TWI835320B - UAV visual navigation system and method for ground feature image positioning and correction - Google Patents

Abstract

A UAV visual navigation method for locating and correcting ground feature images is implemented on a UAV. The landmark image and landmark coordinates of the landmark, and the surrounding coordinates of the surrounding images around the landmark are stored in advance; during flight, based on the previous image of the previous frame Read the correction point image relative to the target landmark coordinates, and determine whether the current image matches any correction point image; if so, compare the correction point image with known coordinates with the current image to determine the current coordinates; if not , then the current coordinates are judged based on the previous image and the comparison between the previous coordinates and the current image; the present invention combines the previous image comparison to quickly generate the current coordinates and the landmark image matching to accurately correct the coordinates, so that the UAV can effectively use computing resources and reduce errors perform flight missions.

Description

UAV visual navigation system and method for ground feature image positioning and correction

A UAV visual navigation system and method, in particular, a UAV visual navigation system and method for ground feature image positioning and correction.

When a drone performs an automatic flight mission, it needs to use a specific automatic navigation method to determine the next flight direction. Among them, visual navigation is a low-cost navigation method for drones. It only needs to be equipped with a camera and an on-board computer that can calculate it.

In a drone visual navigation technology, the features of the continuous images captured by the drone during flight are sequentially matched to determine the movement position of the drone relative to the previous image, and then determine the flight direction and location of the drone. Location. However, the feature extraction and matching of continuous images will cause matching errors or matching of similar objects. The matching of continuous images will also have errors. If not corrected, the errors will become larger and larger, affecting the flight mission of the UAV. result.

In another drone visual navigation technology, the location of the drone is determined by matching known landmark images and their coordinates pre-stored in the database with images captured by the drone. The landmark image system mentioned here is Refers to a partial image of a building, object, regional environment, etc. with a unique appearance viewed from a high altitude. However, if the drone performs multiple landmark image matching based on each captured image, it will consume a lot of computing resources. If the landmark image matching is performed based on a specific period, there is a possibility of missing landmarks; when landmark images and their related information When the number is huge, matching them all will also increase the consumption of computing resources; in addition, how to judge the correctness of the landmark image matching results, or how to select the matching results when matching multiple landmark images in the database, is also a difficult problem.

In summary, existing UAV automatic navigation methods must be further improved.

In view of the fact that existing UAV automatic navigation methods still have problems such as continuous image matching errors, matching errors, computing resource consumption or landmark matching result judgment, the present invention provides a UAV visual navigation method for ground feature image positioning and correction, which is processed by a The device is executed, including the following steps: capture a current image below the drone, and read a previous image and a previous coordinate corresponding to the previous image; based on the previous coordinates and a landmark coordinate, read the coordinates corresponding to the landmark Multiple correction point images; perform image matching on the current image and the correction point images respectively to determine whether the current image matches one of the correction point images; if the current image matches one of the correction point images, according to the current image, match Determine a current coordinate based on the correction point image and its corresponding correction point coordinate; if not, determine the current coordinate based on the current image, the previous image and its corresponding previous coordinate; and determine the current coordinate based on the current coordinate and the landmark coordinate Generate a navigation flight command to make the UAV fly to the landmark coordinates.

In addition, the present invention also provides a UAV visual navigation system for positioning and correcting ground feature images, including: a camera that captures a current image; a memory that is electrically connected to the camera and stores a plurality of previous images and corresponding previous coordinates. , a navigation map, a flight mission information, a plurality of landmark images and a plurality of corresponding landmark coordinates, a plurality of surrounding images corresponding to each landmark image, and a plurality of surrounding coordinates corresponding to the surrounding images; A processor, electrically connected to the memory, executes the aforementioned UAV visual navigation method of ground feature image positioning and correction.

