Non-interfering entity detection method and unmanned vehicle

Abstract

The invention discloses a non-interfering entity detection method and an unmanned vehicle. The method comprises the steps that a detection result of the unmanned vehicle is acquired, and the detection result is obtained by detecting a target environment through sensing equipment of the unmanned vehicle; the detection result is spatially divided according to a target grid model, spatial division feature data are obtained, and the target grid model is provided with grids which are non-uniformly distributed in the height direction and grids which are uniformly distributed in the horizontal direction; and inputting the spatial division feature data into a non-interference entity identification model for identification to obtain an identification result, the identification result being used for representing whether a non-interference entity exists in the target environment. According to the invention, the technical problem of low detection accuracy of non-interfering entities is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned vehicles and perception, in particular, to a non-interference entity detection method and an unmanned vehicle. BACKGROUND

[0002] At present, in the field of unmanned driving, Light Detection and Ranging (LIDAR) perception technology is a core component, which is used to obtain three-dimensional (3D) point cloud data of the surrounding environment of an unmanned vehicle in real time, and then to detect, classify and locate obstacles.

[0003] In the related art, LIDAR perception 3D segmentation (non-interference entity recognition) can be based on a Bird's Eye View (BEV) perspective and use a fixed resolution spatial grid to process point cloud data. In the above method, each grid or column independently outputs a class label, such as ground, retaining wall, pedestrian, vehicle or noise. Although the BEV method performs well in some scenarios, the method is insufficient in spatial granularity, which limits its performance in complex environments. Therefore, there is still a technical problem of low detection accuracy of non-interference entities.

[0004] At present, there is no effective solution to the above problems. SUMMARY

[0005] The embodiments of the present application provide a non-interference entity detection method and an unmanned vehicle to at least solve the technical problem of low detection accuracy of non-interference entities.

[0006] According to an aspect of the embodiments of the present application, a non-interference entity detection method is provided, comprising: obtaining a detection result of an unmanned vehicle, wherein the detection result is obtained by a perception device of the unmanned vehicle detecting a target environment; performing spatial division on the detection result according to a target grid model to obtain spatial division feature data, wherein the target grid model has non-uniformly distributed grids in the height direction and uniformly distributed grids in the horizontal direction; inputting the spatial division feature data into a non-interference entity recognition model for recognition to obtain a recognition result, wherein the recognition result is used to indicate whether there is a non-interference entity in the target environment.

[0007] According to an aspect of some embodiments of the present application, there is provided another method for detecting non-interference entity, comprising: detecting a service state and / or a region of an unmanned vehicle; and controlling the unmanned vehicle to start a target perception mode in a case where the service state and / or the region meets a target condition, wherein the target perception mode is used to detect whether there is a non-interference entity in a current environment.

[0008] According to another aspect of some embodiments of the present application, there is also provided a device for detecting non-interference entity, comprising: an obtaining unit configured to obtain a detection result of an unmanned vehicle, wherein the detection result is obtained by a perception device of the unmanned vehicle detecting a target environment; a dividing unit configured to divide the detection result according to a target grid model to obtain spatial division feature data, wherein the target grid model has non-uniformly distributed grids in a height direction and uniformly distributed grids in a horizontal direction; and an identifying unit configured to input the spatial division feature data into a non-interference entity identification model to obtain an identification result, wherein the identification result is used to indicate whether there is a non-interference entity in the target environment.

[0009] According to another aspect of some embodiments of the present application, there is also provided another device for detecting non-interference entity, comprising: a detecting unit configured to detect a service state and / or a region of an unmanned vehicle; and a controlling unit configured to control the unmanned vehicle to start a target perception mode in a case where the service state and / or the region meets a target condition, wherein the target perception mode is used to detect whether there is a non-interference entity in a current environment.

[0010] According to another aspect of some embodiments of the present application, there is also provided a computer readable storage medium, which stores a plurality of instructions, wherein the instructions are adapted to be loaded and executed by a processor to perform any one of the above methods. According to another aspect of some embodiments of the present application, there is also provided an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to perform any one of the above methods. According to another aspect of some embodiments of the present application, there is also provided an unmanned vehicle, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to perform any one of the above methods. According to another aspect of some embodiments of the present application, there is also provided a computer program product, comprising a computer program, wherein the computer program is used to perform any one of the above methods when executed by a processor.

[0011] In the embodiment of the present application, if it is necessary to identify whether there is a non-interference entity in the target environment where the unmanned vehicle is located, the detection result of the target environment detected by the perception device of the unmanned vehicle can be obtained. The detection result can be spatially divided according to the target grid model to obtain spatial division feature data. The spatial division feature data can be input into the non-interference entity model for identification to obtain the identification result of whether there is a non-interference entity in the target environment.

[0012] In the above embodiment of the present application, considering the non-interference entity which does not affect the driving of the unmanned vehicle in addition to the actual obstacle, the non-interference entity may be irregularly distributed in height and has a certain interval with the actual entity obstacle on the ground. Therefore, by using a non-uniformly distributed grid in the height direction, the space can be more finely divided, especially in the height interval where the non-interference entity frequently occurs, a finer grid division is used, so that even if the non-interference entity and the entity obstacle are close in the horizontal direction, they can be distinguished in the vertical direction. By using the non-uniform height distribution grid, the detection result can be converted into spatial division feature data more suitable for non-interference entity identification. The spatial division feature data is input into the non-interference entity identification model, which can more accurately distinguish the non-interference entity and the entity obstacle by learning the features in the non-uniform height distribution grid. Compared with the model based only on the uniform grid, the non-interference entity identification model has higher sensitivity and accuracy in processing the change of features in the height direction, thereby reducing the false detection, achieving the technical effect of improving the detection accuracy of the non-interference entity, and solving the technical problem of low detection accuracy of the non-interference entity. BRIEF DESCRIPTION OF DRAWINGS

[0013] The accompanying drawings, which are included to provide a further understanding of the present application and are incorporated in and constitute a part of this application, illustrate embodiments of the present application and serve to explain the present application. In the drawings:

[0014] Figure 1 is a flowchart of a non-interference entity detection method according to an embodiment of the present application;

[0015] Figure 2 is a flowchart of another non-interference entity detection method according to an embodiment of the present application;

[0016] FIG. 3(a) is a schematic diagram of a LIDAR point cloud data in the related art;

[0017] FIG. 3(b) is a schematic diagram of a 3D segmentation false detection of a retaining wall, ground and other entity substances caused by geothermal in the related art;

[0018] Figure 4is a schematic diagram of a grid structure according to an embodiment of the present application;

[0019] Figure 5 is a flow chart of a method for solving the false detection of geothermal gas by LIDAR sensing in mine unmanned driving according to an embodiment of the present application;

[0020] Figure 6 is a schematic diagram of a highly non-uniform zooming grid structure according to an embodiment of the present application;

[0021] Figure 7 is a schematic diagram of a structure during model training according to an embodiment of the present application;

[0022] Figure 8 is a structural schematic diagram of a non-interfering entity detection device according to an embodiment of the present application;

[0023] Figure 9 is a structural schematic diagram of another non-interfering entity detection device according to an embodiment of the present application;

[0024] Figure 10 is a schematic diagram of an electronic device for a non-interfering entity detection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the personnel in the technical field better understand the present application scheme, the technical scheme in the embodiment of the present application will be described clearly and completely below in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0027] According to the embodiment of the present application, a non-interfering entity detection method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0028] Figure 1 is a flowchart of a non-interfering entity detection method according to an embodiment of the present application, as shown in Figure 1 , the method comprises the following steps:

[0029] Step S102, obtaining the detection result of the unmanned vehicle.

[0030] In the technical solution provided in step S102 of the embodiment of the present application, the detection result can be obtained by the perception device of the unmanned vehicle. That is, the detection result can refer to the data or information obtained by the perception device of the unmanned vehicle after detecting the target environment, which can also be called perception data. The detection result can be three-dimensional point cloud data generated by the LIDAR sensor after scanning the surrounding environment of the unmanned vehicle, which contains information such as the position, shape and size of the objects in the surrounding environment. The three-dimensional point cloud data can also be called LIDAR point cloud data or LIDAR data. The detection result can also include data from cameras, radars, ultrasonic sensors and other perception devices. The above detection result constitutes the perception result obtained by perceiving the condition of the target environment where the unmanned vehicle is located. The perception device can be a hardware device installed on the unmanned vehicle for collecting various information in the target environment. In the embodiment of the present application, the perception device can be a LIDAR. The LIDAR can construct a three-dimensional image of the target environment by emitting laser pulses and measuring the reflection time. In addition to the LIDAR, the unmanned vehicle can also be equipped with cameras, millimeter wave radars, ultrasonic sensors and other perception devices, which are not limited here.

[0031] Optionally, the unmanned vehicle refers to a car or other traffic tool that can automatically travel without the intervention of a driver. In the embodiments of the present application, the unmanned vehicle involved can be a special vehicle working in a mine environment, or a common unmanned vehicle running on urban roads. The challenges and problems to be solved of the unmanned vehicle in different application scenarios are different, but the method of the present application can be generally applicable to improve the detection ability of the vehicle to non-interference entities in different scenarios. The target environment refers to the external world that the unmanned vehicle needs to perceive and understand. In the context of the present application, the target environment can be a mine scene, or a city traffic environment, etc. which is easy to produce non-interference entities. The non-interference entity refers to a natural phenomenon or man-made phenomenon that may be misidentified as an actual obstacle in certain specific environment, especially in the perception device of the unmanned vehicle, but in fact these phenomena do not pose a safety threat or substantial obstacle to the normal travel of the vehicle. The existence of non-interference entities mainly increases the complexity and false detection risk of the perception device, that is, the non-interference entity will be misdetected by the perception device as a road obstacle that hinders the travel of the vehicle, but the non-interference entity is not a real road obstacle. In the scenarios of mine unmanned driving and urban road automatic driving, identifying and filtering non-interference entities is crucial to improve the perception accuracy and vehicle operation efficiency. For mine scenarios, non-interference entities can be fog, geothermal gas, and dust, etc. For ordinary urban road scenarios, non-interference entities can be fog, rain, snow, dust, etc.

[0032] It should be noted that the above non-interference entities are only for illustration and are not specifically limited here. They can be pre-set according to the application scenario in which the actual unmanned vehicle is located. As long as the non-interference entity is easy to be perceived as an entity obstacle by the unmanned vehicle in the actual application scenario, it is within the protection scope of the embodiments of the present application, which will not be illustrated one by one here.

