A patrol vehicle positioning system based on NRTK and visual information

Abstract

This invention discloses a patrol vehicle positioning system based on NRTK and visual information, including a priori mapping subsystem and a visual SLAM positioning subsystem. The priori mapping subsystem runs first, and after obtaining the point cloud map, the visual SLAM positioning subsystem runs. Before running, the visual SLAM positioning subsystem loads all the point cloud map information at once. This invention provides a patrol vehicle positioning system based on NRTK and visual information, which can ensure high-precision positioning of the patrol vehicle while reducing its operating costs and increasing economic benefits.

Description

Technical Field

[0001] This invention relates to the field of positioning system technology, and in particular to a patrol vehicle positioning system based on NRTK and visual information. Background Technology

[0002] Patrol vehicles in on-street parking lots need to patrol the parking area for extended periods, constantly photographing the parking situation to generate billing information. Since these patrol vehicles need to record parking locations and durations, their own location tracking capabilities are crucial.

[0003] Current high-precision positioning methods for patrol vehicles either rely on specialized equipment or multiple sensors working together. Regarding specialized equipment, methods currently include Bluetooth beacons, embedded metal parking space tags, RFID electronic tags, and parking space QR codes. These methods pre-bind specialized equipment with signal transmission or unique identification characteristics to parking spaces in on-street parking lots. When the patrol vehicle passes a parking space, the equipment emits a signal, or the patrol vehicle detects the signal or characteristics of the equipment to identify the parking space and estimate its own position. Positioning methods relying on multiple sensors simultaneously process information collected by various high-precision sensors, such as inertial measurement units, lidar, or high-precision GPS signal receivers, to jointly estimate the patrol vehicle's position.

[0004] The positioning method for patrol vehicles that relies on specialized equipment requires the production and installation of a large number of Bluetooth beacons, embedded metal parking space signs, RFID electronic tags, etc., and may even require some modification to the existing road surface. This results in high operating costs. However, if this positioning method is to improve accuracy, it needs to rely on high-precision sensors, such as precise inertial measurement units, high-precision lidar, or high-precision GPS signal receivers. These high-precision devices are often expensive to manufacture or require continuous leasing services (such as high-precision GPS positioning services), which greatly increases the manufacturing cost of patrol vehicles and thus reduces their economic benefits. Summary of the Invention

[0005] The purpose of this invention is to provide a patrol vehicle positioning system based on NRTK and visual information, which can ensure high-precision positioning of the patrol vehicle while reducing the operating cost of the patrol vehicle and increasing economic benefits.

[0006] The technical solution adopted by the inspection vehicle positioning system based on NRTK and visual information disclosed in this invention is as follows:

[0007] A patrol vehicle positioning system based on NRTK and visual information includes a prior mapping subsystem and a visual SLAM positioning subsystem. The prior mapping subsystem runs first, and after obtaining the point cloud map, the visual SLAM positioning subsystem runs. Before running, the visual SLAM positioning subsystem loads all the point cloud map information at once.

[0008] The input data of the prior mapping subsystem is the image and the corresponding NRTK location information. NRTK provides accurate global latitude and longitude coordinates, which can be converted into spatial three-dimensional coordinates through the geocentric coordinate system. The image and its corresponding NRTK location information can be used together to perform constrained motion structure recovery, thereby obtaining the position, attitude and spatial point cloud map of the image. At this point, the prior mapping subsystem has completed its work.

[0009] The visual SLAM localization subsystem constructs its own continuous position and attitude estimation based on real-time acquired image information and visual SLAM technology. At the same time, it determines the current location on the map based on the point cloud map and provides the visual SLAM with common constraints on position and attitude estimation. During the operation of the visual SLAM subsystem, the images are also transmitted to the cloud server intermittently.

[0010] The CLIP model in the cloud server calculates the CLIP spatial representation vector of the image and compares it with the text description of key locations along the path to provide a rough observation of the geographic location.

