-D cameras to achieve environmental perception and autonomous navigation. These technologies construct environmental maps and estimate vehicle poses through feature point extraction, matching, and optimization algorithms, achieving significant results in scenes with rich textures and uniform illumination. However, their performance is highly dependent on the texture features of the environmental surface and lighting conditions. In low-texture (e.g., white walls, solid-color ground) or low-light (e.g., nighttime, tunnels) scenes, problems such as feature matching failures and sparse point clouds easily occur, leading to decreased map accuracy or even tracking loss.

[0003] To overcome the aforementioned limitations, polarization-based 3D imaging technology has been introduced into SLAM systems as a physical optical sensing method. Polarization information can capture the anisotropic characteristics of light wave vibration direction. By analyzing the polarization state (such as degree of polarization and polarization angle) of reflected light from an object's surface, the surface normal direction can be inverted, thereby reconstructing the object's 3D shape. Since polarization characteristics are related to the surface's micro-geometry, it has a natural advantage in sensing low-texture targets and is less affected by changes in ambient lighting. However, polarization-based 3D imaging has two inherent limitations: first, the azimuth ambiguity problem, i.e., there is a 180° uncertainty when calculating the normal direction based on the polarization angle, requiring prior assumptions or external constraints to resolve the ambiguity; second, polarization information can only provide relative surface orientation (normal), and cannot directly obtain absolute depth values, making it difficult to meet the requirements of scale consistency and spatial accuracy in SLAM.

[0004] In existing technologies, researchers have attempted to fuse polarization imaging with depth sensors (such as TOF and structured light) to leverage their complementary advantages. However, most fusion schemes remain at the level of simple data overlay, failing to fundamentally address the constraint of polarization normal ambiguity on fusion quality. For example, some schemes employ heuristic rules (such as selecting the candidate polarization normal with the smallest angle to the depth normal) or local smoothing assumptions for deambiguation. These methods lack stability under complex surfaces or noise interference, easily introducing accumulated errors. Furthermore, traditional fusion strategies are mostly based on fixed weights or empirical thresholds, making it difficult to adaptively adjust the reliability weights of different sensors in different scenarios, resulting in poor fusion performance in low signal-to-noise ratio regions.

[0005] In summary, existing visual SLAM technologies face challenges in perceptual capabilities under low-texture and low-light conditions, while the fusion of polarization and depth information is limited by the robustness and adaptability of polarization normal ambiguity resolution methods. Overcoming the limitations of traditional deambiguation strategies and achieving intelligent collaborative fusion of polarization and depth information under geometric constraints has become a key bottleneck in improving the localization and mapping accuracy of autonomous vehicles in extreme environments. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, this invention provides a method and system for unmanned vehicle localization and mapping based on reinforcement learning-based polarization normal deambiguation and depth information fusion. This method simultaneously acquires polarization and depth images in low-texture, low-light scenarios. It utilizes a reinforcement learning model based on the Dueling DQN architecture to adaptively deambiguously process the polarization normals. The state vector of the reinforcement learning model integrates polarization normal candidates, depth normals, neighborhood surface curvature, and depth gradient information. By comprehensively considering the reward mechanism of normal angle error, surface smoothness, and depth consistency error, it intelligently selects the optimal polarization normal direction, thereby resolving the azimuth ambiguity problem in polarization imaging. Simultaneously, this invention designs an adaptive fusion strategy based on confidence evaluation. By calculating depth confidence and polarization confidence separately, it performs pixel-level weighted fusion of the depth image and polarization-estimated depth, effectively combining the advantages of two modalities of data and improving the accuracy and completeness of depth information. Finally, based on the fused high-precision depth information and RGB image, simultaneous localization and mapping are performed, significantly improving the localization accuracy and mapping quality of unmanned vehicles in low-texture, low-light scenarios. The system of this invention includes an unmanned vehicle, a depth camera, a polarization camera, a data acquisition card, a computing card, and a display. The surface normal deambiguation module, the image fusion module, and the localization and mapping module are deployed in the computing card, which work together to realize the entire process from data acquisition to high-precision localization and mapping, providing reliable technical support for the autonomous navigation of unmanned vehicles in complex environments.