[0033] In this embodiment, if it is necessary to detect whether there is a non-interference entity in the target environment around the unmanned vehicle during the travel of the unmanned vehicle, the detection result of the target environment detected by the unmanned vehicle can be obtained.

[0034] Optionally, during the travel of the unmanned vehicle in the target environment, the perception device (such as LIDAR, camera, radar, etc.) can be activated, and the target environment can be detected in real time by the perception device. The static and dynamic information in the target environment, such as road layout, obstacles, pedestrians, vehicles, etc. can be captured by the perception device, wherein the LIDAR generates high-precision three-dimensional point cloud data (LIDAR data) by emitting laser pulses and receiving reflected signals. The information and data collected by the above various perception devices, which can assist in determining whether there is a non-interference entity in the target environment, are summarized as the above detection result.

[0035] In step S104, the detection result is spatially divided according to the target grid model to obtain spatial division feature data.

[0036] In the technical solution provided in the above step S104 of the embodiment of the present application, the target grid model has non-uniformly distributed grids in the height direction and uniformly distributed grids in the horizontal direction.

[0037] Optionally, the target grid model is used to process the detection result, especially to spatially divide the LIDAR point cloud data. The target grid model is characterized by the combination of non-uniformly distributed grids in the height direction and uniformly distributed grids in the horizontal direction. The non-uniformly distributed grids mean that the size and density of the grids can vary at different heights, while the horizontal direction grids maintain consistent size and spacing, forming a standard two-dimensional grid.

[0038] The strategy of using non-uniformly distributed grids in the height direction is because at certain heights, the characteristics of three-dimensional point cloud data can be more intensive or important, especially when identifying random noise points generated by geothermal gas. For example, noise points generated by geothermal gas are often located in a specific height range, so using finer grids in this height range can more accurately capture the distribution details of the noise points, which helps subsequent feature extraction and obstacle identification. At other height intervals, using larger grid sizes can reduce computational load while not sacrificing necessary identification accuracy, effectively utilizing computing resources. The grids in the horizontal direction remain uniformly distributed because the distribution of three-dimensional point cloud data is generally uniform at the same horizontal plane, and there is no need for excessive fine division. Uniform grids simplify the data processing process and also ensure stable identification of solid obstacles (such as retaining walls and rock piles) in the horizontal plane.

[0039] Optionally, the spatial division feature data is obtained by processing the detection result through the target grid model. The spatial division feature data not only retains the spatial position information of the original three-dimensional point cloud data, but also enhances the organization and structuralization of the three-dimensional point cloud data through grid division, making subsequent data processing and analysis more orderly and efficient.

[0040] Optionally, the three-dimensional point cloud data is assigned to the corresponding grid cells by spatially dividing the detection result through the target grid model. The point cloud data in each grid cell is classified, counted, and aggregated to form a set of spatial division feature data. The spatial division feature data can include not only the position information of the point cloud, but also additional attributes such as the density, intensity, and type distribution of the point cloud, providing a rich information foundation for subsequent non-interference entity identification models to identify whether there is a non-interference entity.

[0041] Optionally, since geothermal gas noise points are usually distributed within a certain height range, a non-uniformly distributed grid model can provide higher resolution within this height range, thereby more clearly identifying and distinguishing noise points from solid obstacles. For example, when the grid model uses a smaller grid size in the height range where geothermal gas is commonly found, even if the noise points and solid obstacles are very close in the horizontal direction, they can be distinguished in the vertical direction by grid division. This strategy improves the accuracy of non-interference entity identification and reduces the likelihood of false positives for non-interference entities. By using a non-uniformly distributed grid in the height direction while maintaining uniform distribution in the horizontal direction, the embodiments of the present application find a balance between computational resource consumption and identification accuracy. By densely sampling in the height range of high-frequency information and reducing the sampling density in the height range where information is relatively sparse or less important, the computational resources are not excessively consumed, and the identification accuracy in critical areas is guaranteed.

[0042] In this embodiment, after obtaining the detection result of the unmanned vehicle, the detection result can be spatially divided according to the target grid model to obtain spatial division feature data.

[0043] Optionally, the detection result, i.e., the point cloud data obtained by the unmanned vehicle sensing device (such as LIDAR), is spatially divided according to the target grid model to obtain spatial division feature data. This means that each point cloud data point is assigned to a specific three-dimensional grid cell. For non-uniform distribution in the height direction, there are multiple grid levels in the height direction, and each grid level can have different resolutions, thereby being able to capture various features in the point cloud data in different height regions in more detail.

[0044] Optionally, after spatially dividing the point cloud data, the point cloud data in each grid cell is aggregated, counted, or processed in some way to extract the features of the grid cell as spatial division feature data. The above-mentioned spatial division feature data can include point density, point type (such as geothermal gas noise points, vehicles, pedestrians, ground, etc.), average intensity, and other features. The spatial division feature data is a collection of the above-mentioned various features, not only retaining the original spatial position information of the point cloud, but also adding a comprehensive description of the points in each grid, thereby providing more rich and structured information for subsequent identification and classification.

[0045] Optionally, by performing non-uniform grid division in the height direction, higher resolution can be provided in the height section where geothermal gas noise points are frequent, helping the model to more accurately distinguish noise points from solid obstacles and reduce false positives. Non-uniform distribution means using larger grids in areas that do not require high resolution, such as the ground and distant obstacles, thereby reducing computational complexity, optimizing the use of computing resources, and making the entire system more efficient. The above target grid model can better adapt to the specific needs of mine unmanned driving scenarios, because the height distribution of noise points generated by geothermal gas is regular, and non-uniform grids can provide finer observation in these key height intervals.

[0046] In the embodiments of the present application, the spatial division feature data is generated by spatially dividing the detection results through the target grid model, which is the key to identifying and filtering random noise points generated by geothermal gas and improving the perception accuracy of mine unmanned vehicles. Combined with the characteristics of the target environment, intelligent non-uniform height grid division is achieved, achieving a good balance between recognition accuracy and computational resource optimization, thereby improving the overall system performance.

[0047] In step S106, the spatial division feature data is input into the non-interference entity recognition model for recognition, and the recognition result is obtained.

[0048] In the technical solution provided by step S106 in the embodiments of the present application, the recognition result can be used to indicate whether there is a non-interference entity in the target environment. The non-interference entity recognition model is a deep learning model that can be used to identify and classify entities in the target environment, especially to distinguish non-interference entities that cause false positives in the perception equipment of unmanned vehicles, such as random noise points generated by geothermal gas. In the embodiments of the present application, the non-interference entity recognition model can be a 3D segmentation model. The non-interference entity recognition model can receive spatial division feature data and determine the category of point cloud data in each spatial grid through its learned pattern recognition ability. The recognition result is obtained by the non-interference entity recognition model after analyzing the input spatial division feature data, reflecting the type distribution and location information of entities in the target environment, and can also be called a 3D segmentation result, which can clearly indicate which areas are entity obstacles and which are non-interference entities, such as random noise points generated by geothermal gas. The above recognition result is crucial for unmanned vehicles, as it directly affects the decision-making process of the vehicle, including path planning, obstacle avoidance, etc.

[0049] For example, in a high humidity environment, such as a mountainous area or a foggy day, the LIDAR is likely to identify water vapor particles in the air as obstacles. Although the signals reflected by the water vapor particles can cause abnormal points in the point cloud data, the water vapor particles do not affect the actual driving path of the vehicle. The non-interference entity identification model can distinguish the above-mentioned water vapor particles from actual obstacles to avoid false alarms. In winter, when the temperature is low, the temperature of the mined ore is high, which can generate geothermal gas. The random noise points caused by the above-mentioned geothermal gas appear as floating and widely distributed abnormal points in the point cloud data, which are likely to be misidentified as entity obstacles (such as the ground, retaining walls, etc.). The perception device of the mine unmanned vehicle needs to be able to accurately distinguish these noise points generated by geothermal gas from real obstacles. On construction, mining or dry unpaved roads, dust may be generated when the vehicle is driving. Particulate matter in the dust may also be captured by the laser radar sensor, causing abnormal points in the point cloud data. However, the above-mentioned dust does not constitute an actual obstacle to the vehicle driving, and the identification system needs to be able to mark it as a non-interference entity to avoid unnecessary path adjustment or deceleration. It should be noted that the above-mentioned non-interference entity is only an example and is not specifically limited here.

[0050] In this embodiment, after the detection structure is spatially divided according to the target grid model to obtain the spatial division feature data, the spatial division feature data can be input into the non-interference entity identification model to identify whether there is a non-interference entity in the target environment, and obtain an identification result.

[0051] Alternatively, the non-interference entity identification model can take the spatial division feature data, i.e., the features of each grid cell, as input to predict which category the feature belongs to. The category of the entity can include entity obstacles (such as vehicles, pedestrians, rock piles, boulders, etc.) and non-interference entities (such as random noise points generated by geothermal gas). The obtained identification result can not only be a category label, but also include a confidence score, which can represent the degree of certainty of the classification made by the non-interference entity identification model. The identification result is the final classification of the entity in the target environment by the non-interference entity identification model, which reflects the actual situation of the target environment, i.e., which areas have entity obstacles and which areas of the point cloud are composed of non-interference entities such as geothermal gas noise. Through the non-interference entity identification model, a category label is assigned to each grid, and confidence information is provided to help judge the reliability of the classification.

[0052] The embodiments of the present application will be described in detail below in combination with the above steps.

[0053] As an optional embodiment, the method further comprises: determining non-interference entity distribution attribute information corresponding to the target environment, wherein the non-interference entity distribution attribute information is used to represent the non-interference entity distribution characteristics in the case that the non-interference entity exists in the target environment; and determining the grid resolution of the target grid model in the height direction based at least on the non-interference entity distribution attribute information.

[0054] In this embodiment, the non-interference entity distribution attribute information can be used to represent the non-interference entity distribution characteristics in the case that the non-interference entity exists in the target environment, such as that the geothermal gas generally floats in the near area.

[0055] Optionally, the non-interference entity distribution attribute information refers to a detailed record of the distribution mode and characteristics of the non-interference entity (such as geothermal gas, fog, dust, etc.) in a specific target environment. The non-interference entity distribution attribute information can include key parameters such as the frequency, position, height, and density of the non-interference entity, which are very important for optimizing whether the non-interference entity exists in the environment around the unmanned vehicle. Based on the above non-interference entity distribution attribute information, the grid resolution of the target grid model in the height direction can be adjusted accordingly to improve the recognition accuracy of the non-interference entity and the efficiency of the overall perception system. That is, since different non-interference entities have different distribution characteristics, the target grid model that meets the corresponding distribution characteristics can be set accordingly.