When the UAV visual navigation method for locating and correcting ground feature images of the present invention determines the current coordinates of the UAV, it first performs image matching with the plural correction point images corresponding to the current image and the landmark coordinates to determine whether it meets one of the corrections. If the point image matches one of the correction point images, the current coordinates will be determined based on the matching correction point image; if it does not match any of the correction point images, then the current coordinates will be determined based on the current image relative to the previous image, that is, the captured previous frame of the current image. Calculate the displacement vector relative to the previous image and determine the current coordinates. Since the previous coordinates of the previous image are known, the current coordinates can be calculated based on the previous coordinates, that is, the displacement vector. Finally, based on the displacement between the current coordinates and the landmark coordinates, a flight instruction is generated to cause the UAV to fly toward the landmark coordinates.

The present invention combines two visual navigation technologies that determine the relative movement position based on previous images to estimate the current position, and perform position correction based on pre-stored correction point images corresponding to landmark coordinates, to determine the current coordinates of the UAV and generate flight instructions. When the drone is far away from the landmark coordinates, it cannot match the correction point image near the landmark coordinates, so the current coordinates are determined by the relative movement of the current image and the previous image; when the drone flies near the landmark coordinates, the current image begins to match When the calibration point image is reached, the current coordinates are determined based on the calibration point coordinates corresponding to the corresponding calibration point image.

Since the correction point images and the corresponding correction point coordinates are known, when the current coordinates are determined based on the corresponding correction point images, the current coordinates are corrected and will not be affected by visual navigation that continues to estimate the current coordinates based on previous images. The method accumulates larger and larger errors. In other words, when the UAV passes through the landmark coordinates, its current coordinates can be corrected by a certain image and certain coordinates, thereby ensuring that the current coordinates calculated by the UAV through visual navigation and the actual coordinates remain at a certain level during the flight. Within the error, the positioning accuracy of drone visual navigation is improved.

11:Camera

12:Storage

13: Processor

M: Navigation map

T1, T2: Landmark

M0’: previous image

M0: current image

M1, M2: Landmark images

M11~M14,M21~M24: surrounding images

P1,P2: landmark coordinates

P11~P14,P21~P24: surrounding coordinates

X: current coordinates

X’: previous coordinates

R: expected path

Figure 1 is a block diagram of a UAV visual navigation system for ground feature image positioning and correction according to the present invention.

Figure 2A is a schematic diagram of the navigation map of the UAV visual navigation method for ground feature image positioning and correction according to the present invention.

Figure 2B is a schematic diagram of the navigation map of the UAV visual navigation method for ground feature image positioning and correction according to the present invention.

Figure 3 is a schematic flow chart of the UAV visual navigation method for ground feature image positioning and correction according to the present invention.

Figure 4A is a detailed flow chart of step S102 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 4B is a schematic diagram of an application state of step S102 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 5A is a detailed flow chart of step S103 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 5B is a schematic diagram of an application state of step S103 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 5C is a further detailed flowchart of step S103 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 6A is a further detailed flowchart of step S104 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

6B is a schematic diagram of an application state of step S104 of the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 7A is a further detailed flowchart of step S105 in the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

7B is a schematic diagram of an application state of step S105 of the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 8A is a schematic flowchart of a preferred embodiment of the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 8B is a schematic diagram of the application state of step S109 of the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Figure 8C is a schematic diagram of another application state of step S109 of the UAV visual navigation method for positioning and correcting ground feature images according to the present invention.

Please refer to Figure 1 and Figures 2A-2B. The UAV visual navigation system for ground feature image positioning and correction of the present invention is implemented on an unmanned aerial vehicle (hereinafter referred to as UAV) and includes a camera 11, a storage 12 and a Processor 13. The camera 11 and the processor 13 are electrically connected to the memory 12. The camera 11 is used to continuously capture images directly below the drone during flight. The latest captured image is a current image. M0, and after capturing a current image, the previously captured image is stored in the memory 12 as a previous image M0'. Since the camera 11 captures images directly below the drone, the center point coordinates in the current image can be set to the location of the drone, that is, the current coordinate X to be determined, and the center point coordinates in the previous images is the previous position of the drone, that is, the previous coordinate X' that has been judged. In addition, the memory 12 is also used to store a navigation map M, a flight mission information, a plurality of landmark images M1, M2 and corresponding plurality of landmark coordinates P1, P2, and a plurality of surrounding images M11~ corresponding to each landmark image M1, M2. M14, M21~M24 and the complex surrounding coordinates P11~P14, P21~P24 corresponding to the surrounding images M11~M14, M21~M24. The aforementioned drone is, for example, a quadcopter, an octocopter, a single-copter, or a fixed-wing aircraft, and the invention is not limited thereto.

