Ground unmanned platform intelligent sensing method for field scene

Abstract

The invention relates to a ground unmanned platform intelligent sensing method for a field scene, and belongs to the field of automatic driving. According to the invention, a dynamic obstacle sensing flow based on geometric information is constructed through fusion of a laser radar and a millimeter-wave radar; utilizing visual semantic segmentation to construct a static environment taking semantic information as a main part and a special obstacle perception flow; and finally, through an innovative semantic raster map fusion mechanism, unifying the geometric raster map and the semantic raster map, and providing an environment model with both geometric accuracy and semantic richness for planning control. According to the invention, through laser radar-millimeter wave radar fusion and the SORT adaptive algorithm, the tracking accuracy of the dynamic obstacle is improved, the tracking delay is reduced, and the problem of changeable motion modes of the field target can be effectively solved. According to the method, the visual sensor data is collected, semantic segmentation and laser radar correction are performed, the misjudgment rate of special targets such as grasses and water pits is reduced, and the sensing reliability in a complex environment is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving, specifically relating to an intelligent perception method for unmanned ground platforms in field scenarios. Background Technology

[0002] With breakthroughs in artificial intelligence and autonomous control technologies, ground-based unmanned platforms are rapidly evolving from theoretical innovation to practical application. Intelligent perception technology, as one of the key technologies for autonomous driving of ground-based unmanned platforms, primarily addresses how to detect and identify obstacles and drivable areas. This provides the planning and control module with accurate obstacle information and drivable area data, ensuring safe vehicle operation and improving mission execution efficiency.

[0003] Currently, the high-precision perception technology for structured roads in the field of urban autonomous driving is relatively mature. However, autonomous driving in complex outdoor environments differs significantly from urban structured roads. Outdoor scenarios are mostly unstructured road surfaces, and ground unmanned platforms are mostly low-speed vehicles. Furthermore, there is a lack of readily available high-precision maps, traffic signs, and other auxiliary information. How to achieve integrated intelligent perception of ground unmanned platforms in complex environments by combining sensors such as inertial navigation, lidar, visible light cameras, and millimeter-wave radar remains a challenging problem.

[0004] Currently, the intelligent sensing technology of unmanned ground platforms in field scenarios still faces the following bottlenecks: LiDAR has a false detection rate of over 30% when dealing with special obstacles such as grass and puddles. A single sensor cannot effectively detect and track targets, and the dynamic target tracking delay exceeds 100ms.

[0005] To address the challenges of complex road conditions, varied obstacles, and limited environmental features in wilderness scenarios, as well as the shortcomings of existing technologies such as poor robustness of single sensors and lack of specificity in multi-sensor fusion, this invention proposes an intelligent perception method for unmanned ground platforms in wilderness environments. The core concept lies in solving the intelligent perception problem of wilderness environments through a collaborative perception framework that separates dynamic and static, geometric and semantic information. Specifically, it utilizes the fusion of LiDAR and millimeter-wave radar to construct a dynamic obstacle perception flow primarily based on geometric information; it employs visual semantic segmentation to construct a static environment and specific obstacle (such as grass and puddles) perception flow primarily based on semantic information; and finally, it unifies the two through an innovative semantic raster map fusion mechanism, providing an environmental model for planning and control that combines geometric accuracy with semantic richness. Summary of the Invention

[0006] (a) Technical problems to be solved The technical problem to be solved by this invention is how to provide an intelligent perception method for ground unmanned platforms in field scenarios, so as to solve the problems of complex road conditions, variable obstacles, and few environmental features in field scenarios, as well as the problems of poor robustness of single sensors and lack of specificity of multi-sensor fusion in existing technologies.

[0007] (II) Technical Solution To address the aforementioned technical problems, this invention proposes an intelligent perception method for unmanned ground platforms in field scenarios, comprising the following steps: S1. Align the timestamps of the lidar, millimeter-wave radar, and visible light camera to the combined inertial navigation clock; calculate the rotation matrix R1 and translation vector T1 from the lidar coordinate system to the ground unmanned platform coordinate system, and calculate the rotation matrix R2 and translation vector T2 from the camera to the lidar coordinate system. S2. For dynamic obstacles, detection and tracking are achieved through the fusion of lidar and millimeter-wave radar: the lidar data is processed to obtain obstacle set information, the millimeter-wave radar data is processed to obtain dynamic obstacle information, obstacle information is fused according to the characteristics of millimeter-wave radar and lidar sensors, and dynamic obstacles are tracked in real time. S3. For static obstacles, real-time acquisition of visible light camera images is used, which are input into the pre-trained model and output as semantic segmentation results. The point cloud of static obstacles from the lidar is used to correct semantic misjudgment areas. S4. Perform inverse perspective transformation on the semantic segmentation results to convert them into a two-dimensional semantic raster map; project the static and dynamic obstacle point clouds detected by the LiDAR onto the same coordinate system to generate an occupied raster map, and use a logical OR strategy to generate the final fused raster map.