[0056] For example, the random noise points generated by geothermal gas are mostly concentrated in the near-ground height and frequently appear in areas such as loading sites and dump sites where the temperature changes dramatically. Therefore, a higher grid resolution can be set for the above-mentioned areas and height sections. The water droplets in the fog are uniformly distributed, but there is a change trend with the height, and they are usually more concentrated within a certain height range above the ground. Therefore, for a fog environment, a finer grid resolution can be set within the corresponding height range. The distribution of dust is related to vehicle driving and wind direction, and is usually distributed near the vehicle driving track and above the dust source. For dust recognition, a finer resolution can be set at the height near the vehicle driving route.

[0057] Optionally, the non-interference entity distribution attribute information corresponding to the target environment is determined. The collection and analysis of this non-interference entity distribution attribute information can be mainly based on observation and data analysis under specific environmental conditions. For example, in the mine unmanned driving scenario, through continuous monitoring and recording of LIDAR data, it can be found that the random noise points generated by geothermal gas are mainly concentrated in the near area within a certain height range. The acquisition of this attribute information can be completed through statistical analysis of historical data, specific environment simulation test, or field measurement, etc.

[0058] Optionally, once the distribution characteristics of non-interfering entities are obtained, the grid resolution of the target grid model in the height direction can be adjusted according to the above distribution characteristics. The purpose of the above step is to optimize the processing of point cloud data, improve the recognition accuracy of non-interfering entities, and at the same time, reduce the consumption of computing resources as much as possible.

[0059] Optionally, according to the distribution characteristics of non-interfering entities, it can be determined that finer grid resolution is needed in a certain height interval to more accurately distinguish floating random noise points. In other height intervals (height regions) between the above-mentioned specific height intervals that need finer grid resolution, due to the low frequency of non-interfering entities, a larger grid size can be used, thereby reducing the data processing amount and computing cost without affecting the overall recognition accuracy. In the near area where non-interfering entities frequently occur, the target grid model can apply a super-resolution setting, which means that in this near area, the volume of the grid unit is smaller and the resolution is higher. Point cloud data can be analyzed in more detail to improve the ability to recognize geothermal gas noise points and reduce the false detection rate. The distribution of non-interfering entities may change over time and conditions, so the grid resolution can also be dynamically adjusted to achieve good recognition results in different times and environments.

[0060] As an optional embodiment, the target grid model has a non-uniformly distributed grid in the height direction, which meets the following conditions: the grid resolution of the grid in the first grid area in the height direction is higher than that of the grid in the second grid area in the height direction, and the first grid area is located closer to the ground area than the second grid area.

[0061] In this embodiment, in a laser radar perception system in a three-dimensional space, point cloud data can be processed into three-dimensional voxels or grids. The above-mentioned grid is to discretize the continuous three-dimensional space into small units with boundaries, which is convenient for calculation and analysis. The resolution of the grid model determines the size of each grid unit, which in turn affects the ability to capture the details of the point cloud data.

[0062] Optionally, the grid distribution of the target grid model in the height direction is non-uniform, which means that the grid resolution in some height regions (first grid area) is higher than that in other height regions (second grid area). Moreover, the first grid area is closer to the ground. This embodiment aims to improve the recognition accuracy of non-interfering entities in a specific height interval, while maintaining high efficiency in the use of computing resources. Since geothermal gas, fog, dust and other non-interfering entities often occur at a height close to the ground, by setting high-resolution grids in these areas, point cloud data can be analyzed in more detail, thereby reducing the false detection rate.

[0063] For example, the noise of geothermal gas mainly floats in the near-ground height, especially in the loading area. Therefore, the first grid area of the loading area needs higher resolution to accurately distinguish the geothermal gas noise and the solid obstacles and reduce the probability of mistaking the geothermal gas as a solid obstacle. The fog group often appears locally on the transportation road and is usually located near the vehicle driving height. Therefore, in the transportation road, the first grid area near the vehicle driving height needs higher resolution to accurately identify the influence of fog on the laser radar signal and avoid mistaking the fog as an obstacle.

[0064] For another example, dust (construction or non-paved road) often occurs during vehicle driving, especially in the construction area or dry non-paved road, and the distribution of dust particles is closely related to the vehicle motion trajectory. Therefore, the first grid area should have higher resolution near the height of the vehicle driving area in order to more effectively identify the point cloud anomalies caused by dust and reduce the false influence on the perception system. Unlike dust and geothermal gas, fog can be distributed in the entire height range, especially in low-visibility weather. Although the distribution of fog is not limited to the area near the ground, the grid resolution in the entire height range should be relatively high because of the universal and important influence of fog on the perception system.

[0065] Optionally, according to the specific environment where the vehicle is located and the real-time distribution of non-interference entities, the non-uniform distribution characteristics of the target grid model in the height direction are dynamically adjusted. For example, when entering the loading area, the grid resolution of the first grid area (near-ground height) can be automatically adjusted to be significantly improved. By combining high-resolution and low-resolution grid areas, effective identification of non-interference entities can be achieved without affecting the overall perception quality and computational efficiency. The high-resolution area is used for key environment monitoring, and the low-resolution area is used for general environment analysis to save computing resources.

[0066] As an optional embodiment, determining the non-interference entity distribution attribute information corresponding to the target environment includes: obtaining the business scenario or region information corresponding to the target environment; and determining the specified non-interference entity distribution attribute information corresponding to the business scenario or region information as the non-interference entity distribution attribute information.

[0067] In this embodiment, the business scenario or specific region of the target environment is identified and information is collected. The acquisition of the business scenario or region information can be based on the current position of the vehicle, time, work task and other key parameters. For example, in the mine unmanned driving scenario, the vehicle encounters different non-interference entities in different work stages (such as loading, transportation, and unloading) and locations (such as the loading area, the transportation road, and the dump). The above information can also be obtained through the positioning system of the vehicle, the vehicle-mounted environment monitoring system, and the pre-set business process data.

[0068] Optionally, once the business scenario or area information is obtained, it can be used to understand the non-interference entity distribution characteristics in the target environment, providing a basis for subsequent resolution adjustment. The analysis process includes but is not limited to: referring to historical operation records and perception data, understanding the frequency and distribution pattern of non-interference entities such as geothermal gas, fog, and dust in a specific scenario or area. Considering the natural conditions of the target environment such as temperature, humidity, and wind direction, the possible location and height of the non-interference entity are predicted. Based on the operation task of the unmanned vehicle, the types of non-interference entities that the vehicle may encounter and their height distribution characteristics during different task or operation stages are analyzed.

[0069] Optionally, after completing the in-depth analysis of the business scenario or area information, the specified non-interference entity distribution attribute information related to the above business scenario or area information is determined. It can be explicitly identified that in the current business scenario or area, which height and location are the active area of specific non-interference entities such as geothermal gas, fog, and dust. The distribution characteristics of non-interference entities in the active area, such as density, height range, and frequency of occurrence, are summarized. The above analysis results are integrated into the specified non-interference entity distribution attribute information, providing data support for adjusting the grid resolution in the next step.

[0070] Optionally, according to the determined specified non-interference entity distribution attribute information, the grid resolution of the target grid model in the height direction is adjusted to meet the needs of non-interference entity recognition. In the height range where non-interference entities are active, a higher grid resolution is set to improve the recognition accuracy; while in the non-active area, the resolution can be appropriately reduced to save computing resources.

[0071] For example, taking the loading area of a mine unmanned vehicle as an example, the vehicle performs mine loading operations in the loading area. After the loading area is located in underground mining operations, the temperature of the mine material is higher than the ambient temperature, and geothermal gas is particularly active in this area, mainly appearing at near-ground height. Through historical data review and environmental condition evaluation, it is found that the distribution density of random noise points generated by geothermal gas is highest at near-ground height (0-2m), and the frequency of occurrence increases during the loading process. According to the above analysis, an ultra-resolution grid is set in the area near the ground height (0-2m), and a lower resolution grid is set in the remaining height, thereby reducing the false detection rate of geothermal gas random noise points while ensuring efficient use of computing resources.

[0072] As an optional embodiment, determining the grid resolution of the target grid model in the height direction based on at least the non-interference entity distribution attribute information includes: determining the grid resolution of the target grid model in the height direction based on the non-interference entity distribution attribute and the computing resources of the unmanned vehicle, wherein the average value of the grid resolution in the height direction is positively correlated with the computing resources.

[0073] In this embodiment, the average value of the grid resolution in the height direction has a positive correlation with the computing resources, that is, the greater the resolution, the higher the computing resources.

[0074] Optionally, the distribution attribute information of non-interference entities in the target environment is analyzed in depth, including but not limited to the distribution density, height range, and appearance frequency of these entities in space, and other key features. The environmental conditions of a specific business scenario or area (such as a mine loading area, a city road construction area, etc.) are understood to accurately predict the active area and height of non-interference entities.

[0075] Optionally, the computing resources equipped on the unmanned vehicle are evaluated, including processor performance, memory capacity, power consumption limit, real-time data processing capability, etc. Considering the limitations of computing resources and the demand for real-time and accuracy of the perception system, a resource utilization strategy that meets the requirements is determined.

[0076] Optionally, the positive correlation between the average value of the grid resolution in the height direction and the computing resources is established, which means that higher resolution requires corresponding computing power support to avoid resource bottlenecks. Based on the distribution attributes of non-interference entities and the actual situation of computing resources, the grid resolution of the target grid model in the height direction is dynamically adjusted. In the active height range of non-interference entities, high resolution is set to enhance the recognition ability; in the inactive area, low resolution is used to save computing resources, achieving a balance between resources and performance.

[0077] Optionally, after determining the active height range of non-interference entities (such as geothermal gas), such as 0-2m, the grid resolution of the target grid model in this height range will be set to super-resolution to finely capture point cloud data and avoid false detection due to insufficient spatial granularity. In the case of limited computing resources, the grid resolution of non-critical areas can be appropriately reduced without affecting the recognition ability of critical areas to ensure the efficient operation of the entire system in processing large-scale point cloud data. When designing or adjusting the grid model, considering the positive correlation between computing resources and resolution, the computing load under different resolution settings can be evaluated to ensure that the real-time processing capability of the model matches the computing resources, avoiding delay or processing bottlenecks caused by excessive computing load.