[0011] As a preferred approach, the construction of the point cloud map requires the use of motion structure recovery technology, which can be considered as an optimization task aimed at minimizing the reprojection errors of all images. The optimization objective is to obtain N spatial points, assuming they are... in Let a spatial point p i Two-dimensional image points mapped to the camera at the initial pose The mapping relationship is as follows:

[0012]

[0013] in It is a matrix determined by the physical properties of the camera, when the spatial point p i Mapped to a pose Location is Point of view from the camera Therefore, the mapping relationship is:

[0014]

[0015] in It is p iThe reciprocal of the z-axis coordinate after spatial transformation, when a spatial point is observed by M cameras, is the pose of the M cameras. Location The parameters will form mutual constraints, and finally only one optimization objective needs to be constructed to complete the nonlinear optimization:

[0016]

[0017] in It is p i The actual coordinates mapped to the j-th camera. Furthermore, since monocular vision cannot recover depth, it is generally not optimized. It is not an exact value, but rather a ratio: the pivot frame's... Consider it as 1, and the depth of other frames is a multiple of it. Since we need to recover a map containing N map points, the final optimization objective is:

[0018]

[0019] Instead of starting optimization with random parameters, we first use techniques such as feature point tracking, epipolar geometry, and triangulation to obtain rough initial values, and then start optimization from these initial values.

[0020] Additional absolute position information is needed. Assume that the precise position information of L cameras has already been obtained through NRTK signals or actual measurements. At this point, the optimization objective can be modified to:

[0021]

[0022] In the final optimization results To form a sparse point cloud map, It can also provide location reference for certain situations when used in subsequent map applications.

[0023] As a preferred option, the location information obtained by the NRTK is represented using longitude, latitude, and altitude. The mapping method is performed in the Cartesian coordinate system. Therefore, it is necessary to convert the Earth's longitude, latitude, and altitude coordinates into Cartesian coordinates. The selected Cartesian coordinate system is the geocentric-geocentric coordinate system. In the geocentric-geocentric coordinate system, the Earth's center of mass is located at the origin, the Z-axis points towards the North Pole of the Earth's rotation axis, the X-axis is located in the equatorial plane and points towards the Prime Meridian, while the Y-axis is perpendicular to the XZ plane and points eastward, forming a right-handed coordinate system. In this coordinate system, the Earth's rotation causes the position of all ground points to be fixed in three-dimensional space, which is called geocentric. That is, the coordinates of any fixed point on Earth in this coordinate system will not change with time.

[0024] In the Earth-centered Earth-fixed coordinate system, the location of any point can be represented by three coordinate values: X, Y, and Z. Specifically, the latitude φ, longitude λ, and altitude h of a point on or near the Earth's surface can be converted to Earth-centered Earth-fixed coordinates using a specific transformation formula. The radius N of the circumpolar orbit at that point can then be calculated using the following formula:

[0025]

[0026] Where a is the Earth's semi-major axis and e is the Earth's first eccentricity, calculated using the following formula:

[0027]

[0028] b is the Earth's minor semi-axis. The following formula converts geographic coordinates to geocentric coordinates:

[0029]

[0030] Each t j All of these can be represented as [X, Y, Z]. T .

[0031] As a preferred embodiment, the visual SLAM localization subsystem incorporates features that can describe the global state of the image. These global features can be obtained by aggregating local features, resulting in N D-dimensional descriptors for each image. We can first cluster these N descriptors to obtain K cluster centers. The overall features of the image are Therefore, V is calculated as follows:

[0032] Where V(k) represents the k-th vector in the overall features, α k (x i ) is an indicator function, if x i The cluster center is c k , then α k (x i ) = 1, otherwise α k (x i ) = 0. V is an aggregated local descriptor vector that uses vectors learned from a neural network to replace the cluster centers obtained from clustering. Furthermore, due to α k (x i Since is a non-differentiable function, parameter learning is not possible. Therefore, it needs to be modified as follows:

[0033]

[0034] Where w k b k To learn the parameters, this modification actually replaces the original indicator function with a continuous probability function; therefore, the new VLAD calculation method is as follows:

[0035]

[0036] With both global and local features available, the system first retrieves the closest image from the map based on the global features, and then performs precise localization based on the local features.

[0037] As a preferred option, the CLIP model is mainly used for language enhancement functions in map positioning: images are collected during the patrol vehicle's journey and transmitted to a cloud server to calculate their feature vectors using the CLIP model; simultaneously, the CLIP model is used to calculate the feature vectors of each text in the path text database, and correlation calculation is performed. Based on the correlation score, it is determined whether the patrol vehicle has reached a certain landmark location; if so, coordinate information of the vicinity of the location is provided to assist in positioning; otherwise, pure visual positioning is performed normally.

[0038] As a preferred embodiment, the path text database is a text description database of the inspection route of an inspection vehicle. It contains several human-specified natural language entries to represent key locations on the route, collectively referred to as path text, and each path text entry is assigned a geographic coordinate.