[0007] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0008] A method for high-precision local localization and mapping of unmanned vehicles based on reinforcement learning polarization normal deambiguation and depth information fusion includes:

[0009] In low-texture, low-light scenes, simultaneously acquire polarization and depth images of the target scene during the autonomous vehicle's operation;

[0010] Based on the polarization image, multiple polarization normal candidates for each pixel are extracted, and based on the depth image, depth normal, neighborhood surface curvature, and depth gradient information are extracted.

[0011] The polarization normal candidate, depth normal, neighborhood surface curvature and depth gradient information are constructed into a state vector and input into the reinforcement learning agent based on Dueling DQN. Based on the reward function that integrates the angle error between the polarization normal and the depth normal, the neighborhood normal smoothing loss and the polarization normal and the depth gradient direction consistency error, the optimal polarization normal selection action is output to obtain the unambiguous surface normal.

[0012] The polarization-estimated depth is generated based on the unambiguous surface normal, and the depth confidence and polarization confidence are calculated respectively. The depth confidence is obtained based on the reflection intensity and depth gradient magnitude of the depth image, and the polarization confidence is obtained based on the polarization degree of the polarization image and the residual angle of the polarization model fitting.

[0013] Based on the depth confidence and polarization confidence, the depth values ​​of the depth image and the polarization estimated depth are fused at the pixel level to obtain a fused depth map;

[0014] Based on the fused depth map and the synchronized RGB image, synchronous localization and mapping are performed, and the point cloud map of the target scene and the running trajectory of the unmanned vehicle are output.

[0015] The synchronously acquired polarization and depth images are transmitted in parallel through multi-threaded control, and the frame rates of the polarization and depth cameras are adjusted synchronously using a continuous soft-trigger method. The polarization camera, depth camera, acquisition card for deploying image acquisition algorithms, and computing card for deploying core algorithms are all adapted to the image acquisition and data processing requirements of low-texture, low-light scenes.

[0016] Furthermore, the state vector construction, decision-making, and training process of the reinforcement learning agent includes: forming a state vector from polarization normal candidates, depth normals, neighborhood surface curvature, and depth gradient information; calculating action value through a Dueling DQN architecture containing value branches and advantage branches; outputting the optimal choice based on a negative weighted summation reward function that integrates the angle error between the polarization normal and the depth normal, the neighborhood normal smoothing loss, and the consistency error between the polarization normal and the depth gradient direction; iteratively training until policy convergence based on historical data in the offline stage; and directly inputting the real-time state vector in the online stage to obtain unambiguous surface normals.

[0017] Furthermore, the polarization confidence is evaluated by combining the pixel polarization degree with the residual angle of the polarization model fitting, and the depth confidence is evaluated by combining the depth camera reflection intensity with the gradient magnitude of the depth map. Both are normalized to characterize the reliability of the information source. During fusion, the depth value of the depth image and the polarization estimated depth are dynamically weighted according to the dual confidence, and the fused depth map is obtained by weighted averaging. A small compensation amount is introduced during the fusion process to avoid calculation anomalies.

[0018] Furthermore, the execution of synchronous localization and mapping specifically includes: filtering key frames through tracking threads, optimizing the map and removing redundant frames through the local mapping module, identifying similar historical scenes using the bag-of-words model to achieve loop closure detection, and then correcting the pose of key frames and the position of map points through global optimization to eliminate accumulated errors and scale drift, thereby ensuring the geometric consistency of the point cloud map.