In an embodiment of the present invention, the navigation map M includes a real-life ground map of the area where the flight mission is to be performed. The navigation map M is generated, for example, by projecting a real world map into a square world map in advance, and then using Quadkey calculation or similar map partition calculation method to classify the square world map into regions according to the area where the flight mission is to be performed. Select the navigation area, generate the navigation map, and give the navigation map a normalized coordinate system.

For ease of reference, the following description will be made in the form of a simplified diagram of the navigation map M, ground features, and landmarks T1 and T2.

Please refer to FIG. 2A . In the navigation map M, by selecting a plurality of unique ground features as landmarks T1 and T2 in advance, the flight mission is a flight route planned within the scope of the navigation map M. In one embodiment, the flight route is composed of at least one and preferably a plurality of landmarks T1, T2, etc. in the navigation map M, and the flight mission information includes the sequential numbers of the landmarks T1, T2. When the drone takes off, the goal is, for example, to make the drone fly from an initial position to the landmarks T1, T2, etc., until the last landmark. In the present invention, the flight mission is taken as an example of flying to landmarks T1 and T2 in sequence.

Please refer to FIG. 2B. In one embodiment, based on the known landmarks T1 and T2, the landmark image M1 of the landmark T1, the landmark image M2 of the landmark T2, and the landmark coordinate P1 of the landmark T1 are collected and stored in advance. , landmark coordinates P2 of landmark T2, surrounding images M11~M14 of landmark T1, surrounding images M21~M24 of landmark T2, and surrounding coordinates P11~P14 of surrounding images M11~M14, and surrounding coordinates P21~P24 of surrounding images M21~M24. .

The landmarks T1 and T2 are set with known landmark coordinates P1 and P2 in the navigation map M, and the landmark images M1 and M2 are based on the landmark coordinates P1 and P2 as the center points P1 and P2 before performing the flight mission. Ground images collected in advance. In addition, a plurality of surrounding images M11~M14 and M21~M24 adjacent to the landmarks T1 and T2 are also collected in advance. In the embodiment of FIG. 2B , for example, there are four surrounding images M11 ~ M14 and M21 ~ M24 located above, below, left and right of the landmarks T1 and T2 . more detailed Said that the surrounding images M11~M14 and M21~M24 are, for example, plural ground surfaces captured based on the landmark coordinates P1 and P2 in the north, south, east, and west of the navigation map M, and at a specific distance away from the center point. image, so the center points of the surrounding images M11~M14, M21~M24 are the surrounding coordinates P11~P14, P21~P24. In addition, you can also arbitrarily select surrounding areas with more obvious ground features in the adjacent landmark coordinates P1 and P2 to capture multiple surrounding images, and the number of surrounding images can also be determined according to the comparison requirements, as long as the reference coordinates of these surrounding images It only needs to be known for certain, and the present invention is not limited thereto.

Please refer to Figure 3. The UAV visual navigation method for ground feature image positioning and correction of the present invention is executed by the processor 13 and includes the following steps: S101: Capture a current image M0 and read a previous image M0' And a previous coordinate X' corresponding to the previous image M0'; wherein, the previous coordinate The relative position of a landmark coordinate on the navigation map M, read the plural correction point images M1, M12, M13 corresponding to the landmark coordinate P1; S103: Match the current image M0 with the correction point images respectively to determine Whether the current image M0 matches one of the correction point images M1, M12, and M13; S104: If the current image M0 matches one of the correction point images, determine based on the current image M0 and the coordinates of a correction point corresponding to the matching correction point image. A current coordinate X; S105: If not, determine the current coordinate X based on the current image M0, the previous image M0' and the corresponding previous coordinate flight instructions.