[0039] The beneficial effects of the NRTK and visual information-based patrol vehicle positioning system disclosed in this invention are as follows: This system mainly includes two subsystems: a priori mapping subsystem and a visual SLAM positioning subsystem. The priori mapping subsystem generates a high-precision on-street parking map by collecting image information and high-precision NRTK positioning information and using motion structure recovery technology. The visual SLAM positioning subsystem determines its own position in real time based on images and maps when the patrol vehicle is patrolling. The priori mapping subsystem only needs to be run once, and its results can be used by the visual SLAM positioning subsystem for a long time. Therefore, NRTK related services do not need to be continuously leased, and the multiple sensors involved in this system do not need to be precisely calibrated. It has the characteristics of low cost, simple implementation, and high accuracy. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the structure of an inspection vehicle positioning system based on NRTK and visual information according to the present invention.

[0041] Figure 2 This is a schematic diagram of the sliding window and keyframes in a visual SLAM system for a patrol vehicle positioning system based on NRTK and visual information according to the present invention. Detailed Implementation

[0042] The present invention will be further described and illustrated below with reference to specific embodiments and the accompanying drawings:

[0043] Please refer to Figure 1 A patrol vehicle positioning system based on NRTK and visual information includes a priori mapping subsystem and a visual SLAM positioning subsystem. The priori mapping subsystem runs first, and after obtaining the point cloud map, the visual SLAM positioning subsystem runs. Before running the visual SLAM positioning subsystem, all point cloud map information is loaded at once.

[0044] The input data for the prior mapping subsystem are images and their corresponding NRTK positioning information. NRTK provides precise global latitude and longitude coordinates (with errors at the centimeter level), which can be converted into three-dimensional spatial coordinates using the Earth-Centered, Earth-Fixed (ECEF) coordinate system. The images and their corresponding NRTK position information can be used together for constrained structure from motion (SFM) reconstruction, thereby obtaining the image's position, orientation, and spatial point cloud map. At this point, the prior mapping subsystem has completed its work.

[0045] The construction of point cloud maps requires the use of motion structure recovery techniques, which can be viewed as an optimization task aimed at minimizing the reprojection errors of all images. The optimization objective is to obtain N spatial points, assuming they are... in Let a spatial point p i Two-dimensional image points mapped to the camera at the initial pose The mapping relationship is as follows:

[0046]

[0047] in It is a matrix determined by the physical properties of the camera, when the spatial point p i Mapped to a pose Location is Point of view from the camera Therefore, the mapping relationship is:

[0048]

[0049] in It is p i The reciprocal of the z-axis coordinate after spatial transformation, when a spatial point is observed by M cameras, is the pose of the M cameras. Location The parameters will form mutual constraints, and finally only one optimization objective needs to be constructed to complete the nonlinear optimization:

[0050]

[0051] Recovery depth is generally not optimized. It is not an exact value, but rather a ratio: the pivot frame's... Consider it as 1, and the depth of other frames is a multiple of it. Since we need to reconstruct a map containing N map points, the final optimization objective is...

[0052]

[0053] Instead of starting optimization with random parameters, we first use techniques such as feature point tracking, epipolar geometry, and triangulation to obtain rough initial values, and then start optimization from these initial values.

[0054] So far, the recovered maps lack a real-world physical scale and cannot be used directly. Additional absolute position information must be added, such as the t-axis of some cameras. j Assume that the precise location information of L cameras has already been obtained through NRTK signals or actual measurements. At this point, the optimization objective can be modified to...

[0055]

[0056] In the final optimization results To form a sparse point cloud map, It can also provide location reference for certain situations when used in subsequent map applications.

[0057] NRTK acquires location information using longitude, latitude, and altitude, and the mapping method is performed in a Cartesian coordinate system. Therefore, it is necessary to convert the Earth's latitude, longitude, and altitude coordinates to Cartesian coordinates. The chosen Cartesian coordinate system is the geocentric-fixed coordinate system. In the geocentric-fixed coordinate system, the Earth's center of mass is located at the origin, the Z-axis points towards the North Pole of the Earth's rotation axis, the X-axis lies in the equatorial plane and points towards the Prime Meridian, and the Y-axis is perpendicular to the XZ plane, pointing eastward, forming a right-handed coordinate system. In this coordinate system, the Earth's rotation causes the position of all points on the ground to be fixed in three-dimensional space; therefore, it is called geocentric-fixed, meaning that the coordinates of any fixed point on Earth in this coordinate system do not change over time.