[0019] Furthermore, a high-precision local localization and mapping system for unmanned vehicles based on reinforcement learning-based polarization normal deambiguity and depth information fusion includes: an unmanned vehicle; a polarization camera and a depth camera mounted on the unmanned vehicle for simultaneously acquiring polarization and depth images of the target scene in low-texture, low-light scenarios; a data processing unit connected to the polarization camera and the depth camera, configured to: extract polarization normal candidates, depth normals, neighborhood surface curvature, and depth gradient information to construct a state vector; output unambiguous surface normals through reinforcement learning logic based on Dueling DQN; calculate depth confidence and polarization confidence; perform pixel-level weighted fusion of depth values ​​and polarization estimated depths; perform synchronous localization and mapping based on the fused depth map and the synchronized RGB image; and a display unit connected to the data processing unit for outputting a point cloud map and the unmanned vehicle's trajectory.

[0020] Furthermore, the polarization camera is a polarization camera adapted to low-texture, low-light scenes, and the depth camera is a depth camera adapted to the scene; the data processing unit includes an acquisition module, a deambiguation module, a fusion module, and a mapping module, and each module works together to complete the entire process from image acquisition to localization and mapping.

[0021] Furthermore, the unmanned vehicle positioning and mapping system, based on the above method, includes a polarization camera and a depth camera mounted on the unmanned vehicle, a data acquisition card connected to the polarization camera and the depth camera, a computing card connected to the data acquisition card, and a display connected to the computing card.

[0022] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0023] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0024] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0025] First, this invention effectively overcomes the performance bottleneck of a single visual sensor in low-texture, low-light scenarios, solving problems such as feature extraction difficulties and pose estimation drift caused by lack of texture and insufficient lighting in traditional visual SLAM. Through the collaborative acquisition of data by a polarization camera and a depth camera, combined with hardware selection and a synchronization control mechanism adapted to the scenario, this invention can stably acquire effective environmental information in environments lacking obvious texture and with harsh lighting conditions. This provides high-quality data support for subsequent localization and mapping, improving the environmental adaptability and operational reliability of autonomous vehicles in complex real-world scenarios.

[0026] Secondly, this study specifically addresses the inherent azimuth ambiguity and absolute depth deficiency issues in polarization-based 3D imaging technology. By introducing a reinforcement learning agent based on the Dueling DQN architecture, and combining geometric features extracted from depth images (depth normals, neighborhood surface curvature, and depth gradients) to adaptively filter polarization normal candidates, the 180° uncertainty in the normal direction can be automatically eliminated without relying on manually designed prior constraints, ensuring the accuracy of surface geometric information. Simultaneously, through a weighted fusion strategy of depth confidence and polarization confidence, the surface topography information from polarization imaging is organically combined with the absolute depth information from the depth camera. This retains the geometric representation advantages of polarization technology in low-texture scenes while compensating for its inability to obtain absolute depth, providing more comprehensive and accurate 3D environmental information for subsequent localization and mapping.

[0027] Furthermore, it improves the accuracy and completeness of depth information, laying the foundation for high-precision output of localization mapping. This invention quantifies the reliability of depth images and polarization-estimated depth separately through confidence assessment, and dynamically allocates fusion weights based on the confidence level to achieve pixel-level adaptive fusion: in areas with reliable depth information (such as high reflectivity and low gradient areas), depth camera data is emphasized, while in areas with better polarization information (such as low texture and weak illumination areas), the contribution of polarization-estimated depth is strengthened. This effectively reduces the impact of noise or error from a single data source on the results, ensuring that the fused depth map has both absolute depth accuracy and maintains the smoothness and realism of surface morphology.