Please refer to Figures 4A and 4B together. In an embodiment of the present invention, when step S102 is executed, the following sub-steps are included: S1021: Read a landmark image M1 based on the landmark coordinate P1, and read a plurality of surrounding images M11~M14 related to the landmark image M1. These surrounding images M11~M14 have corresponding surrounding coordinates P11~P14; S1022: According to an expected path R between the previous coordinate M11~M14 are the correction point images M1, M12, and M13.

In step S102, when the previous coordinate read by the drone is X', the processor 13 first reads the corresponding landmark image M1 and related surrounding images M11~M14 according to the landmark coordinate P1 of the current target landmark T1, and An expected path R between the previous coordinate X' and the landmark coordinate P1 is calculated. In the embodiment of FIG. 4B , the expected path R is calculated as a connection between the previous coordinate X' and the landmark coordinate P1. When selecting the possible corrected point image, the surrounding coordinates P11 to P14 are selected from the landmark image M1 and the surrounding images M11 to M14, and the vertical distance from the expected path R is within a preset distance. Images M11~M14 are calibration point images. In addition, since the landmark coordinate P1 is the end point of the expected path R, the landmark image M1 will inevitably be selected as one of the correction point images. Taking the example shown in Figure 4B as an example, the vertical distance between the surrounding coordinates P11~P14 and the expected path R is less than the preset distance is the surrounding images M12 and M13, so the correction point images are the surrounding images M12, M13 and the landmark image. M1.

By selecting the correction point image and correction point coordinates through S102 and its sub-steps S1021~S1022, surrounding images M11, M14, M21~M24 with low passing probability can be dynamically eliminated based on the previous coordinate X' and the target landmark coordinate P1. When determining the matching between the current image M0 and the calibration point images M1, M12, and M13 to determine whether it is close to the landmark T1, the processor 13 does not need to compare all surrounding images M11~M14, M21~M24 to achieve reduction The purpose of computing resource consumption.

Please refer to Figures 5A and 5B. In an embodiment of the present invention, when step S103 is executed, the following sub-steps are performed on each correction point image: S1031: Perform feature point extraction and feature point matching on the current image M0 and the correction point images M1, M12, and M13 respectively to generate plural matching feature point pairs; S1032: Calculate the matching credibility of the matching feature point pairs and Matching angle; S1033: Determine whether the number of matching feature point pairs is greater than a quantity threshold, whether the average matching credibility of each matching feature point pair is greater than a credibility threshold, and whether the average matching of each matching feature point pair is greater than a credibility threshold Whether the angle meets an angle condition; S1034: If so, determine that the current image M0 meets the correction point M13.

In step S103, as shown in Figure 5B, taking the surrounding image M13 of one of the correction point images as an example, the image on the left is the current image M0 at a certain moment, and the image on the right is the surrounding image M13. The complex dotted lines connecting both sides are is the connecting line L of the matching feature point pairs after feature point extraction and matching. Among them, the matching credibility is a credibility value of each matching feature point pair calculated through the matching algorithm, the matching number is the total number of all matching feature point pairs generated, and the matching angle is each matching feature point. An angle relative to a horizontal direction. Among them, the specific value of the credibility value has different value ranges depending on the matching algorithm used, and the credibility threshold can be set by the user. The matching angle can vary according to the settings of the coordinate axis and horizontal direction. Take the current image M0 and the correction point image M13 side by side in Figure 5B as an example. When the connecting line L of the matching feature point pair is closer to the left and right horizontal lines in the image, that is, the matching angle is close to 180°, it means that the current image M0 is more consistent. This correction point image is M13. Preferably, a deep learning model is used to extract and match feature points between the current image M0 and the correction point images M1, M12, and M13. The interpretability and matching accuracy of the feature point extraction of the deep learning model are better than traditional methods. good.