[0058] In the Earth-centered Earth-fixed coordinate system, the location of any point can be represented by three coordinate values: X, Y, and Z. Specifically, the latitude φ, longitude λ, and altitude h of a point on or near the Earth's surface can be converted to Earth-centered Earth-fixed coordinates using a specific transformation formula. The radius N of the circumpolar orbit at that point can then be calculated using the following formula:

[0059]

[0060] Where 'a' is the Earth's semi-major axis, approximately 6,378,137.0 meters, and 'e' is the Earth's first eccentricity, calculated using the following formula:

[0061]

[0062] b is the Earth's minor semi-axis, approximately 6,356,752.3 meters. The following formula converts geographic coordinates to geocentric coordinates:

[0063]

[0064] Each t j All of these can be represented as [X, Y, Z]. T .

[0065] The visual SLAM localization subsystem first constructs its own continuous position and pose estimation based on real-time acquired image information using visual SLAM technology. At the same time, it determines the current location on the map based on the point cloud map, providing visual SLAM with common constraints for position and pose estimation.

[0066] During the visual SLAM subsystem operation, images are also periodically transmitted to a cloud server. The CLIP model on the cloud server calculates the CLIP spatial representation vector of the image and compares it with the text description of key locations along the path to provide a rough observation of the geographic location.

[0067] Visual SLAM in the visual localization subsystem is an optimization task with Equation (2) as the objective. All observation points are point cloud map points generated during the prior map creation stage. However, for efficiency reasons, keyframes and sliding windows are often used, such as... Figure 2As shown. The concept of keyframes mainly originates from the fact that during camera movement, some image frames exhibit significant motion, while others show no obvious difference (e.g., when the camera is temporarily stationary). To reduce the number of optimization images in SLAM, images with significant inter-frame motion are selected as keyframes. This avoids disrupting image continuity and conserves computational resources for redundant images. However, the keyframe method still maintains a large optimization image library in long-distance SLAM, eventually causing the system to stop due to resource exhaustion. Therefore, a sliding window method is needed. This method maintains only a fixed number of keyframes for optimization. Newly arrived keyframes are added to the back of the window, while the first keyframe is discarded, ensuring that the current window always maintains the most up-to-date set of keyframes.

[0068] Map-based localization utilizes deep learning-based image retrieval and feature point matching techniques. Traditional localization methods treat the map as a codebook containing numerous map points. When the map is needed, the current image is matched against the codebook content to find the closest match, and then this information is used to solve for the current localization. The basic elements of the codebook are typically map point descriptors, such as ORB and SIFT. These descriptors are generally calculated based on local image information, such as brightness gradients and texture features. This local information allows for fast and accurate matching when the scene is consistent. However, when encountering scenes with changing lighting or surface variations, the accuracy of this local information decreases significantly. Therefore, using only local image information is insufficient; features that describe the entire image must be incorporated.

[0069] Global features of an image can be obtained by aggregating local features, thus generating N descriptors of dimension D for an image. We can first cluster these N descriptors to obtain K cluster centers. The overall features of the image are Therefore, V is calculated as follows:

[0070] Where V(k) represents the k-th vector in the overall features, α k (x i ) is an indicator function, if x i The cluster center is c k , then α k (x i ) = 1, otherwise α k (x i ) = 0. V is an aggregated local descriptor vector that uses vectors learned from a neural network to replace the cluster centers obtained from clustering. Furthermore, due to α k(x i Since is a non-differentiable function, parameter learning is not possible. Therefore, it needs to be modified as follows:

[0071]

[0072] Where w k b k To learn the parameters, this modification actually replaces the original indicator function with a continuous probability function; therefore, the new VLAD calculation method is as follows:

[0073]

[0074] With both global and local features available, the system first retrieves the closest image from the map based on the global features, and then performs precise localization based on the local features.

[0075] The CLIP model is pre-tuned based on the inspection vehicle's route. The fine-tuning method involves collecting images of key locations along the vehicle's fixed route, such as intersections of two main roads or near landmark buildings. Each image is manually labeled with a natural language text, such as "intersection of Road A and Road B" or "plaza directly in front of Building C." The CLIP model is then fine-tuned using supervised training. Furthermore, a database is built to record the geographic coordinates corresponding to each natural language text. During actual operation, if a captured image matches a natural language text, the corresponding geographic coordinates are provided as an observation for the visual SLAM localization subsystem.