[0028] Finally, the overall accuracy and geometric consistency of the autonomous vehicle's localization and mapping were ensured. The localization and mapping module, through the collaborative work of keyframe selection via tracking threads, map optimization via local mapping modules, similar scene identification via loop closure detection modules, and global error correction via global optimization, effectively eliminates accumulated errors and scale drift during long-term mapping processes. This ensures consistent geometric structure of the generated point cloud map, while simultaneously outputting more accurate autonomous vehicle trajectories. This avoids localized error accumulation leading to positioning loss or mapping deviations, providing more stable technical support for core functions such as autonomous vehicle navigation and obstacle avoidance. Attached Figure Description

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0030] Figure 1 This is a system framework diagram of an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the image acquisition algorithm flow according to an embodiment of the present invention;

[0032] Figure 3 These are scene point cloud maps and unmanned vehicle operation trajectory maps according to embodiments of the present invention. Detailed Implementation

[0033] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0034] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0036] This invention proposes a high-precision local localization and mapping system and method for unmanned vehicles based on the fusion of polarization images, depth information, and reinforcement learning. The system is implemented using an unmanned vehicle, a depth camera, a polarization camera, a data acquisition card, a computing card, and a display. The depth camera and polarization camera are mounted on the unmanned vehicle, and the data acquisition card connects them. The image acquisition algorithm is deployed in the data acquisition card. The core modules of the invention—the surface normal deambiguation module, the image fusion module, and the localization and mapping module—are implemented on the computing card.

[0037] Its core functionality for achieving high-precision local localization and mapping of unmanned vehicles based on reinforcement learning-based polarization normal deambiguation and depth information fusion includes: an image acquisition module, a surface normal deambiguation module, an image fusion module, and a localization and mapping module. Building upon the traditional polarization and depth information fusion framework, a Dueling DQN agent is introduced to dynamically decide between polarization normal candidates and depth normals. A reward function is used to feed back the error in the normal angle and the surface consistency error, achieving adaptive deambiguation of polarization normals. In the image fusion module, the depth map and the polarization-estimated depth map are weighted and fused based on confidence levels, improving the realism and smoothness of the 3D reconstructed surface of the target object and reducing the trajectory error of the unmanned vehicle in local low-texture, low-light scenes.

[0038] The method employs reinforcement learning for adaptive deambiguation, where the agent automatically selects the optimal normal direction to obtain deambigued surface normals, thus resolving the ambiguity of azimuth angles. Furthermore, it utilizes an adaptive weighted fusion strategy that combines depth maps with polarization-estimated depths recovered from deambigued polarization normals. This effectively improves the accuracy and completeness of depth information, thereby significantly enhancing the localization and mapping accuracy of autonomous vehicles in low-texture and low-light scenarios.

[0039] like Figure 1 As shown, in the implementation of this embodiment of the invention, the unmanned vehicle is equipped with a computing card, a data acquisition card, a polarization camera, and a depth camera to collect environmental information around the scene;

[0040] The image acquisition module includes an acquisition card, a depth camera, a polarization camera, a multi-threaded parallel acquisition module, and a frame synchronization control module. The polarization camera and depth camera are used to acquire polarization and depth images of the surrounding scene. The acquisition card is connected to the polarization camera and depth camera and is used to deploy image acquisition algorithms. The multi-threaded parallel acquisition module is used to control the operation of the polarization camera and depth camera. The frame synchronization control module is used to control the polarization camera and depth camera to achieve synchronous acquisition.

[0041] The surface normal deambiguation module is connected to the image acquisition module. Through reinforcement learning, it adaptively deambiguates the surface normals, neighborhood surface curvature, and gradient information extracted from the depth image. It performs feature state analysis and action decision-making on the polarization normal candidates acquired by the polarization camera, so that the agent can automatically select the optimal normal direction and obtain the deambigued surface normal result.

[0042] The image fusion module is connected to the surface normal deambiguation module. It characterizes the reliability of each pixel information source by calculating the depth confidence and polarization confidence. In the fusion unit, the depth map and the polarization estimated depth map are weighted and averaged according to the confidence to obtain the pixel-level adaptive fusion result.

[0043] The localization and mapping module is connected to the image fusion module. The localization and mapping module includes a tracking thread module, a local mapping module, a loop closure detection module, map building and location recognition. It uses high-precision depth images and RGB images as input to build a point cloud map of a local low-texture, low-light scene and to locate the unmanned vehicle.