After the feature point extraction and matching of the current image M0 and the correction point images M1, M12, and M13 are completed, three values are further calculated: the number of matching feature point pairs, the average matching credibility of all matching feature point pairs, and all matching features. The average matching angle of the point pair, and only when the above three values meet the requirements of greater than the quantity threshold, greater than the credibility threshold, and meet the angle conditions, the judgment will be made. The current image M0 matches the correction point image, thereby improving the matching accuracy of the current image M0 and the correction point image.

Please refer to Figure 5C. In a better embodiment, when it is determined that the number of matching feature point pairs is greater than the number threshold, the average matching credibility of each matching feature point pair is greater than the credibility threshold, and After the average matching angle of each matching feature point pair meets the angle condition, the following steps are further included: S1033A: Determine whether the number of matching feature point pairs between the current image M0 and the correction point image is greater than the number threshold. Times threshold; S1033B: Determine whether the average matching credibility of the current image M0 and the correction point image is greater than the credibility threshold; S1033C: Determine the current image M0 and the correction point Whether the number of times the average matching angle of the image meets the angle condition is greater than an angle number threshold; if so, the current image M0 is judged to match the correction point image.

In this preferred embodiment, when the current image M0 meets the three numerical conditions, it is further determined that when a plurality of consecutive current images M0 meet the respective thresholds of times for each of the numerical conditions, the correction point image is further used to determine the current coordinate X . It should be understood by those with ordinary knowledge in the field that the execution order of steps S1033A~S1033C can be adjusted and achieve the same technical effect. For example, it can be set that when the number of matching feature point pairs between the current image M0 and the correction point image is greater than the quantity threshold (the quantity threshold is 1), there have been 3 consecutive frames of the current image M0 and the correction point image. The matching credibility is greater than the credibility threshold (the credibility threshold is 3), and the average matching angle between the previous image of 5 consecutive frames and the correction point image meets the angle condition (the credibility threshold is 3). value is 5), it is judged that the current image M0 matches the correction point image to avoid possible misjudgments and ensure the correctness of the correction point image.

In step S103, if it is determined that the current image M0 matches one of the correction point images, the correction point image is used to determine the current coordinate X, that is, step S104 is performed. If it is determined that the current image M0 does not match any correction point image, the current coordinate X is determined based on the previous image M0', that is, step S105 is performed.

Please refer to Figures 6A and 6B. When step S104 is executed, the following sub-steps are included: S1041: Calculate the current image M0 relative to the correction point according to the matching feature point pairs in the current image M0 and the correction point image. A transformation matrix of the point image; S1042: Calculate the current coordinate X according to the correction point coordinates and the transformation matrix of the current image M0 relative to the correction point image.

As shown in FIG. 6B , continuing the example of FIG. 5B , if the current image M0 matches one of the correction point images (M13), the transformation matrix is further calculated through the matching feature point pair generated in step S103. According to the transformation matrix, the center point of the current image M0 on the left, that is, the current coordinate X, is projected to the correction point image M13 on the right. Since the correction point coordinate P13 is known, the current coordinate X can be calculated based on the correction point coordinate P13 and the relative position of the center point of the current image M0 in the projection of the correction point image M13 (X in the right image) . It should be emphasized that, as explained for step S103, the matching feature point pairs between the current image and the correction point image here are generated by using deep learning model feature point extraction, which has a strong computing foundation and better accuracy. Therefore, when calculating the current coordinate X through the correction point image as the basis for position correction of the drone, it can provide good accuracy and correction effect.

Please refer to Figures 7A and 7B. Before performing step S105, the following steps are first performed: S105A: Extract feature points of the current image M0 and the previous image M0' respectively to generate plural feature points respectively; S105B: Match the feature points in the current image M0 and the feature points in the previous image M0' to calculate a transformation matrix of the current image M0 relative to the previous image M0'; and when executing S105, the following content is executed: S105': Calculate the current coordinate X according to the previous coordinate X' and the transformation matrix of the current image M0 relative to the previous image M0'.