[0078] For example, in the mine unmanned scenario, if the vehicle enters the loading area where geothermal gas frequently occurs, the grid resolution near the ground height can be automatically adjusted to significantly improve, and when the vehicle travels on the transportation road that is relatively clean and not affected by geothermal gas, the grid resolution can be appropriately reduced to save computing resources.

[0079] In the embodiments of the present application, by dynamically determining the resolution of the target grid model in the height direction based on the non-interference entity distribution attribute information and the actual situation of the computing resources, the performance and efficiency of the autonomous vehicle perception system can be effectively improved. The above strategy ensures that in the case of limited resources, high resolution can be run in key areas to improve the identification accuracy of non-interference entities, while reducing the resolution in non-key areas to reduce the computing load, achieving a balance between resources and performance. In implementation, environmental changes and computing resource states can be continuously monitored, and resolution settings can be adjusted in a timely manner to adapt to different business scenarios and environmental conditions.

[0080] As an optional embodiment, the non-interference entity identification model includes a segmentation model, which is trained by spatial division result samples of the target environment of the autonomous vehicle on an initial segmentation model. The spatial division result samples are obtained by spatial division of detection result samples in the target environment by multiple grid models, and different grid models have different grid resolutions in the height direction. In step S106, the spatial division feature data is input into the non-interference entity identification model for identification to obtain an identification result, including: inputting the spatial division feature data into the segmentation model for identification to obtain an identification result.

[0081] In this embodiment, the non-interference entity identification model can include a segmentation model. The segmentation model can be trained by spatial division result samples of the target environment of the autonomous vehicle on an initial segmentation model, which can be a Segmentation Head (SegHead) model. The spatial division result samples are obtained by spatial division of detection result samples in the target environment by multiple grid models. Different grid models have different grid resolutions in the height direction. The spatial division result samples can be a point cloud dataset. The multiple grid models can be used to train to improve the learning of height spatial features. The detection result samples can be point cloud data, images, or any perception data.

[0082] Optionally, the non-interference entity identification process mainly relies on a special segmentation model, which is trained by using spatial division result samples of the target environment of the autonomous vehicle, and can accurately identify non-interference entities.

[0083] Optionally, the generation of spatial division result samples relies on detection result samples collected in the target environment, which are then spatially divided by multiple grid models. Each grid model has different grid resolutions in the height direction, meaning that the grid model can capture detailed information at different height levels. The application of multiple grid models ensures the diversity and richness of the spatial division result samples, covering the feature representation of non-interference entities at different height levels, providing a comprehensive data basis for subsequent model training.

[0084] Optionally, the spatial division result sample constitutes a basic data set for training the segmentation model, and the spatial division result sample contains point cloud data sets captured at different height resolutions and the results after spatial division of the data sets. The segmentation model is obtained by continuously optimizing the initial state through learning of the spatial division result sample. During the training process, the model gradually learns how to distinguish between real entity obstacles and non-interference entities such as geothermal gas, fog, or dust. By using multiple grid models with different height-direction grid resolutions, the training process can significantly improve the model's learning of height spatial features, especially in height regions that are prone to non-interference entities. After the segmentation model is trained, its structure is optimized to more efficiently handle complex features related to non-interference entities while reducing the consumption of computing resources.

[0085] Optionally, during the identification process of the segmentation model, the spatial division feature data is directly input into the trained segmentation model. The task of the segmentation model is to identify the distribution of non-interference entities and possible entity obstacles based on the input spatial division feature data. It is worth noting that the segmentation task does not require inference of 3D detection or other perception tasks, which means that during model inference, only the segmentation task can be focused on, thereby significantly saving computing resources.

[0086] In the embodiments of the present application, the spatial division result sample makes full use of the grid resolution in different height directions, which helps to comprehensively capture the distribution characteristics of non-interference entities in three-dimensional space and provide more abundant training materials for the segmentation model. By training the segmentation model at different height resolutions, the generalization ability of the model can be improved, making it more adaptable and accurate in identifying non-interference entities in different scenarios. Focusing on the segmentation task during model inference rather than performing redundant detection tasks can effectively save computing resources, which is particularly important for resource-constrained autonomous vehicles.

[0087] As an optional embodiment, the method further comprises: determining detection labels and segmentation labels in the detection result sample in the target environment, wherein the detection labels are used to represent the detection boxes labeled for the object samples in the target environment, and the segmentation labels are used to represent the classification results labeled for the object samples; and training (multi-task joint training, training of multi-task model structure at the same time) the initial detection model and the initial segmentation model based on the detection labels, the segmentation labels, and the multiple grid models to obtain the non-interference entity identification model, wherein the different grid models have different grid resolutions in the height direction.

[0088] In this embodiment, the detection label can be used to represent the bounding box labeled for the object sample in the target environment, which can be a real labeled 3D bounding box. The object sample can be a foreground target in the environment with a labeled ground truth. The detection box can be a 3D detection box (3D detection Box). The segmentation label can be used to represent the classification result labeled for the object sample, which can be labeled as a non-interference entity or an entity obstacle. Different grid models have different grid resolutions in the height direction.

[0089] Optionally, the detection label is a detection box labeled for the object sample in the target environment, which is usually a 3D bounding box, used to mark foreground targets in the environment, such as pedestrians, vehicles, obstacles, etc. The segmentation label is the result of classifying the object sample, indicating whether a point cloud data belongs to an entity obstacle or a non-interference entity (such as geothermal gas, fog, dust). Different grid models have different resolutions in the height direction, which means they have different levels of refinement when processing point cloud data at different height levels. The resolution in the height direction is particularly important when dealing with non-interference entities, as it helps to distinguish non-interference entities floating at different heights from entity obstacles.

[0090] Optionally, based on the detection label and the segmentation label, the initial detection model and the initial segmentation model are trained in a multi-task joint manner using multiple grid models. The above training method can enable the model to learn object detection and classification capabilities simultaneously, especially to better learn the distribution characteristics of non-interference entities in the height direction. The goal of the training process is to enable the model to distinguish entity obstacles (such as rock piles, boulders, etc.) from non-interference entities (such as geothermal gas, fog, etc.) from point cloud data, and to strengthen this ability through grid models with different height resolutions.

[0091] Optionally, before training, the detection result samples are preprocessed to ensure that each sample is attached with accurate detection labels (3D bounding boxes) and segmentation labels (classification results). It is very important to ensure the consistency of the detection label and the segmentation label, that is, the point cloud data in each detection box should be consistent with the segmentation label of the box. This requires manual or semi-automatic inspection and correction of the data set. During the training process, the architecture of the model may need to be adjusted to adapt to the different resolutions of the multiple grid models. For example, introduce advanced modules such as deformable convolution and attention mechanism to enhance the model's sensitivity to different height details. The limitation of computing resources and the efficiency of model inference also need to be considered during the training process. Although high-resolution grid models can help improve recognition accuracy, excessive refinement may increase the computational burden, so a balance between accuracy and efficiency needs to be found. The training process can be an iterative process, including model training, validation, and adjustment. After each iteration, the performance of the model is tested using the validation set, and the model parameters and structure are further optimized based on the results.

[0092] As an optional embodiment, the object samples include obstacle samples and non-interference entity samples, and the initial detection model and the initial segmentation model are trained based on the detection label, the segmentation label and the plurality of grid models to obtain a non-interference entity recognition model, including: training the initial detection model and the initial segmentation model by using the detection label, the segmentation label and the plurality of grid models; adjusting the trained initial detection model and the trained initial segmentation model by using attribute information of the obstacle samples and noise point information of the non-interference entity samples to obtain the non-interference entity recognition model.

[0093] In this embodiment, the attribute information can include information such as the position, direction and size of the obstacle. This embodiment describes how to train and optimize the non-interference entity recognition model by using obstacle samples and non-interference entity samples, combining a multi-task learning framework of detection and segmentation tasks, and different height resolution grid models.

[0094] Optionally, the samples in the point cloud dataset are divided into two categories: obstacle samples and non-interference entity samples. The obstacle samples include actual physical obstacles such as stones, stone piles, vehicles, etc., while the non-interference entity samples refer to noise points generated by environmental factors such as geothermal gas, fog, dust. A multi-task learning framework is adopted, which includes an initial detection model and an initial segmentation model. These two models are responsible for 3D detection and 3D segmentation tasks respectively. The detection label and the segmentation label of the complete dataset are used to synchronously train the multi-task model structure. The purpose of this is to enable the model to learn the spatial distribution patterns of obstacles and non-interference entities in the early stage.

[0095] Optionally, after the preliminary training is completed, the initial detection model and the initial segmentation model are further fine-tuned by using a radar training set containing only obstacles (such as stone piers and stones) and non-interference entities (such as geothermal gas). The position, direction, size and other attribute information of the obstacle samples and the noise point information of the non-interference entity samples are combined to guide the adjustment of the model parameters, so as to enhance the performance of the model in identifying non-interference entities.

[0096] Optionally, during the initial model training stage, the detection label and the segmentation label are used simultaneously, which helps the model to understand the geometric morphology of the obstacles and the point cloud classification knowledge, laying a solid foundation for subsequent fine-tuning. By applying grid models with different height resolutions, the model can learn the distribution characteristics of different obstacles and non-interference entities in the height direction, which is particularly important in dealing with height-related noise such as geothermal gas. In the fine-tuning stage, the focus is on training using specific samples (obstacles and non-interference entities) to enable the model to focus more on solving the non-interference entity recognition problem in specific environments and improve the recognition accuracy.

[0097] Optionally, the radar training set used in the fine-tuning stage should be carefully selected from those samples that best reflect the impact of non-interference entities, such as the frequently occurring loading area point cloud data of geothermal gas, to ensure that the model can fully learn the skills of distinguishing entities from noise points in such environments. During the fine-tuning process, the adjustment of model parameters should focus on improving the ability to identify non-interference entity noise, especially those noise points at specific height distributions, such as ensuring that the model has higher sensitivity to geothermal gas noise at near-ground height. When designing and training the model, the computational resource limitations of autonomous vehicles should be considered to make the model as accurate and lightweight as possible, enabling real-time perception and rapid decision-making.