[0044] The computing card is used to deploy the surface normal deambiguation module, image fusion module, and localization and mapping module, and is responsible for performing the corresponding computing tasks; the display is used to present the point cloud map of the local low-texture and low-light scene and the running trajectory of the unmanned vehicle.

[0045] In a preferred embodiment, the polarization camera and the depth camera are fixedly connected and mounted on the unmanned vehicle; the acquisition card connects the polarization camera and the depth camera and deploys image acquisition algorithms to control the camera to acquire images; the surface normal deambiguation module, the image fusion module, and the localization and mapping module are deployed on the computing card and are responsible for performing the corresponding calculation tasks; the display is connected to the computing card and is used to display the point cloud map of the local low-texture low-light scene and the running trajectory of the unmanned vehicle.

[0046] In a preferred embodiment, the image acquisition module includes a polarization camera and a depth camera module, a multi-threaded parallel acquisition module, an acquisition card, and a frame synchronization control module. The acquisition card is connected to the polarization camera and depth camera module, the multi-threaded parallel acquisition module, and the frame synchronization control module, respectively. The multi-threaded parallel acquisition module controls the transmission of polarization and depth images in a multi-threaded parallel manner. The frame synchronization control module sends signals to the depth camera and polarization camera using a continuous soft-trigger method, while controlling parameters such as the frame rate. The multi-threaded parallel acquisition module and the frame synchronization control module are used together to obtain synchronized polarization and depth images.

[0047] In a preferred embodiment, the surface normal deambiguation module establishes a reinforcement learning model with polarization normals, depth normals, neighborhood surface curvature, and gradient information as state vectors. It iteratively updates the normal selection strategy based on the reward function, thereby achieving adaptive deambiguation of surface normals. The implementation of this module's function specifically includes the following steps:

[0048] Step A1: For each pixel, extract the candidate polarization normal direction vector, depth normal direction vector, neighborhood surface curvature, and depth gradient information from the polarization image and depth image to construct a state vector;

[0049] Step A2: Based on the Dueling DQN reinforcement learning network, iterative training is performed in the offline training stage according to the state vector, the selected action and the reward function composed of normal angle error, smoothness error and depth consistency error to obtain a convergent deambiguation policy;

[0050] Step A3: During the online operation phase, the state vector of the current pixel is input into the trained DuelingDQN network, which outputs the optimal polarization normal selection action and obtains the unambiguous surface normal of the pixel accordingly.

[0051] As a preferred embodiment, the state vector of the reinforcement learning agent is defined as follows: ,in Indicates candidate 1 for polarization normal; Indicates candidate 2 for polarization normal; Represents the curvature of the neighborhood surface; Representing deep gradient information, the agent adopts a Dueling DQN network structure. Based on the state vector, it calculates the action value function Q(s,a) through the value branch V(s) and the advantage branch A(s,a) respectively, and outputs the polarization normal to select the action.

[0052] To achieve adaptive deambiguity of polarization normals, the reward function of the reinforcement learning agent is designed as a multinomial reward function that comprehensively considers normal direction accuracy, surface geometric continuity, and depth consistency, and its definition is as follows:

[0053]

[0054] In the formula, Let be the reward value at time t. For weight parameters, This refers to the angle error between the polarization normal and the depth normal. The smoothing loss is applied to the neighborhood normal. This represents the consistency error between the polarization normal and the depth gradient direction.

[0055]

[0056] In the formula, Represents the polarization normal vector. The angle error, which is the depth normal vector, reflects the consistency of the direction between the deambiguous normal and the depth geometric normal.

[0057]

[0058] In the formula, Let p be the normal vector of the current pixel. Its neighborhood The normals of adjacent pixels within.

[0059] .

[0060] In a preferred embodiment, the image fusion module includes a confidence estimation unit and an adaptive weighted fusion unit. The confidence estimation unit includes a polarization confidence estimation unit and a depth confidence estimation unit. The polarization confidence unit takes the polarization degree calculated from the multi-angle polarization image and the residual angle of the polarization model fitting as input, and calculates the polarization confidence according to the following formula:

[0061]

[0062] In the formula, The polarization degree at pixel x. This is the preset polarization normalization coefficient. To fit the residual angle to the polarization model, For weight parameters, This is the Sigmoid function.