[0008] (III) Beneficial Effects This invention proposes an intelligent perception method for unmanned ground platforms in field scenarios. The proposed intelligent perception method improves the tracking accuracy of dynamic obstacles and reduces tracking latency through lidar-millimeter-wave radar fusion and SORT adaptive algorithm, which can effectively cope with the problem of variable target movement patterns in the field.

[0009] The intelligent perception method proposed in this invention collects visual sensor data and combines semantic segmentation of the DeepLabv3+ algorithm after training and optimization with LiDAR correction to reduce the misjudgment rate of special targets such as grass and puddles, thus ensuring the reliability of perception in complex environments. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the intelligent sensing technology roadmap for unmanned ground platforms in field scenarios according to the present invention. Figure 2This is a schematic diagram of a typical intelligent perception scenario for unmanned ground platforms in the field. Detailed Implementation

[0011] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0012] This invention relates to the field of autonomous driving, specifically to the application of multi-sensor fusion and semantic segmentation technology in intelligent perception of ground unmanned platforms in field scenarios, and particularly to an intelligent perception method for ground unmanned platforms in complex field environments.

[0013] Wilderness scenarios are characterized by complex road conditions, varied obstacles, and limited environmental features. This paper proposes an intelligent perception method for unmanned ground platforms in wilderness environments. The core concept lies in effectively addressing the perception challenges of high tracking delays for dynamic obstacles, high misjudgment rates for special obstacles (such as grass and puddles), and the lack of prior map support through a collaborative perception framework that separates dynamic and static, geometric and semantic information. Specifically, it utilizes the fusion of LiDAR and millimeter-wave radar to construct a dynamic obstacle perception flow primarily based on geometric information; it employs visual semantic segmentation to construct a static environment and special obstacle perception flow primarily based on semantic information; and finally, it unifies the two through an innovative semantic grid map fusion mechanism. The planning and control module, based on navigation positioning, odometer, waypoint tasks, and the semantic grid map, realizes path planning and real-time obstacle avoidance for the unmanned ground platform in wilderness scenarios, thereby achieving autonomous obstacle avoidance and autonomous driving.

[0014] The technical problem this invention aims to solve is how to achieve intelligent perception, including obstacle detection and drivable area detection, using vehicle-mounted sensors in unstructured road environments where dynamic obstacles are variable, static obstacles have complex semantics, and prior maps are lacking. The specific technical solution is as follows: This invention provides an intelligent perception method for unmanned ground platforms in field scenarios, the method comprising: S1. Align the timestamps of the lidar, millimeter-wave radar, and visible light camera to the combined inertial navigation clock; calculate the rotation matrix R1 and translation vector T1 from the lidar coordinate system to the ground unmanned platform coordinate system, and calculate the rotation matrix R2 and translation vector T2 from the camera to the lidar coordinate system. S2. For dynamic obstacles, detection and tracking are achieved through the fusion of lidar and millimeter-wave radar: the lidar data is processed to obtain obstacle set information, the millimeter-wave radar data is processed to obtain dynamic obstacle information, obstacle information is fused according to the characteristics of millimeter-wave radar and lidar sensors, and dynamic obstacles are tracked in real time. S3. For static obstacles, real-time acquisition of visible light camera images is used, which are input into the pre-trained model and output as semantic segmentation results. The point cloud of static obstacles from the lidar is used to correct semantic misjudgment areas. S4. Perform inverse perspective transformation on the semantic segmentation results to convert them into a two-dimensional semantic raster map; project the static and dynamic obstacle point clouds detected by the LiDAR onto the same coordinate system to generate an occupied raster map, and use a logical OR strategy to generate the final fused raster map.

[0015] Example 1: ① Sensor selection and basic functions The integrated inertial navigation system (INS) can acquire vehicle position and attitude information using the Global Navigation Satellite System (GNSS) and the Inertial Measurement Unit (IMU). When satellite signals are available, vehicle positioning accuracy can reach the centimeter level; when satellite signals are unavailable, vehicle position and attitude changes can be calculated through laser-inertial odometry fusion, continuously outputting high-precision navigation and positioning information for a short period of time.