[0098] As an optional embodiment, the attribute information of the obstacle samples and the noise point information of the non-interference entity samples are used to adjust the trained initial detection model and the trained initial segmentation model to obtain a non-interference entity recognition model, including: using the attribute information of the obstacle samples and the noise point information of the non-interference entity samples to adjust the trained initial detection model and the trained initial segmentation model to obtain a target detection model and a target segmentation model; using the attribute information of the obstacle samples and the noise point information of the non-interference entity samples to adjust the target segmentation model to obtain a segmentation model; and constructing the segmentation model and the target detection model as the non-interference entity recognition model.

[0099] In this embodiment, the initial detection model and the initial segmentation model are adjusted and optimized based on the attribute information of the obstacle samples and the noise point information of the non-interference entity (such as geothermal gas) samples. The key to the above steps is that the attribute information of the obstacle samples includes the position information, direction information, size information, category information of the obstacles, and the mutual relationship information between the obstacles and the non-interference entities. The above attribute information helps the model to identify the geometric features and spatial layout of the target, thereby more accurately performing object detection and classification. The noise point information of the non-interference entity samples mainly focuses on the manifestation of the non-interference entity in the point cloud data, including its distribution law (such as height, distance) relative to the obstacles, morphological features (such as dispersion degree, density), and the difference between the noise points and the actual obstacle point cloud. Through learning of the noise point information, the model can gradually distinguish between real obstacles and non-interference entities, reducing false positives.

[0100] Optionally, the initial detection model and the initial segmentation model are further adjusted and fine-tuned using the attribute information of the obstacle samples and the noise information of the non-interfering entity samples: through continuous training, the model can more accurately locate and identify obstacles, while learning to ignore the noise generated by non-interfering entities. The adjustment process may involve modifying the network structure, optimizing the loss function or adjusting the hyperparameters to improve detection accuracy and robustness. In particular, the radar training set containing only obstacles such as stone blocks and geothermal gas is used to fine-tune the model specifically for the segmentation task. The above process helps the model establish clearer boundaries between specific obstacles and non-interfering entities, improving the accuracy of segmentation.

[0101] Optionally, the adjusted target detection model and segmentation model are combined to form a non-interfering entity recognition model. The purpose of the above model design is to ensure that the model can accurately identify entity obstacles while effectively distinguishing noise points such as geothermal gas in the segmentation task, avoiding misidentification as obstacles. The detection model is responsible for locating obstacles, while the segmentation model further refines the classification boundaries between obstacles and background, noise. Through model integration, comprehensive perception of the environment can be achieved, reducing false positives and missed detections, and improving the safety and reliability of unmanned vehicles.

[0102] Optionally, during the fine-tuning phase, it is crucial to select radar training sets containing specific obstacles and non-interfering entities. These data sets should cover various possible environmental conditions to ensure that the model has good generalization ability. A multi-stage training strategy is adopted, using a complete data set for preliminary training and a specific data set for fine-tuning. The above strategy helps the model to further learn and refine identification skills in specific environments based on mastering basic perception capabilities. During the training process, performance indicators such as accuracy, recall rate and F1 score should be continuously monitored to ensure that the model maintains or improves overall recognition ability after adjustment, especially in terms of non-interfering entity recognition.

[0103] As an optional embodiment, the initial detection model and the initial segmentation model are trained using detection labels, segmentation labels, and multiple grid models, including: determining a first loss between the output result of the initial detection model and the detection label in the detection result sample, and a second loss between the output result of the initial segmentation model and the segmentation label in the detection result sample; using multiple grid models to divide the space of the detection result sample to obtain a spatial division result sample; training the initial detection model using the first loss and the spatial division result sample, and adjusting the gradient corresponding to the second loss using the first loss, and training the initial segmentation model using the adjusted gradient and the spatial division result sample.

[0104] In this embodiment, the output result can be a 3D detection box. The first loss can be a loss function of the 3D detection box. In the process of training the initial detection model and the initial segmentation model, it is crucial to determine the loss between the model output and the actual label.

[0105] Optionally, the difference between the output result (i.e., the 3D detection box) of the initial detection model and the true detection label given in the detection result sample is calculated to obtain the first loss function. A 3D detection loss function can be used to quantify the degree of coincidence between the model prediction box and the actual box. The first loss not only guides the learning of the detection task, but also serves as the basis for adjusting the gradient of the segmentation model. The difference between the output result of the initial segmentation model for 3D spatial segmentation of the point cloud data and the actual segmentation label is calculated to obtain the second loss function, which is used to supervise the learning of the segmentation task. The segmentation loss is used to ensure that the model can accurately distinguish different objects or backgrounds in the point cloud.

[0106] Optionally, in order to improve the sensitivity and learning ability of the model to spatial features, multiple grid models with different height resolutions are used to divide the space of the detection result sample (i.e., the point cloud data containing objects and environmental noise): spatial division, the point cloud data is mapped to multiple grid models, each model has different grid resolutions in the height direction, some may be more fine-grained (such as 0.05m), and some may be more coarse-grained (such as 0.2m). By voxelizing the data at different resolutions, spatial division result samples can be generated, which contain the distribution and features of the point cloud at different height resolutions.

[0107] Optionally, the initial detection model is trained using the first loss (detection loss) and the spatial division result sample. The training here not only involves the adjustment of the parameters of the detection model itself, but also affects the training process of the segmentation model through the first loss, ensuring that the two models can be optimized collaboratively. The first loss is used to adjust the gradient corresponding to the second loss (segmentation loss). In the backpropagation process, the loss function of the detection task will simultaneously affect the weight update of the detection model and the segmentation model. In this way, even if the detection task is completed, the gradient of its loss function can still guide the learning of the segmentation model, helping the model to distinguish obstacles from non-interfering entities, such as noise generated by geothermal gas.

[0108] Optionally, the loss function selection of the first loss and the second loss can meet the requirements of the respective tasks while ensuring that the loss function can effectively guide the learning of the model. For example, the loss function of 3D detection may need to take into account the geometry of the frame, while the loss function of 3D segmentation may focus more on the accuracy of point cloud classification. Under the multi-task learning framework, the models of the detection task and the segmentation task are trained together, realizing the sharing of parameters, which can reduce the complexity of the model, improve the training efficiency and generalization ability, and at the same time enhance the adaptability of the model to complex environments through mutual supervision. By adjusting the grid model of different height resolutions, a balance point between computing resources and model performance can be found to avoid loss of recognition accuracy due to computing resource limitations.

[0109] Figure 2 is a flowchart of a non-interference entity detection method according to an embodiment of the present application, as shown in Figure 2 The method comprises the following steps:

[0110] Step S202, detecting the business state of the unmanned vehicle and / or the area where it is located.

[0111] In this embodiment, the current business state of the unmanned vehicle and the specific environmental area where it is located are evaluated to determine whether it is necessary to start a special target perception mode. This step embodies an important feature of the intelligent perception system, i.e. situational awareness and decision making. The business state can refer to the operation or task the vehicle is currently performing, such as loading, unloading, driving, parking, etc. Detecting the business state means that the system has the ability to identify the immediate needs of the vehicle, which is crucial for activating a specific perception mode in a specific situation. In addition to the business state, the location or area where the vehicle is located, such as the loading position, the dump site, the road, etc. can also be identified. This is because the influence of non-interference entities such as geothermal gas is usually closely related to specific geographical environments, and some areas (such as places with frequent geothermal activity) are more likely to encounter the above entities than other areas. By combining the information of the business state and the area where it is located, a comprehensive judgment can be made. For example, if the vehicle is in the loading position and is performing a loading task, the influence of geothermal gas on the perception system can be considered, and appropriate measures can be taken accordingly.

[0112] Step S204, in the case where the business state and / or the area where it is located meet the target conditions, controlling the unmanned vehicle to start a target perception mode, wherein the target perception mode is used to detect whether there is a non-interference entity in the current environment.

[0113] In this embodiment, once the detected business state and the region where the vehicle is located meet the set target conditions, the system will take action and start the target awareness mode. The target awareness mode is specifically designed to detect whether there are non-interfering entities in the current environment, such as noise points generated by geothermal gas, which is crucial for reducing false detections and improving the accuracy of perception. The target condition refers to a series of pre-set trigger conditions. For example, when the vehicle is located in a known region where geothermal gas frequently occurs, and is performing a business operation that may be affected by geothermal gas (such as loading), it is considered to meet the target condition. In the target awareness mode, the focus is on detecting non-interfering entities, which can include more detailed analysis of point cloud data or the use of identification algorithms specific to this type of entity. The goal is to effectively filter or mark non-interfering entities, such as abnormal point clouds generated by geothermal gas, without affecting normal obstacle detection.

[0114] The embodiments of the present application will be described in detail below in combination with the above steps.

[0115] As an optional embodiment, the target condition includes at least one of the following: in the case of detecting the business state of the unmanned vehicle, the target condition is that the business state is a loading business state, a dumping business state, or a road transportation state; in the case of detecting the region where the unmanned vehicle is located, the target condition is that the region is a loading position region, a dumping position region, or a transportation road region.

[0116] In this embodiment, the determination of the target condition is directly related to the business state of the unmanned vehicle and the environment region where it is located. The business state includes but is not limited to the loading business state, the dumping business state, and the road transportation state, which correspond to the running modes of the vehicle in different operation scenarios. Specifically, the influence of geothermal gas and dust.

[0117] For example, for the loading business state, in coal mining business, especially when excavating in winter, the generation of geothermal gas becomes a significant environmental factor. Due to the high temperature of the mined material during the excavation process and the low temperature of the external environment, the temperature difference effect leads to the release of a large amount of geothermal gas. The above-mentioned geothermal gas appears as abnormal noise points in the point cloud data, which can interfere with the normal operation of the perception system and cause false identification of entity obstacles. Therefore, when the vehicle is detected to be in the loading business state, it will be evaluated whether to start the target awareness mode to deal with the interference caused by geothermal gas and ensure the accuracy of perception. For the dumping business state, dumping operations are usually accompanied by a large amount of dust, especially in dry or windy conditions. Dust can also generate point cloud noise similar to geothermal gas, affecting the performance of 3D segmentation and obstacle detection. When the dumping business state is identified, it can be checked whether the conditions for starting the target awareness mode are met to reduce the negative impact of dust on the perception results.