[0063] The depth confidence estimation unit takes depth reflection intensity and depth gradient as inputs and calculates the depth confidence according to the following formula:

[0064]

[0065] In the formula, Let x be the reflection intensity of the depth camera at pixel x. This is the normalization coefficient for the preset reflection intensity. This represents the gradient magnitude of the depth map at pixel x. These are the weight parameters.

[0066] The adaptive weighted fusion unit performs pixel-level adaptive weighted fusion of the depth map and the polarization-estimated depth based on polarization confidence and depth confidence to obtain a fused depth map. The adaptive weighted fusion unit fuses the depth according to the following formula:

[0067]

[0068] In the formula, This represents the coarse depth value of the depth camera at pixel x. The depth is estimated by polarization obtained from the gradient integration of the unambiguous polarization normal. To prevent tiny constants with a denominator of zero.

[0069] In a preferred embodiment, the localization and mapping module includes a tracking thread module, a local mapping module, a loop closure detection module, map construction, and location recognition. The tracking thread module estimates the camera pose of each frame and determines whether a new image frame can be used as a new keyframe through feature matching. The local mapping module incrementally processes new keyframes, optimizes the local map, and removes redundant keyframes. The loop closure detection module identifies historical similar scenes using a bag-of-words model, calculates the similarity transformation matrix between closed loop frames, and corrects accumulated errors. After detecting a loop, the global optimization module performs a global optimization process to correct the poses of all keyframes and the positions of map points, eliminating accumulated errors and correcting scale drift, thereby achieving geometric consistency of the entire map.

[0070] Based on the above systematic implementation scheme, the present invention also provides a method for high-precision local localization and mapping of unmanned vehicles based on reinforcement learning polarization normal deambiguation and depth information fusion, including the following steps:

[0071] Step S1: Acquire synchronous polarization images and depth images of local low-texture, low-light scenes during the autonomous vehicle's operation using the image acquisition module;

[0072] Step S2: Calculate the zenith angle and azimuth angle of the polarization normal candidate for each pixel based on the polarization map, estimate the depth normal and depth gradient information based on the depth image, and extract the surface curvature features of the pixel neighborhood.

[0073] Step S3: Input the information obtained in step S2 into a state vector and input it into the agent in the surface normal deambiguation module. Output the polarization normal selection action according to the converged deambiguation strategy to obtain the unambiguous surface normal set of the corresponding pixel.

[0074] Step S4: In the image fusion module, the polarization estimated depth is calculated based on the deambiguous polarization normal. The confidence unit evaluates the polarization confidence and depth confidence respectively. The depth map and the polarization estimated depth are then weighted and fused by the adaptive weighted fusion unit to obtain the fused depth map.

[0075] Step S5: Input the high-precision depth image and the synchronized RGB image obtained in step S4 into the localization and mapping module, and finally output the point cloud map of the local low-texture and low-light scene and the trajectory of the unmanned vehicle.

[0076] As a preferred embodiment, the image acquisition module includes a polarization camera and a depth camera module, a multi-threaded parallel acquisition module, an acquisition card, and a frame synchronization control module. The polarization camera is a Flir BFS-U3-51S5P model, with a maximum resolution of 2448×2048 and a maximum adjustable frame rate of 75 FPS; the depth camera is a Kinect v2 based on the TOF principle, with a maximum resolution of 1920×1080 and a maximum adjustable frame rate of 30 FPS; the acquisition card is a Jetson TX2.

[0077] The acquisition card is connected to the polarization camera and depth camera modules, the multi-threaded parallel acquisition module, and the frame synchronization control module, respectively. The multi-threaded parallel acquisition module controls the transmission of polarization and depth images in a multi-threaded parallel manner. The frame synchronization control module sends signals to the depth camera and polarization camera using a continuous soft-trigger method, while controlling parameters such as the frame rate. The multi-threaded parallel acquisition module and the frame synchronization control module are used together to obtain synchronized polarization and depth images, providing the necessary conditions for subsequent modules.