[0016] LiDAR (Light Detection and Ranging) offers advantages such as high resolution and real-time performance. It can be used in traditional methods, such as downsampling, ground point cloud segmentation, and clustering, to detect obstacles. Alternatively, it can employ deep learning methods, involving dataset processing, model training and optimization, and model deployment and inference, to achieve obstacle detection. In addition to obstacle detection, LiDAR can also acquire laser odometry information using algorithms such as Normal Distributions Transform (NDT) or Iterated Closest Points (ICP) for navigation and positioning.

[0017] Visible light cameras are characterized by low cost and good visualization. They can detect obstacles through target detection algorithms, including common targets such as people, vehicles, and traffic cones, as well as special targets such as water barriers and barricades. They can detect drivable areas through semantic segmentation algorithms, and recognize and segment object types based on models trained on labeled data.

[0018] Millimeter-wave radar is characterized by its low cost and good anti-interference ability. It calculates the position and speed of obstacles by measuring the time difference between the electromagnetic waves it emits and the electromagnetic waves reflected by the object. This sensor makes up for the shortcomings of visual solutions being more sensitive to light and lidar having poor detection performance in rain and fog environments.

[0019] ②Time synchronization and joint calibration The PPS+GPRMC method is used to align the timestamps of the lidar, millimeter-wave radar, and visible light camera to the combined inertial navigation clock, ensuring that the synchronization error is ≤1ms and achieving time synchronization between sensors.

[0020] The laser point cloud data of the three-dimensional feature calibration board (a flat plate with a large number of regular holes for extrinsic parameter calibration) in 30 different poses was collected. The laser point cloud was registered with the accurate three-dimensional model of the calibration board using the ICP algorithm. The rotation matrix R1 and translation vector T1 from the laser radar coordinate system to the vehicle coordinate system were calculated to ensure that the calibration error is ≤0.05m, thus realizing the extrinsic parameter calibration of the laser radar. Three-dimensional markers that can be precisely located in the laser point cloud are placed at the corner points of the traditional chessboard plane to construct a "three-dimensional target". By extracting the pixel coordinates of the chessboard corner points in the image and the three-dimensional center coordinates of the corresponding three-dimensional markers in the laser point cloud, the correspondence between the image (2D) and the laser point cloud (3D) is obtained. The rotation matrix R2 and translation vector T2 from the camera to the lidar coordinate system are solved to ensure the spatial alignment accuracy between the image and the laser point cloud, and to achieve joint calibration of the camera and lidar.

[0021] ③ Construction of dynamic obstacle perception flow For dynamic obstacles such as moving people and vehicles, detection and tracking are achieved through the fusion of "LiDAR and millimeter-wave radar". The specific process is as follows: LiDAR data processing: Voxel network filtering and downsampling are employed, with voxel sizes of 0.1m × 0.1m × 0.1m, balancing the accuracy and speed of LiDAR point cloud data processing. To effectively address the interference from undulating unstructured road surfaces in the field, ground point cloud segmentation based on plane fitting and radial analysis is used to remove road surface interference. Euclidean clustering based on adaptive thresholds is used for obstacle aggregation to obtain geometric information such as obstacle location and size, with a near-range threshold of 0.3m and a far-range threshold of 0.8m. When the distance between two obstacle point clouds is less than the threshold for the corresponding scene, the two point clouds are determined to belong to the same obstacle and are aggregated to obtain obstacle set information.

[0022] Millimeter-wave radar data processing: The intermediate frequency signal containing obstacle information is processed by fast Fourier transform to obtain a range-amplitude map; the obstacle's velocity is calculated by the phase change rate to obtain dynamic obstacle velocity information.

[0023] Data Fusion and Tracking: Based on the characteristics of millimeter-wave radar and lidar sensors, millimeter-wave radar detection data is prioritized in close-range scenarios (less than 20m). In medium-range scenarios (20m to 50m), the Hungarian algorithm is used to match millimeter-wave radar and lidar detection data, fusing geometric and velocity information. In long-range scenarios (greater than 50m), lidar detection data is output. Addressing the variable movement patterns of targets in outdoor scenarios, the SimpleOnline and Realtime Tracking (SORT) algorithm, which uses a single uniform velocity model prediction method, is adaptively adjusted. Specifically, by fusing direct velocity measurement information from millimeter-wave radar, the rate of change of obstacle velocity is calculated in real time. When the rate of change of velocity exceeds 0.5m / s², the algorithm switches from a "uniform velocity model" to a "uniform acceleration model," significantly reducing the probability of incorrect matching during the data association stage and improving the reliability of the system in tracking dynamic obstacles.