[0118] Optionally, the target condition also relates to the environmental area where the unmanned vehicle is located, including the loading site area, the dumping site area, and the transportation road area. The above-mentioned areas are usually closely related to specific business states, and each has different environmental characteristics and potential interference sources. In the loading site area, the vehicle performs the mineral loading operation, which is a scene where geothermal gas frequently occurs. When it is detected that the vehicle is located in the loading site area, it will be evaluated whether the condition for starting the target perception mode is met, so as to more accurately filter the noise generated by geothermal gas and reduce the misidentification of solid obstacles (such as stone piles and stone blocks). In the dumping site area, the dumping site is where the vehicle unloads the mineral, and a large amount of dust will be generated during the operation. When it is identified that the vehicle is located in the dumping site area, the environmental conditions can be checked to determine whether the target perception mode needs to be started to enhance the filtering ability of the system to the dust noise and ensure safe unloading. In the transportation road area, the vehicle may encounter various obstacles, and the influence of geothermal gas and dust is relatively small. However, the environmental changes during transportation can still be monitored to identify whether the perception mode needs to be adjusted to deal with potential obstacles or unexpected interference.

[0119] Optionally, by continuously monitoring the business state of the vehicle and the environmental area where the vehicle is located, based on the preset target condition, it is intelligently judged whether to start the target perception mode. The starting of the above-mentioned target perception mode will activate a special algorithm or adjust the parameters of the existing algorithm to more effectively identify and filter non-interference entities such as noise points generated by geothermal gas and dust. Different business states and environmental areas require different perception strategies. For example, in the loading site where geothermal gas frequently occurs, the perception system may use a higher resolution grid model and stricter noise point filtering rules; while in the dumping site area, it may focus more on the filtering of dust and the fine identification of solid obstacles. In the target perception mode, the point cloud data can be checked more carefully, and the point cloud characteristics of each solid obstacle, such as the number of points, the height distribution, and the confidence, are used to judge whether there is a false detection of non-interference entities. This checking mechanism provides more reliable data support for the decision of the vehicle, which helps to reduce unnecessary stopping, taking over or obstacle avoidance, and improves the safety and efficiency of the operation.

[0120] As an optional embodiment, the target perception mode is used to identify the spatial division feature data of the unmanned vehicle by using a non-interference entity identification model to obtain an identification result, the identification result is used to represent whether there is a non-interference entity in the current environment, and the spatial division feature data is obtained by spatially dividing the detection result of the unmanned vehicle according to a target grid model, the target grid model has non-uniformly distributed grids in the height direction and uniformly distributed grids in the horizontal direction.

[0121] In this embodiment, the target awareness mode utilizes a non-interference entity recognition model to recognize the spatial partition feature data of the autonomous vehicle to determine whether there is a non-interference entity in the current environment. The spatial partition feature data is obtained by mapping the detection results of the autonomous vehicle to a specific target grid model. The target grid model has the following characteristics: non-uniform distribution in the height direction, that is, the target grid model uses non-uniformly distributed grids in the height direction, that is, the resolution of the grid is higher in some height intervals, and the resolution is lower in other intervals. The purpose of the above design is to better capture and distinguish entities and non-interference entities appearing at different heights, such as noise points generated by geothermal gas. Because in the area where geothermal gas frequently occurs, noise points often appear in specific height intervals, by increasing the resolution of these areas, the sensitivity of the model to height features can be enhanced, and false detection can be reduced. Uniform distribution in the horizontal direction, that is, in contrast to the non-uniform distribution in the height direction, the grid distribution in the horizontal direction is uniform, which means that in the x and y axes, the size and distribution of the grid are fixed. This ensures the consistency of the model in the horizontal direction, maintains the stable recognition ability of the obstacle plane distribution, and simplifies the calculation and reduces resource consumption.

[0122] Optionally, the detection results of the autonomous vehicle, such as 3D point cloud data, are voxelized according to the target grid model to obtain spatial partition feature data. These feature data contain the distribution of point clouds at different heights and the uniform grid representation in the horizontal plane. The non-interference entity recognition model extracts key features such as the density, distribution pattern, and relationship with entity obstacles based on the input spatial partition feature data. These features help the model distinguish between entity obstacles and non-interference entities such as noise points caused by geothermal gas or dust. Through analysis of the input data, the recognition result is output to indicate whether there is a non-interference entity in the current environment. The above recognition result is crucial for the decision of the autonomous vehicle, such as whether to take additional obstacle avoidance or speed reduction measures when encountering a loading site or a dump site area where geothermal gas frequently occurs.

[0123] As an optional embodiment, the method further comprises: detecting the current environment by the perception device of the autonomous vehicle to obtain a perception data set; recognizing the spatial partition feature data of the autonomous vehicle by using the non-interference entity recognition model to obtain a recognition result; obtaining a classification result corresponding to the perception data set and a confidence corresponding to the classification result according to the recognition result; and outputting a target verification result corresponding to the classification result and the confidence, wherein the target verification result includes a first verification result for indicating that the classification result is a false detection.

[0124] In this embodiment, the perception devices on the autonomous vehicle, such as LiDAR, camera, millimeter wave radar, etc., conduct comprehensive detection of the current environment, collecting point cloud data, image data or other sensor data in the environment to form a perception dataset. The above perception dataset is the basis for the autonomous vehicle to understand the environment, identify obstacles and plan a path. The non-interference entity recognition model identifies the spatial division feature data in the perception dataset, which is obtained by spatial division and feature extraction of the point cloud data through the target grid model. The task of the non-interference entity recognition model is to distinguish between entity obstacles (such as rocks, vehicles) and non-interference entities (such as noise points generated by geothermal gas) in the point cloud and output the recognition result. The non-interference entity recognition model can analyze the point cloud features in each grid, including but not limited to the density, distribution pattern of points, relative position to known entity obstacles, etc., and classify each grid using trained model weights and algorithm logic. Accurate identification and differentiation of entity obstacles and non-interference entities, such as noise points generated by geothermal gas, can reduce false positives and ensure the safe and efficient operation of the autonomous vehicle.

[0125] Optionally, the recognition result includes the classification result of each grid or point cloud region, as well as the confidence corresponding to the classification result. The confidence reflects the degree of certainty of the model for the classification result, and is a value between 0 and 1. The larger the value, the higher the confidence of the model. The non-interference entity recognition model classifies each grid as an entity obstacle, a non-interference entity, or other categories such as background. For each classification result, the model also outputs a confidence. By analyzing the classification results and confidences of different grids, it can be determined which regions have high-confidence entity obstacles and which regions have classification results that may be affected by noise points and have low confidence.

[0126] Optionally, according to the classification result and the confidence, a target verification result is output for further analysis and decision-making. The target verification result can include multiple types, but the first verification result for indicating a false positive in the classification result is particularly mentioned: checking the grids with low confidence in the classification result, especially those grids that are classified as entity obstacles but may actually be non-interference entities. If it is found that the confidence of the classification result of a certain grid is lower than a predetermined threshold, and the grid contains a large number of possible noise points (such as those caused by geothermal gas), the system will mark the classification result as a false positive. When a classification result is determined to be a false positive, i.e., the grid is incorrectly identified as an entity obstacle rather than an actual non-interference entity, the first verification result is output. This result is particularly important because it directly relates to the safety and operational efficiency of the autonomous vehicle, such as unnecessary obstacle avoidance or stopping due to frequent false positives.

[0127] The above technical solutions are described in detail below.

[0128] FIG. 3(a) is a schematic diagram of a 3D segmentation effect in the related art, as shown in FIG. 3(a), a perception segmentation model has important significance in the field of unmanned driving, and the perception inference calculation is performed on the real-time data of a LIDAR (laser radar), and the category (such as a retaining wall, a vehicle, a ground, dust, a pedestrian, etc.) of each point of the LIDAR point cloud data is judged.

[0129] FIG. 3(b) is a schematic diagram of a 3D segmentation false detection of retaining walls, grounds and other solid substances caused by geothermal in the related art, as shown in FIG. 3(b), a mine unmanned driving scene usually faces the challenge of a harsh environment, especially the influence of geothermal gas. The geothermal gas is generated due to the low temperature in winter and the high temperature after the mine material excavator is excavated. The geothermal gas seriously affects the performance of the laser radar sensor, causes a large number of abnormal irregular random noise points in the laser radar data, causes the false detection of the LIDAR perception 3D segmentation, and thus causes the geothermal gas to be falsely detected as a solid substance, such as a ground 301, a retaining wall 302, an obstacle, etc. The downstream algorithm for clustering the semantic point cloud is falsely detected as a solid obstacle. Thus, the unnecessary abnormal braking, obstacle avoidance or takeover of the mine unmanned vehicle is caused, the frequent braking causes damage to the vehicle, a large number of takeovers affect the running efficiency of the unmanned driving, and the obstacle avoidance detour for the false detection obstacle easily causes a safety risk. Continuous observation can determine that the floating discrete points are all noise points 303, which are easily falsely detected as solid substances. The mainstream LIDAR perception 3D segmentation method mainly uses BEV, and each spatial grid (voxel) or column (pillar) independently outputs one category, such as a ground, a retaining wall, a noise point, a pedestrian, a vehicle, etc. The spatial fine-grained splitting is not enough, and different types of point clouds are output according to the same one type, such as the noise points and the ground in the voxel. The point clouds of all voxels may be uniformly output as the ground, thus causing the false detection of the noise points. In actual observation, there is a certain interval between the geothermal gas noise points and the solid medium.

[0130] The mine unmanned environment is prone to produce geothermal gas in winter due to low temperature and high temperature of mine materials, which affects the sensor performance of LIDAR, thereby affecting the 3D segmentation perception ability. The geothermal gas causes a large number of abnormal irregular noise points of LIDAR, which are floating, have a large distribution range, and are easy to adhere to real barrier wall ground obstacles and other medium entities. The LIDAR 3D segmentation perception is prone to misjudge the noise points as ground, barrier wall, obstacle and other medium entities, thereby causing misjudgment of obstacle types such as stone piles, affecting the operation efficiency of mine unmanned driving, and causing wear of vehicle tires and other parts due to frequent braking. Considering the 3D segmentation misjudgment caused by voxel fine granularity, a non-uniform variable resolution grid method is adopted to perform super-resolution splitting on the height space, thereby reducing the medium entity misjudgment caused by geothermal gas noise points due to insufficient spatial fine granularity. Meanwhile, the variable resolution grid balances between calculation resources and inference accuracy, thereby improving the feasibility of deployment. Considering the 3D segmentation misjudgment caused by the upper limit of the 3D segmentation model itself, a multi-task training is adopted to increase the loss of 3D detection Box to synchronously supervise the training, thereby avoiding misjudgment of geothermal gas noise points as medium entities such as stone piles. Considering the 3D segmentation misjudgment still existing during inference, the medium entities and noise points in specific regions such as loading positions and dumping sites are verified.