[0076] The CLIP model is mainly used for language enhancement functions in map positioning: it collects images while the patrol vehicle is driving, and transmits the images to the cloud server to calculate their feature vectors using the CLIP model; at the same time, it uses the CLIP model to calculate the feature vector of each text in the path text database, performs correlation calculation, and determines whether the patrol vehicle has driven to a certain landmark location based on the correlation score; if so, it provides the coordinate information of the vicinity of the location to assist in positioning, otherwise it performs pure visual positioning normally.

[0077] The route text database is a text description database of the patrol vehicle's inspection routes. It contains several human-defined natural language entries representing key locations along the routes, collectively referred to as route text, and each route text is assigned a geographic coordinate. Examples include "intersection of road A and road B" and "bus stop D for bus route C." For instance, based on the prior map construction process, if an image of the "intersection of road A and road B" is located near the (X1, Y1) position, then the route text is assigned the geographic coordinates (X1, Y1). As described above, all these steps are completed during the prior map creation phase.

[0078] The advantage of using the CLIP model for text augmentation in localization is that pure visual camera localization is susceptible to fluctuations due to dynamic changes in the environment. For example, the visual description of an intersection may change drastically when it is full of parked cars and when it is empty. However, with the addition of text description, the localization is anchored to certain static objects. For example, the text "intersection of road A and road B" focuses on the name and topology of the roads, which makes the camera localization more robust.

[0079] This invention proposes a method for locating patrol vehicles that does not require altering road surfaces or installing specialized equipment, thereby eliminating the need to manufacture and install large quantities of Bluetooth beacons, embedded metal parking space signs, RFID electronic tags, and other similar equipment. This not only significantly reduces initial investment costs but also avoids the high expenses associated with modifying existing road surfaces. Compared to multi-sensor-based positioning methods, this method eliminates the need for high-precision inertial measurement units, high-precision lidar, or continuous leasing of high-precision GPS positioning services, thus substantially reducing equipment costs. By reducing the need for fixed equipment, patrol service providers can deploy and manage patrol vehicles more flexibly, further saving on maintenance and operating costs.

[0080] The vehicle positioning method of this invention simplifies sensor configuration requirements significantly by requiring only a camera and short-term high-precision GPS signal service, while long-term positioning can be achieved solely with the camera. This method not only reduces initial installation complexity but also provides greater flexibility, facilitating future technology upgrades and system expansion. Inspection service providers can quickly adjust equipment configurations as needed, improving operational flexibility and adaptability.

[0081] This invention simplifies operation and reduces the requirements for sensor extrinsic parameters and time alignment, thereby lowering the technical barriers. This allows even operators without specialized technical backgrounds to easily use and maintain the inspection vehicle, further reducing labor and training costs. This simplified design enables the inspection vehicle to be deployed more quickly, improving overall efficiency and economic benefits.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A patrol vehicle positioning system based on NRTK and visual information, characterized in that, It includes a prior mapping subsystem and a visual SLAM localization subsystem. The prior mapping subsystem runs first, and after obtaining the point cloud map, the visual SLAM localization subsystem runs. Before running the visual SLAM localization subsystem, all point cloud map information is loaded at once. The input data of the prior mapping subsystem is the image and the corresponding NRTK location information. NRTK provides accurate global latitude and longitude coordinates, which can be converted into spatial three-dimensional coordinates through the geocentric coordinate system. The image and its corresponding NRTK location information can be used together to perform constrained motion structure recovery, thereby obtaining the position, attitude and spatial point cloud map of the image. At this point, the prior mapping subsystem has completed its work. The visual SLAM localization subsystem constructs its own continuous position and pose estimation based on real-time acquired image information using visual SLAM technology. At the same time, it determines the current location on the map based on the point cloud map and provides the visual SLAM with common constraints for position and pose estimation. During the operation of the visual SLAM subsystem, images are also transmitted to the cloud server intermittently. The CLIP model in the cloud server calculates the CLIP spatial representation vector of the image and compares it with the text description of key locations along the path to provide a rough observation of the geographic location. The construction of the point cloud map requires the use of motion structure recovery technology, which can be considered as an optimization task to minimize the reprojection error of all images. The optimization objective is to obtain N spatial points, assuming they are... in Let a spatial point p i Mapping a 2D image point u to the camera at the initial pose i = The mapping relationship is as follows: in It is a matrix determined by the physical properties of the camera, when the spatial point p i Mapped to a pose Location is From the camera's perspective Therefore, the mapping relationship is: in It is p i The reciprocal of the z-axis coordinate after spatial transformation, when a spatial point is observed by M cameras, is the pose of the M cameras. Location The parameters will form mutual constraints, and finally only one optimization objective needs to be constructed to complete the nonlinear optimization: in It is p i The actual coordinates mapped to the j-th camera. Furthermore, since monocular vision cannot recover depth, it is generally not optimized. It is not an exact value, but rather a ratio: the pivot frame's... Consider it as 1, and the depth of other frames is a multiple of it. Since we need to recover a map containing N map points, the final optimization objective is: Instead of starting optimization with random parameters, we first use feature point tracking, epipolar geometry, and triangulation to obtain rough initial values, and then start optimization from these initial values. Additional absolute position information is needed. Assume that the precise position information of L cameras has already been obtained through NRTK signals or actual measurements. At this point, the optimization objective can be modified to: In the final optimization results To form a sparse point cloud map, It can also provide location reference for certain situations when used in subsequent map applications.