[0024] ④ Static Perception Flow Construction For static obstacles such as grass and puddles in the wild, semantic segmentation is achieved through "visual-LiDAR fusion". The specific process is as follows: Using the pre-trained open-source model DeepLabv3+, a large number of real-world RGB images were collected. The LabelMe annotation tool was used to assign category labels to each pixel in each training image based on a predefined category list, such as 0-background, 1-road, 2-sky, 3-grass, 4-puddles, 5-trees, 6-pedestrians, 7-vehicles, etc. The trained and optimized DeepLabv3+ model was deployed on a ground-based unmanned platform to achieve static target semantic labeling and boundary recognition, obtaining semantic segmentation results. A label map corresponding to each pixel of the input image is output, with its coordinate system being the image pixel coordinate system. Using the camera-LiDAR joint calibration parameters obtained in the previous calibration, the static obstacle point cloud detected by the LiDAR is projected onto the image pixel coordinate system. If there is no LiDAR point cloud in the "grass / puddles" area in the semantic segmentation result, it is marked as "impassable" by default; if the LiDAR point cloud shows an area as an obstacle but the semantic segmentation does not label it, it is supplemented with the label "unknown obstacle" to correct semantic misjudgment.

[0025] ⑤ Semantic raster map fusion The semantic segmentation results are subjected to inverse perspective transformation to convert them into a two-dimensional semantic raster map with a raster resolution of 0.2m × 0.2m. Impassable categories (such as grass, puddles, etc.) are marked as "1", and drivable areas are marked as "0". The point clouds of obstacles (including dynamic and static) detected by the LiDAR are projected onto the same coordinate system to generate an occupied raster map. Obstacle areas are marked as "1", and obstacle-free areas are marked as "0". A logical OR strategy is used to generate the final fused raster map: for any raster, if it is marked as "1" in the LiDAR occupied raster map, or if it is marked as impassable category "1" in the visual semantic raster map, then the raster is marked as "impassable" (value 1) in the fused map; only when both are marked as "0" is it marked as "drivable" (value 0). To improve the accuracy of fusion perception, this invention introduces a scene-adaptive perception optimization mechanism before generating the aforementioned binary grid map: In low-light scenes, due to decreased visual reliability, the confidence level of "impassable" can be appropriately increased when generating the semantic grid map, while the size threshold of clustered obstacles can be appropriately reduced when generating the occupied grid map, thereby increasing the "weight" of the LiDAR; In sunny scenes, the opposite strategy is executed, reducing the judgment threshold of the semantic grid to utilize richer semantic information, while restoring the size threshold of the LiDAR, thereby increasing the "weight" of visual semantics.

[0026] Example 1: The intelligent perception method for ground-based unmanned platforms in complex field environments proposed in this invention has the following specific implementation steps: ① Sensor installation and debugging A combined inertial navigation system is installed in the middle of the wire-controlled chassis of the ground unmanned platform, and a lidar, millimeter-wave radar and a visible light camera are installed in front of the platform to ensure that the front of the sensors are not obstructed.

[0027] Power on the sensor and confirm normal data acquisition via the host computer. The combined inertial navigation system outputs position and attitude information, the lidar outputs point cloud data, the millimeter-wave radar outputs range-velocity information, and the camera outputs RGB images.

[0028] Following the steps of time synchronization and joint calibration, complete the time synchronization and sensor joint calibration, and verify that the synchronization error is ≤1ms and the spatial calibration error is ≤0.05m.

[0029] ② Dynamic obstacle detection and tracking Real-time acquisition of LiDAR and millimeter-wave radar data; LiDAR data is processed by voxel filtering, ground segmentation and Euclidean clustering to obtain obstacle position and size; millimeter-wave radar data is processed by FFT to obtain obstacle position and velocity information. Data is fused in segments based on distance, and combined with the absolute position of the combined inertial navigation vehicle to calculate the absolute position, size, and velocity of obstacles; Run the adaptive SORT algorithm to achieve dynamic obstacle tracking, and output the ID, absolute position, velocity and size information of the tracked target, with an update frequency of ≥10Hz.