[0131] The embodiment of the present application proposes a mine unmanned LIDAR perception geothermal gas misjudgment method, which optimizes the resolution of height space by adopting a non-uniform variable resolution grid method, thereby improving the ability of the model to distinguish geothermal gas floating noise points and entity obstacles. In addition, a multi-task training strategy is introduced, which synchronously guides the training of the 3D segmentation model by using the loss function of 3D detection Box, thereby strengthening the accuracy of the model in identifying obstacles such as stone piles and stone blocks, and reducing the misidentification of geothermal gas as obstacles. Finally, the scheme also considers the characteristics of specific business states and regions, designs a target perception mode, and further filters possible misjudgments through a verification mechanism to ensure the reliability of the identification result in regions such as loading positions and dumping sites where geothermal gas frequently occurs. Overall, the present scheme shows innovation and practicality in improving perception accuracy, avoiding misjudgment and adapting to complex environments, provides strong technical support for the safe operation of mine unmanned driving, realizes the technical effect of improving the detection accuracy of non-interference entities, and solves the technical problem of low detection accuracy of non-interference entities.

[0132] The method will be further introduced below.

[0133] A plurality of different types of laser radars are installed on a mine unmanned operating vehicle to obtain LIDAR data in real time. Figure 4 is a schematic diagram of a grid structure according to an embodiment of the present application, as Figure 4As shown, the LIDAR output can be processed into this grid structure for computation.

[0134] Figure 5 is a flowchart of a method for solving the problem of false detection of geothermal gas by LIDAR perception in mine unmanned driving according to an embodiment of the present application, as shown, the method can include the following steps: Figure 5

[0135] Step S501, acquire perception data. In this embodiment, point cloud data of the surrounding environment is collected. The above-mentioned data contains distance, angle and intensity information of objects around the vehicle, which is the basis for subsequent processing and analysis.

[0136] Step S502, perform non-uniform voxel division on the perception data to obtain grid data with non-uniform resolution. In this embodiment, the point cloud data is converted into grid data through non-uniform voxel division. Compared with traditional uniform division, non-uniform voxel division can flexibly adjust the resolution according to the importance of spatial features, especially in the height direction, using super-resolution division. This process increases the sensitivity of the model to height space details, helping to more finely distinguish between ground and high thermal gas noise, reducing false detection.

[0137] Step S503, send the grid data into the backbone network to learn spatial layout and object category information. In this embodiment, the voxelized grid data is sent into the backbone network, which is the core part of a deep learning architecture, responsible for extracting spatial features and category information of the input data. The backbone network abstracts and processes data through multiple layers of neural networks to generate highly condensed feature representations, which are crucial for subsequent detection and segmentation tasks. By applying super-resolution in the height direction, the backbone network can better understand and learn the distribution patterns of point cloud data in three-dimensional space, improving its adaptability to complex environments.

[0138] Step S504, classify voxels according to the extracted information using the segmentation head model. In this embodiment, the segmentation head model (SegHead) receives features from the backbone network and focuses on classifying each voxel. Its goal is to classify voxels into specific categories, such as ground, vehicle, pedestrian, obstacle or noise, etc. SegHead can convert feature maps into class predictions for each voxel based on deep learning techniques such as fully convolutional networks (FCN), a process called 3D semantic segmentation.

[0139] ​Step S505, output the type of voxel. In this embodiment, after the processing of the segmented head model, the final output is the classification of the type of each voxel, accompanied by the corresponding confidence score. These information provides detailed classification and location information of objects in the entire perception environment, which helps the unmanned system to make accurate decisions. For example, if a certain voxel is classified as a noise point and the confidence is lower than the set threshold, it can be considered as a false detection caused by geothermal gas, so as not to trigger unnecessary obstacle avoidance or stopping action of the vehicle, reducing the operating cost and improving the safety. In order to solve the false detection of 3D segmentation caused by the fine granularity of voxel, a non-uniform zoom grid method is adopted to finely divide the height space, reduce the false detection of medium entity caused by the lack of spatial fine granularity of geothermal gas noise point. At the same time, the variable resolution grid is adopted to balance between computing resources and inference accuracy, and improve the feasibility of deployment.

[0140] The LIDAR perception 3D segmentation method is mainly based on BEV, and each spatial grid (voxel) or pillar outputs 1 category, such as ground, barrier, noise, pedestrian, vehicle, etc. The spatial fine granularity is not enough, which may cause different types of point clouds to be output as the same type, such as noise and ground in voxel, which may output all voxel point clouds as ground, resulting in false detection of noise. In actual observation, there is a certain interval between geothermal gas noise and entity medium.

[0141] The 3D segmentation voxel is uniformly divided in x, y and z directions, and the voxelization operation (Voxelize) is performed according to the uniform voxel to generate the input of the corresponding Backbone network, for example, the perception range is X: (-50~50m), Y: (-50~50m), Z: (-6~6m), the size of the grid voxel is (X: 0.1m, Y: 0.1m, Z: 0.2m), and the generated feature output dimension is (1000*1000*60). The fine granularity of the mine unmanned scene may cause the model to learn insufficient characteristics of the height, and the same grid (multiple radar points) outputs one category, so as to easily identify the floating and chaotic geothermal gas noise as entity medium (including obstacles such as stone piles and blocks, ground, barrier, etc.).

[0142] In this embodiment, the uneven variable resolution grid method is adopted, and in the model training stage, the high uneven grid is divided as follows, and the high variable spatial height grid is changed to, for example, 0.2 m, 0.1 m, or even super-resolution 0.05 m, and the spatial feature is changed to three groups of features (1000*1000*60), (1000*1000*120, 1000*1000*240), which can be used together to train the learning of the high spatial feature of the promotion model, so as to have the ability to distinguish the geothermal gas floating noise.

[0143] Figure 6 is a schematic diagram of a high uneven zoom grid structure according to an embodiment of the application, as Figure 6 As shown in the model inference stage, the super-resolution output grid result can ensure the geothermal gas recognition accuracy, but often the hardware computing device resources of the unmanned vehicle are limited and cannot support the super-resolution output. In order to balance between the computing resources and the super-resolution output capability, the high uneven zoom method is adopted, that is, the high altitude interval (such as the near area) which needs to guarantee the accuracy adopts super high resolution, and the non-core area adopts larger resolution, and can be selected and matched according to the split in the training. Under the limited computing resources, the 3D segmentation accuracy is ensured, so that the model structure corresponding to the geothermal gas can be deployed on the unmanned vehicle.

[0144] In order to solve the 3D segmentation false detection caused by the upper limit of the 3D segmentation model itself, multi-task training is adopted, and the loss of 3D detection Box is increased to synchronously supervise the training, so as to avoid the geothermal gas noise being misdetected as stone block and stone pile and other medium entities. The annotation true value of the 3D detection algorithm is the foreground target (such as pedestrians, vehicles, stone piers and stone blocks and other obstacles, which can be wrapped by 3D boxes) in the environment. When the multi-task joint training is performed, the 3D detection supervises the learning of the stone piers and stone blocks and other obstacles, and the loss is also used for the gradient back propagation of the 3D segmentation model structure, so as to improve the recognition ability of the stone piers and stone blocks and other medium entities and the recognition ability of the geothermal gas noise (misdetected stone piers and stone blocks and other medium entities).

[0145] Figure 7 is a schematic diagram of a structure during model training according to an embodiment of the application, as Figure 7As shown, the structure can include a point cloud training set 701, a shared backbone network 702, a detection task module 703, and a segmentation task module 704. Among them, the point cloud training set 701 is the basic data source of the entire training process, containing a large amount of labeled lidar scan point cloud data. These data are usually collected by professional data acquisition systems under various conditions, including normal environment and scenes with geothermal gas influence. The label information includes the category of objects in the point cloud (such as ground, retaining wall, noise, vehicle, pedestrian, etc.) and the accurate position of the object (for the detection task, usually represented by a 3D bounding box). The quality and diversity of the training set directly affect the final effect and generalization ability of model training. The shared backbone network 702 plays a core role in the network training structure, extracting high-level feature representations from point cloud data. In the embodiment of the application, Backbone is shared, meaning that it can provide feature input for both the detection task network (detection task module 703) and the segmentation task network (segmentation task module 704). The above design helps the model to learn more general and rich features that can support both detection and segmentation tasks, improving the comprehensive performance and resource utilization of the model.

[0146] Optionally, as Figure 7 shown, the detection task module 703 focuses on identifying and locating objects in the point cloud, based on the features extracted by Backbone, using techniques such as bounding box regression and class prediction to determine the position and class of each object in space. For the mine unmanned driving scene, this includes identifying rock blocks, rock piles, vehicles, and other solid obstacles, as well as potential noise points. By optimizing the loss of 3D detection, the detection task Net can provide more accurate object positioning information, helping to avoid misidentifying geothermal gas noise as an obstacle. The segmentation task module 704, i.e. SegHead, is responsible for 3D semantic segmentation of point cloud data, i.e. determining the class of each voxel. It also uses features extracted by Backbone, but focuses more on refining the classification of each voxel to ensure that it can accurately distinguish between ground, obstacles, and noise points. By using an uneven zoom grid method and a super-resolution division strategy, the segmentation task Net can improve its ability to learn spatial features and reduce the false detection rate of geothermal gas noise.

[0147] Optionally, the multi-task model structure is trained with the detection label and the segmentation label of the complete data set at the same time, and the whole training is completed; then the multi-task model structure is fine-tuned by using the radar training set containing only obstacles such as stone piers and stone blocks and geothermal gas; finally, the segmentation branch task is trained by using the radar training set containing only obstacles such as stone piers and stone blocks and geothermal gas, and the segmentation branch model is fine-tuned. The 3D detection characteristics of obstacles such as stone piers and stone blocks are integrated into the supervision of the model structure during model training. Thus, the ability of the model to distinguish geothermal gas and obstacles such as stone piers and stone blocks is improved. During model inference, the structure does not need to infer the detection task, only the segmentation task, which can save the occupation of computing resources.