2. The inspection vehicle positioning system based on NRTK and visual information as described in claim 1, characterized in that, The location information obtained by NRTK is represented using longitude, latitude, and altitude. The mapping method is performed in the Cartesian coordinate system. Therefore, it is necessary to convert the Earth's longitude, latitude, and altitude coordinates into Cartesian coordinates. The selected Cartesian coordinate system is the geocentric-geocentric coordinate system. In the geocentric-geocentric coordinate system, the Earth's center of mass is located at the origin, the Z-axis points towards the North Pole of the Earth's rotation axis, the X-axis is located in the equatorial plane and points towards the Prime Meridian, and the Y-axis is perpendicular to the XZ plane, forming a right-handed coordinate system. In this coordinate system, the Earth's rotation causes the position of all ground points to be fixed in three-dimensional space, which is called geocentric. That is, the coordinates of any fixed point on Earth in this coordinate system will not change with time. In the Earth-centered Earth-fixed coordinate system, the location of any point can be represented by three coordinate values: X, Y, and Z. The latitude φ, longitude λ, and altitude h of a point on or near the Earth's surface can be converted to Earth-centered Earth-fixed coordinates using a specific transformation formula. The radius N of the circumpolar orbit at that point can then be calculated using the following formula: Where 'a' is the Earth's semi-major axis and 'e' is the Earth's first eccentricity, calculated using the following formula: b is the Earth's minor semi-axis. The following formula converts geographic coordinates to geocentric coordinates: Each t j All of these can be represented as [X,Y,Z] T .

3. The inspection vehicle positioning system based on NRTK and visual information as described in claim 1, characterized in that, In the visual SLAM localization subsystem, features that can describe the global state of an image are incorporated. These global features can be obtained by aggregating local features, resulting in N descriptors of dimension D for each image. We can first cluster these N descriptors to obtain K cluster centers. The overall features of the image are The calculation method for V is as follows: Where V(k) represents the k-th vector in the overall features, α k (x i ) is an indicator function, if x i The cluster center is c k Then α k (x i ) = 1, otherwise α k (x i ) = 0, V is an aggregated local descriptor vector that uses vectors learned from neural networks to replace the cluster centers obtained from clustering. Furthermore, due to α k (x i Since is a non-differentiable function, parameter learning is not possible. Therefore, it needs to be modified as follows: Where w k ,b k To learn the parameters, this modification actually replaces the original indicator function with a continuous probability function; therefore, the new VLAD calculation method is as follows: With both global and local features available, the system first retrieves the closest image from the map based on the global features, and then performs precise localization based on the local features.

4. The inspection vehicle positioning system based on NRTK and visual information as described in claim 1, characterized in that, The CLIP model is used for language enhancement functions in map positioning: images are collected during the patrol vehicle's operation and transmitted to a cloud server to calculate their feature vectors using the CLIP model; Simultaneously, the CLIP model is used to calculate the feature vector of each text in the path text database, and correlation calculation is performed. Based on the correlation score, it is determined whether the current inspection vehicle has driven to a certain landmark location. If so, the coordinate information of the vicinity of the location is provided to assist in positioning; otherwise, pure visual positioning is performed normally.

5. The inspection vehicle positioning system based on NRTK and visual information as described in claim 4, characterized in that, The route text database is a text description database of the patrol vehicle's patrol routes. It contains several human-specified natural language entries to represent key locations on the route, collectively referred to as route text, and each route text entry is assigned a geographic coordinate.

Images