[0030] ③Static environment semantic segmentation and correction Real-time acquisition of visible light camera images, input into a pre-trained DeepLabv3+ model, and output of semantic segmentation results; Collect static obstacle point clouds from lidar, project them onto the semantic segmentation coordinate system, and correct semantic misjudgment areas; The corrected semantic segmentation results are subjected to inverse perspective transformation to generate a semantic raster map with an update frequency of ≥5Hz.

[0031] ④ Generating merged raster maps Collect point clouds of obstacles from LiDAR and generate an occupied grid map; Following the "logical OR" rule, the semantic raster map and the occupied raster map are merged to generate the final merged raster map; Under conditions of low light intensity, the confidence level of areas identified as "impassable areas" in semantic raster maps generated based on visual information can be appropriately adjusted, the size threshold of obstacles in lidar point cloud clustering can be reduced, and the accuracy of intelligent perception can be improved.

[0032] ⑤ Autonomous driving control The planning and control module receives waypoint tasks and fused grid maps. Based on the path planning algorithm (A* or Dijkstra) and combined with the "impassable" markers in the fused raster map, a feasible path from the current location to the target waypoint is planned; The system adjusts its path in real time based on dynamic obstacle tracking results to achieve autonomous obstacle avoidance, control the ground unmanned platform to travel along the planned path, and complete tasks in the field.

[0033] The beneficial effects of this invention are: Compared with existing technologies, the intelligent perception method for unmanned ground platforms in field scenarios described in this invention has the following beneficial effects: The intelligent sensing method proposed in this invention improves the tracking accuracy of dynamic obstacles and reduces tracking latency through the fusion of lidar and millimeter-wave radar and the SORT adaptive algorithm, which can effectively cope with the problem of variable movement patterns of targets in the field.

[0034] The intelligent perception method proposed in this invention collects visual sensor data and combines semantic segmentation of the DeepLabv3+ algorithm after training and optimization with LiDAR correction to reduce the misjudgment rate of special targets such as grass and puddles, thus ensuring the reliability of perception in complex environments.

[0035] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent perception of unmanned ground platforms in field scenarios, characterized in that, The method includes the following steps: S1. Align the timestamps of the lidar, millimeter-wave radar, and visible light camera to the combined inertial navigation clock; calculate the rotation matrix R1 and translation vector T1 from the lidar coordinate system to the ground unmanned platform coordinate system, and calculate the rotation matrix R2 and translation vector T2 from the camera to the lidar coordinate system. S2. For dynamic obstacles, detection and tracking are achieved through the fusion of lidar and millimeter-wave radar: the lidar data is processed to obtain obstacle set information, the millimeter-wave radar data is processed to obtain dynamic obstacle information, obstacle information is fused according to the characteristics of millimeter-wave radar and lidar sensors, and dynamic obstacles are tracked in real time. S3. For static obstacles, real-time acquisition of visible light camera images is used, which are input into the pre-trained model and output as semantic segmentation results. The point cloud of static obstacles from the lidar is used to correct semantic misjudgment areas. S4. Perform inverse perspective transformation on the semantic segmentation results to convert them into a two-dimensional semantic raster map; project the static and dynamic obstacle point clouds detected by the LiDAR onto the same coordinate system to generate an occupied raster map, and use a logical OR strategy to generate the final fused raster map.

2. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 1, characterized in that, In step S1, a combined inertial navigation system is installed in the middle of the ground unmanned platform's wire-controlled chassis. A lidar, millimeter-wave radar, and a visible light camera are installed in front of the platform to ensure that the sensor front ends are not obstructed. The sensors are powered on and started. The host computer confirms that the data acquisition is normal. The combined inertial navigation system outputs position and attitude information, the lidar outputs point cloud data, the millimeter-wave radar outputs distance-velocity information, and the camera outputs RGB images.

3. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 2, characterized in that, In step S1, the timestamps of the lidar, millimeter-wave radar, and visible light camera are aligned to the combined inertial navigation clock using the PPS+GPRMC method to ensure that the synchronization error is ≤1ms and to achieve time synchronization between the sensors.

4. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 2, characterized in that, In step S1, calculating the rotation matrix R1 and translation vector T1 from the lidar coordinate system to the ground unmanned platform coordinate system includes: collecting lidar point cloud data of a 3D feature calibration board in 30 different poses, registering the lidar point cloud with the accurate 3D model of the calibration board using the ICP algorithm, calculating the rotation matrix R1 and translation vector T1 from the lidar coordinate system to the vehicle coordinate system, ensuring that the calibration error is ≤0.05m, and realizing the external parameter calibration of the lidar.

5. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 2, characterized in that, In step S1, calculating the rotation matrix R2 and translation vector T2 from the camera to the lidar coordinate system includes: placing three-dimensional markers that can be accurately located in the lidar point cloud at the corner points of the checkerboard plane to construct a "three-dimensional target"; obtaining the correspondence between the image and the lidar point cloud by extracting the pixel coordinates of the checkerboard corner points in the image and the three-dimensional center coordinates of the corresponding three-dimensional markers in the lidar point cloud; solving for the rotation matrix R2 and translation vector T2 from the camera to the lidar coordinate system to achieve joint calibration of the camera and lidar.

6. The intelligent perception method for unmanned ground platforms in field scenarios as described in any one of claims 1-5, characterized in that, S2 includes: LiDAR data processing: Voxel network filtering downsampling is used, and ground point cloud segmentation is performed based on plane fitting and radial analysis to remove road surface interference; Euclidean clustering based on adaptive threshold is used to aggregate obstacles to obtain obstacle location and size information. When the distance between two obstacle point clouds is less than the threshold in the corresponding scene, the two point clouds are determined to belong to the same obstacle and are aggregated to obtain obstacle set information. Millimeter-wave radar data processing: The intermediate frequency signal containing obstacle information is processed by fast Fourier transform to obtain a range-amplitude map; the obstacle's velocity is calculated by the phase change rate to obtain dynamic obstacle velocity information; Data fusion and tracking: Based on the characteristics of millimeter-wave radar and lidar sensors, millimeter-wave radar detection data is prioritized in close-range scenarios. In mid-range scenarios, the Hungarian algorithm is used to match millimeter-wave radar and lidar detection data, fusing geometric and velocity information. In long-range scenarios, lidar detection data is output. Considering the variable target motion patterns in outdoor scenarios, the SORT algorithm is adopted. By fusing direct velocity measurement information from millimeter-wave radar, the velocity change rate of obstacles is calculated in real time. When the velocity change rate is greater than 0.5 m / s², the system switches from a "uniform velocity model" to a "uniform acceleration model".

7. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 6, characterized in that, The adaptive thresholds are: 0.3m for close distance and 0.8m for long distance.

8. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 6, characterized in that, S4 includes: using the pre-trained open-source model DeepLabv3+, assigning category labels to each pixel in each training image based on a predefined category list using the LabelMe annotation tool to collect RGB images of actual field scenes; training and optimizing the DeepLabv3+ model; deploying it on a ground-based unmanned platform to achieve static target semantic labeling and boundary recognition, and obtaining semantic segmentation results; outputting a label map corresponding to each pixel of the input image, with its coordinate system being the image pixel coordinate system; using the camera-LiDAR joint calibration parameters obtained in the previous calibration, projecting the static obstacle point cloud detected by the LiDAR onto the image pixel coordinate system; if there is no LiDAR point cloud in the "grass / puddle" area in the semantic segmentation result, it is marked as "impassable" by default; if the LiDAR point cloud shows that a certain area is an obstacle but the semantic segmentation does not label it, it is supplemented with the label "unknown obstacle" to correct semantic misjudgment.

9. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 8, characterized in that, S5 includes: performing an inverse perspective transformation on the semantic segmentation result to convert it into a two-dimensional semantic grid map, with the non-drivable category marked as "1" and the drivable area marked as "0"; projecting the dynamic and static obstacle point clouds detected by the lidar onto the same coordinate system to generate an occupied grid map, with obstacle areas marked as "1" and obstacle-free areas marked as "0"; and generating the final fused grid map using a "logical OR" strategy: for any grid, if it is marked as "1" in the lidar occupied grid map or as non-drivable category "1" in the visual semantic grid map, then the grid is marked as "impassable" in the fused map; only when both are marked as "0" is it marked as "drivable".

10. The intelligent perception method for unmanned ground platforms in field scenarios as described in claim 9, characterized in that, Before generating the grid map, S5 introduces a scene-adaptive perception optimization mechanism: In low-light scenes, due to decreased visual reliability, when generating the semantic grid map, the confidence level of the judgment as "impassable" is appropriately increased, and when generating the occupied grid map, the size threshold of its clustered obstacles is appropriately reduced, thereby increasing the "weight" of the LiDAR; In sunny scenes, the opposite strategy is executed, reducing the judgment threshold of the semantic grid to utilize richer semantic information, while restoring the size threshold of the LiDAR, thereby increasing the "weight" of visual semantics.

Images