[0148] Optionally, in order to solve the false detection of 3D segmentation during inference, the medium entity and noise point verification of specific regions is performed on the region where geothermal gas frequently occurs, such as the loading position and the dump site. The large presence of geothermal gas only exists in regions such as the loading position and the dump site, which involves the change of the temperature of the mining material and thus the surge of geothermal gas. Similarly, the influence on the unmanned vehicle is limited to the processing of the rear of the vehicle during reversing loading or unloading. The performance of the 3D segmentation model cannot distinguish geothermal gas noise, and the false detection is a retaining wall or the ground, and the corresponding confidence is relatively low, and generally has a certain height. The 3D segmentation result corresponds to the positioning (x, y, z) of each point of the laser radar point cloud data, the type, and the confidence (between 0 and 1). In view of the above characteristics, when the vehicle enters the loading position or the dump site region, the unmanned vehicle can obtain the relative business state and the region, at this time, the verification of the 3D segmentation result can be started, and the confidence, height and other attributes are distinguished.

[0149] Optionally, the point cloud data of the ground, the retaining wall, and the obstacle entity medium are clustered, and the corresponding entity obstacle result (related to the shape and height of the obstacle) can be obtained; the confidence of the point cloud higher than the ground by a certain height, such as 0.3 m or more, is counted, such as the mean and mode of the confidence of 0.3 m or more are less than a certain threshold, such as 0.5. And the internal entity obstacle result contains a certain number of noise points (such as the total number of point cloud points M, and the number of noise points reaches M / 5), the obstacle is recalled, and the above obstacle is judged as a false detection, and is not output to the downstream.

[0150] Figure 8 is a structural schematic diagram of a non-interference entity detection device according to an embodiment of the present application, as shown in Figure 8 The device can include an acquisition unit 802, a division unit 804, and an identification unit 806.

[0151] The acquisition unit 802 is configured to acquire a detection result of the unmanned vehicle. The division unit 804 is configured to perform spatial division on the detection result according to a target grid model to obtain spatial division feature data. The identification unit 806 is configured to input the spatial division feature data into a non-interference entity identification model for identification to obtain an identification result.

[0152] Figure 9 FIG. 8 is a structural schematic diagram of another non-interference entity detection device according to an embodiment of the present application, as shown in the figure, the device can include a detection unit 902 and a control unit 904. Figure 9

[0153] The detection unit 902 is configured to detect a service state and / or a region where the unmanned vehicle is located. The control unit 904 is configured to control the unmanned vehicle to start a target perception mode in a case where the service state and / or the region where the unmanned vehicle is located meets a target condition. The device can further include a processor and a memory, and the above units are stored in the memory as program units, and the corresponding functions are implemented by the processor executing the program units stored in the memory. The processor includes a kernel, and the kernel calls the corresponding program units from the memory. The kernel can be one or more, and the same device type of the to-be-stopped device is stopped by adjusting the kernel parameters. The memory can include a non-persistent memory in a computer readable medium, a random access memory (RAM) and / or a non-volatile memory such as a read-only memory (ROM) or a flash memory, and the memory includes at least one memory chip. The processor includes a kernel, and the kernel calls the corresponding program units from the memory. The kernel can be one or more, and the work efficiency of the transaction personnel is improved by adjusting the kernel parameters. The memory can include a non-persistent memory in a computer readable medium, a RAM and / or a non-volatile memory such as a ROM or a flash memory, and the memory includes at least one memory chip.

[0154] According to an embodiment of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a program. The program is executed by a processor to implement the above method. According to an embodiment of the present application, a processor is provided, and the processor is configured to run a program. The program is executed to implement the above method. Figure 10 FIG. 9 is a schematic diagram of an electronic device for a non-interference entity detection method according to an embodiment of the present application, as shown in the figure, the device can include a detection unit 902 and a control unit 904. Figure 10 ​As shown, the embodiments of the present application further provide an electronic device 1000, which comprises a processor 1001, a memory 1002, and a program stored in the memory and capable of running on the processor, and the processor implements the method in any of the above embodiments when executing the program, which will not be repeated here. The device herein can be a server, a personal computer (PC), a tablet computer (PAD), a mobile phone, etc. The present application further provides a computer program product which, when executed on a data processing device, is adapted to execute a program initialized with any of the above method steps. According to another aspect of the embodiments of the present application, a vehicle is further provided, which comprises a memory and a processor, the memory stores a computer program, and the processor is configured to run the computer program to execute any of the method steps.

[0155] The above merely illustrates the embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of the claims of the present application.

Claims

1. A method for detecting a non-interfering entity, characterized in that: include: Obtaining a detection result of the unmanned vehicle, wherein the detection result is obtained by a sensing device of the unmanned vehicle detecting a target environment; Performing spatial division on the detection results according to a target grid model to obtain spatial division feature data, wherein the target grid model has grids that are non-uniformly distributed in a height direction and grids that are uniformly distributed in a horizontal direction; The spatial division feature data is input into a non-interference entity recognition model for recognition to obtain a recognition result, wherein the recognition result is used to indicate whether a non-interference entity exists in the target environment.

2. The method according to claim 1, characterized in that The method further comprises: Determining non-interference entity distribution attribute information corresponding to the target environment, wherein the non-interference entity distribution attribute information is used to represent non-interference entity distribution characteristics when non-interference entities exist in the target environment; Based at least on the non-interference entity distribution attribute information, a grid resolution of the target grid model in the height direction is determined.

3. The method according to claim 1 or 2, characterized in that The target grid model has grids that are non-uniformly distributed in the height direction that satisfy the following conditions: a grid resolution of the grids in the first grid area in the height direction is higher than a grid resolution of the grids in the second grid area in the height direction, and the first grid area is located closer to the ground area than the second grid area.

4. The method according to claim 2, characterized in that The determining of the non-interference entity distribution attribute information corresponding to the target environment includes: Obtaining business scenarios or regional information corresponding to the target environment; According to the service scenario or the area information, corresponding designated non-interference entity distribution attribute information is determined as the non-interference entity distribution attribute information.

5. The method according to claim 2, characterized in that The determining, based at least on the non-interference entity distribution attribute information, a grid resolution of the target grid model in the height direction, includes: Based on the non-interference entity distribution attributes and the computing resources of the unmanned vehicle, a grid resolution of the target grid model in the height direction is determined, wherein an average value of the grid resolution in the height direction is positively correlated with the computing resources.

6. The method according to claim 1, characterized in that The non-interference entity recognition model includes a segmentation model, the segmentation model is obtained by training an initial segmentation model using spatial division result samples of the unmanned vehicle in the target environment, the spatial division result samples are obtained by spatially dividing the detection result samples in the target environment using multiple grid models, and different grid models have different grid resolutions in the height direction. The spatial division feature data is input into the non-interference entity recognition model for recognition, and a recognition result is obtained, including: The spatial division feature data is input into the segmentation model for recognition to obtain the recognition result.

7. The method according to claim 1 or 6, characterized in that The method further comprises: Determining a detection label and a segmentation label in the detection result sample under the target environment, wherein the detection label is used to represent a detection box marked for the object sample under the target environment, and the segmentation label is used to represent a classification result marked for the object sample; Based on the detection label, the segmentation label and multiple grid models, an initial detection model and an initial segmentation model are trained to obtain the non-interference entity recognition model, wherein different grid models have different grid resolutions in the height direction.

8. The method according to claim 7, characterized in that The object samples include obstacle samples and non-interference entity samples, and the training of the initial detection model and the initial segmentation model based on the detection label, the segmentation label and the multiple grid models to obtain the non-interference entity recognition model includes: Training the initial detection model and the initial segmentation model using the detection labels, the segmentation labels, and the multiple grid models; The trained initial detection model and the trained initial segmentation model are adjusted by using the attribute information of the obstacle samples and the noise information of the non-interference entity samples to obtain the non-interference entity recognition model.

9. The method according to claim 8, characterized in that Utilizing the attribute information of the obstacle sample and the noise information of the non-interference entity sample, the adjusted trained initial detection model and the trained initial segmentation model to obtain the non-interference entity recognition model, including: Using the attribute information of the obstacle samples and the noise information of the non-interference entity samples, adjusting the trained initial detection model and the trained initial segmentation model to obtain a target detection model and a target segmentation model; Using the attribute information of the obstacle sample and the noise information of the non-interference entity sample, the target segmentation model is adjusted to obtain the segmentation model; The segmentation model and the target detection model are constructed into the non-interference entity recognition model.

10. The method according to claim 8, characterized in that The training of the initial detection model and the initial segmentation model using the detection label, the segmentation label, and the multiple grid models includes: Determining a first loss between an output result of the initial detection model and a detection label in the detection result sample, and a second loss between an output result of the initial segmentation model and the segmentation label in the detection result sample; Using the multiple grid models, spatially dividing the detection result samples to obtain spatially divided result samples; The initial detection model is trained using the first loss and the spatial division result samples, and the gradient corresponding to the second loss is adjusted using the first loss, and the initial segmentation model is trained using the adjusted gradient and the spatial division result samples.

11. A method for detecting a non-interfering entity, characterized in that: include: Detecting the business status and / or area of ​​the unmanned vehicle; When the business status and / or the area in which the vehicle is located meets the target conditions, the unmanned vehicle is controlled to start a target perception mode, wherein the target perception mode is used to detect whether there is a non-interference entity in the current environment.

12. The method according to claim 11, characterized in that The target conditions include at least one of the following: In the case where the business state of the unmanned vehicle is detected, the target condition is that the business state is a loading business state, a soil dumping business state, or a road transport state; When the area where the unmanned vehicle is located is detected, the target condition is that the area is a loading area, a dumping area or a transportation road area.

13. The method according to claim 11, characterized in that The target perception mode is used to use a non-interference entity recognition model to identify the spatial division feature data of the unmanned vehicle to obtain a recognition result. The recognition result is used to indicate whether there is a non-interference entity in the current environment. The spatial division feature data is obtained by spatially dividing the detection results of the unmanned vehicle according to a target grid model. The target grid model has grids that are non-uniformly distributed in the height direction and grids that are evenly distributed in the horizontal direction.

14. The method according to claim 11 or 13, characterized in that Also includes: A perception data set obtained by detecting the current environment by a perception device of the unmanned vehicle; Using a non-interference entity recognition model to identify the spatial division feature data of the unmanned vehicle to obtain a recognition result; Obtaining a classification result corresponding to the perception data set and a confidence level corresponding to the classification result according to the recognition result; Output a target verification result corresponding to the classification result and the confidence level, wherein the target verification result includes a first verification result indicating that the classification result is a false positive.

15. An unmanned vehicle, characterized in that: include: a memory storing an executable program; A processor, configured to run the program, wherein the program executes the method according to any one of claims 1 to 14 when running.

Images