Among them, steps S105A and S105B are pre-steps for executing step S105 (S105'), for example, they are executed simultaneously with step S103. In addition, in step S103, if the determination result is "yes", step S105 will not be executed. In steps S105A and S105B, the processor 13 extracts the feature points of the current image M0 and the previous image M0', and then generates a transformation matrix, or a projection matrix, based on the feature points of the current image M0 and the previous image M0'. The transformation matrix represents a relative displacement of the current image M0 relative to the previous image M0' or the correction point image. Finally, in step S105', the current coordinate X is calculated using the transformation matrix. The effect produced is shown in Figure 7B, in which the right side is the previous image M0', the left side is the current image M0, and the dotted arrow is a projection of the transformation matrix calculated by the processor 13 based on the feature points of the previous image M0' and the current image M0. . According to the transformation matrix, the center point of the current image M0 is projected into the previous image M0'. Since the previous coordinate X' is known, based on the previous coordinate 'The relative position of the projection can be calculated as the current coordinate X.

Preferably, the feature point extraction of the current image M0 and the previous image M0' or the correction point image is performed through a scale-invariant feature transform (SIFT) image feature point extraction algorithm. The matching of the feature points in the current image M0 and the feature points in the previous image M0' or the correction point image is performed through a K Nearest Neighbors (KNN) algorithm.

Please refer to Figures 8A and 8B together. The present invention further includes the following steps: S107: Calculate the distance between the current coordinate X and the landmark coordinate P1; S108: Determine whether the distance is less than a landmark arrival threshold; S109: If so, determine that the landmark coordinate P1 has been reached, update the landmark coordinates and the landmark image according to a flight mission information, and return to step S101; if not, directly Return to step S101.

Taking the navigation map M and the current coordinate X of the drone shown in Figure 8B as an example, in order to further prevent the drone from leaving early before actually arriving at the landmark coordinate P1, after the current coordinate Is the distance between X and the landmark coordinate P1 close enough? If the distance is less than the landmark arrival threshold, it is determined that the current coordinate According to the flight mission information, the processor 13 will update the current landmark coordinate P1 as the target to the landmark coordinate P2 of the next target, and update and read the corresponding landmark image M2, surrounding images M21~M24, surrounding coordinates P21~P24, etc. information to continue the flight mission to the next landmark coordinate P2.

Please refer to Figure 8C. When the UAV determines that it has arrived at the landmark coordinate P1, the processor 13 will first clear the calibration point image and the calibration point coordinates, and update the landmark coordinate P1 to the new landmark coordinate P2 based on the flight mission information, and Based on the current previous coordinates As shown in FIG. 8C , the current coordinate Select the surrounding images M11~M14, M21~M24 and the landmark image M2 whose vertical distance from the expected path R is within the preset distance as new correction point images. In the example of FIG. 8C , the processor 13 selects the surrounding images M14, M22, M23 and the landmark image M2 as the correction point images.

In addition, in one embodiment, when the processor 13 determines that the landmark coordinates have been reached, it first determines whether there are any landmark coordinates that have not yet arrived based on the flight mission information; if so, it updates the landmark coordinates, landmark images, and surroundings based on the flight mission information. image; if not, the mission will be ended.

In summary, compared with traditional feature point extraction and matching, which is less accurate and traditional deep learning feature point extraction and matching, which consumes more computing resources, the UAV visual navigation method and system for positioning and correcting ground feature images of the present invention It has a high matching speed when performing self-positioning based on the previous image M0', and requires accuracy when matching based on landmark images, thus combining the advantages of the two positioning methods to help the drone fly smoothly on the navigation map M. By setting the correction point adoption mechanism based on the relative positions of the previous coordinates Whether M0 conforms to the calibration point image, this can effectively improve the accuracy of judgment and eliminate most misjudgments. By collecting the surrounding images of the landmark image in advance and including them in the judgment of the correction point image, the position of the drone relative to the landmark image can be effectively determined, helping the matching process and avoiding early misjudgment that the landmark has been reached, causing the drone to turn early. situation occurs.