[0078] In a preferred embodiment, the polarization camera and depth camera are fixedly connected and mounted on the unmanned vehicle, while the acquisition card is wired to both the polarization camera and depth camera. A schematic diagram of the image acquisition algorithm flow is shown below. Figure 2 As shown, the specific steps include:

[0079] Step B1: Configure the camera equipment and set the acquisition parameters for each camera to ensure the quality and consistency of the acquired images;

[0080] Step B2: Run the unmanned vehicle, polarization camera, and depth camera to acquire environmental images of the surrounding scene;

[0081] Step B3: Use the rosbag tool in ROS tools to simultaneously record topic data from both the polarization camera and the depth camera;

[0082] Step B4: Extract image data from the recorded rosbag file and construct an image dataset;

[0083] Step B5: Use the rqt tool in ROS to view the images in real time to ensure that the data acquisition is normal.

[0084] Figure 3 The image shows the scene point cloud map and the unmanned vehicle's running trajectory map of the test example of this invention. The left image is a top view of the sparse point cloud of the scene, showing that the unmanned vehicle acquired rich scene point cloud information during its movement. The right image represents the running trajectory of the unmanned vehicle. As the unmanned vehicle returned to the origin along the loop line, the running trajectory obtained by the localization and mapping algorithm also formed a closed loop.

[0085] Compared with existing technologies, this invention provides a new research scheme for localization and mapping of unmanned vehicles using multi-sensor fusion, and effectively solves the problems of poor mapping effect and inaccurate localization of single sensors under low texture and low light conditions.

[0086] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0087] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0090] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive other forms of a local high-precision local positioning and mapping system and method for unmanned vehicles based on reinforcement learning polarization normal deambiguation and depth information fusion. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. An unmanned vehicle local high-precision positioning and mapping method based on reinforcement learning polarization normal disambiguation and depth information fusion, characterized in that, The application relates to a method for generating a point cloud map of a target scene and a trajectory of an unmanned vehicle. In a low-texture and low-light scene, a polarization image and a depth image of a target scene during driving of an unmanned vehicle are synchronously acquired; A plurality of polarization normal candidates of each pixel are extracted based on the polarization image, and a depth normal, a neighborhood surface curvature and depth gradient information are extracted based on the depth image; The polarization normal candidates, the depth normal, the neighborhood surface curvature and the depth gradient information are constructed into a state vector, input into a reinforcement learning agent based on a Dueling DQN, and an optimal polarization normal selection action is output according to a reward function of an angle error of fusion of the polarization normal and the depth normal, a neighborhood normal smoothness loss and a consistency error of the polarization normal and the depth gradient direction, so that an unambiguous surface normal is obtained; Polarization estimation depth is generated based on the unambiguous surface normal, and a depth confidence and a polarization confidence are respectively calculated; the depth confidence is obtained based on a reflection intensity and a depth gradient amplitude of the depth image, and the polarization confidence is obtained based on a polarization degree and a polarization model fitting residual angle of the polarization image; According to the depth confidence and the polarization confidence, a depth value of the depth image and the polarization estimation depth are pixel-level weighted and fused to obtain a fusion depth map. Synchronous positioning and mapping are performed based on the fusion depth map and a synchronous RGB image, and a point cloud map of the target scene and a running trajectory of the unmanned vehicle are output.

2. The method of claim 1, wherein the method is a local high-precision positioning and mapping method for an unmanned vehicle based on reinforcement learning polarization normal disambiguation and depth information fusion. The synchronously acquired polarization image and the depth image are controlled through multi-thread parallel image transmission, and a continuous soft trigger mode is used to synchronously adjust frame rates of a polarization camera and a depth camera; the polarization camera, the depth camera, a collection card for deploying an image collection algorithm and a calculation card for deploying a core algorithm are all adapted to image collection and data processing requirements of a low-texture and low-light scene.

3. The method of claim 1, wherein the method is a local high-precision positioning and mapping method for an unmanned vehicle based on reinforcement learning polarization normal disambiguation and depth information fusion. The state vector construction, decision and training process of the reinforcement learning agent includes: the polarization normal candidates, the depth normal, the neighborhood surface curvature and the depth gradient information are used to form the state vector; an action value is calculated through a Dueling DQN architecture containing a value branch and an advantage branch, and an optimal selection is output according to a negative weighted summation reward function of the angle error of fusion of the polarization normal and the depth normal, the neighborhood normal smoothness loss and the consistency error of the polarization normal and the depth gradient direction; in an offline stage, historical data are iteratively trained until a policy converges, and in an online stage, real-time state vectors are directly input to obtain the unambiguous surface normal.

4. The method of claim 1, wherein the method is a local high-precision positioning and mapping method for an unmanned vehicle based on reinforcement learning polarization normal disambiguation and depth information fusion. The polarization confidence is evaluated in combination with a pixel polarization degree and a polarization model fitting residual angle, and the depth confidence is evaluated in combination with a reflection intensity of the depth camera and a depth image gradient amplitude; both are normalized to represent information source reliability; during fusion, the depth value of the depth image and the polarization estimation depth are dynamically weighted according to the dual confidences, a fusion depth map is obtained through weighted averaging, and a small compensation is introduced during the fusion process to avoid calculation abnormalities.

5. The method of claim 1, wherein the method is a local high-precision positioning and mapping method for unmanned vehicles based on reinforcement learning polarization normal disambiguation and depth information fusion. The synchronous positioning and mapping specifically include: key frames are screened through a tracking thread, a local mapping module optimizes a map and eliminates redundant frames, a loop detection is realized by identifying a similar historical scene through a bag-of-words model, and then a global optimization is used to correct key frame poses and map point positions, so that cumulative errors and scale drifts are eliminated, and the geometric consistency of the point cloud map is ensured.

6. An unmanned vehicle local high-precision positioning and mapping system based on reinforcement learning polarization normal disambiguation and depth information fusion, characterized in that, The application relates to a method for generating a point cloud map of a target scene and a trajectory of an unmanned vehicle. The application relates to an unmanned vehicle A polarization camera and a depth camera are mounted on the unmanned vehicle to synchronously acquire a polarization image and a depth image of a target scene in a low-texture and low-light scene; A data processing unit is connected with the polarization camera and the depth camera, and is configured to extract a polarization normal candidate, a depth normal, a neighborhood surface curvature, and depth gradient information to construct a state vector, and output an unambiguous surface normal through reinforcement learning logic based on a Dueling DQN. Synchronous localization and mapping are performed based on the fused depth map and a synchronous RGB image.

7. The local high-precision positioning and mapping system for unmanned vehicles based on reinforcement learning polarization normal disambiguation and depth information fusion according to claim 6, characterized in that: The polarization camera is a polarization camera adapted to a low-texture and low-light scene, and the depth camera is a depth camera adapted to the scene.

8. An unmanned vehicle mapping system implementing the method of any one of claims 1-5, wherein, The data processing unit includes an acquisition module, a disambiguation module, a fusion module, and a mapping module, and each module cooperates to complete the whole process from image acquisition to localization and mapping.

9. A computer device, comprising: The system includes a polarization camera and a depth camera mounted on an unmanned vehicle, an acquisition card connected with the polarization camera and the depth camera, a computing card connected with the acquisition card, and a display connected with the computing card.

10. A non-transitory computer-readable storage medium, comprising: The system includes a processor and a storage medium storing a computer program, and the processor executes the computer program to implement the method of any one of claims 1-5. The storage medium stores a computer program, and the computer program is executed by a processor to implement the method of any one of claims 1-5.

Images