S101~S106: Steps

Claims (9)

A UAV visual navigation method for ground feature image positioning and correction, executed by a processor, includes the following steps: acquiring a current image below a UAV, and reading a previous image and a previous image corresponding to the previous image coordinates; wherein, the previous image is the previous frame image captured before the current image is captured; based on the previous coordinates and a landmark coordinate, read a plurality of correction point images corresponding to the landmark coordinates; combine the current image with the Image matching is performed on the other correction point images to determine whether the current image matches one of the correction point images; if the current image matches one of the correction point images, based on the current image, the matching correction point image and a corresponding correction point Determine a current coordinate based on the point coordinates; if not, determine the current coordinates based on the current image, the previous image and the corresponding previous coordinates; and generate a navigation flight command based on the current coordinates and the landmark coordinates to direct the UAV towards This landmark coordinate flight. The UAV visual navigation method for locating and correcting ground feature images as described in claim 1, wherein when performing the step of "reading a plurality of correction point images corresponding to the landmark coordinates based on the previous coordinates and a landmark coordinate", The system includes the following sub-steps: reading a landmark image according to the landmark coordinates, and reading a plurality of surrounding images related to the landmark image, and the surrounding images have corresponding surrounding coordinates; according to the relationship between the previous coordinates and the landmark coordinates An expected path, select landmark images or surrounding images whose distance from the landmark coordinates or surrounding coordinates to the expected path is less than a preset distance threshold as the correction point images. The UAV visual navigation method for locating and correcting ground feature images as described in request item 1, wherein when executing "match the current image with the correction point images respectively to determine whether the current image conforms to one of the corrections" point image", the following sub-steps are performed on each correction point image: feature point extraction and feature point matching are performed on the current image and the correction point image respectively to generate a plurality of matching feature point pairs; calculate the The matching credibility and matching angle of the matching feature point pairs; where the matching angle is an angle value of the connection line of each matching feature point pair relative to a horizontal direction; determine whether the number of matching feature point pairs is greater than a quantity threshold, Whether the average matching credibility of each matching feature point pair is greater than a credibility threshold, and whether the average matching angle of each matching feature point pair meets an angle condition; if so, it is determined that the current image meets the correction point image. The UAV visual navigation method for ground feature image positioning and correction as described in claim 3, wherein when it is determined that the number of matching feature point pairs is greater than the number threshold, the average matching credibility of each matching feature point pair is greater than the The credibility threshold, and the average matching angle of each matching feature point pair meets the angle condition, further including the following steps: Determine whether the number of matching feature point pairs between the current image and the correction point image is greater than the number threshold Greater than a number of times threshold; determine whether the average matching credibility of the current image and the correction point image is greater than the credibility threshold; determine whether the average matching credibility of the current image and the correction point image is greater than a credibility number threshold; determine the average matching credibility of the current image and the correction point image Whether the number of times the matching angle meets the angle condition is greater than an angle number threshold; if so, the current image is judged to match the correction point image. The UAV visual navigation method for locating and correcting ground feature images as described in claim 2, wherein the step of "determining the current coordinates based on the corresponding correction point coordinates of the corresponding correction point image" further includes the following sub-steps: Calculate a transformation matrix between the current image and the correction point image based on the matching feature point pairs in the current image and the correction point image; calculate based on the correction point coordinates and the transformation matrix between the current image and the correction point image the current coordinates. The UAV visual navigation method for locating and correcting ground feature images as described in claim 2, wherein when the corresponding correction point image is the landmark image, the method further includes the following steps: calculating the distance between the current coordinates and the landmark coordinates. ; Determine whether the distance is less than a landmark arrival threshold; if so, determine whether the landmark coordinates have been reached, and update the landmark coordinates and the landmark image according to a flight mission information. The UAV visual navigation method for locating and correcting ground feature images as described in claim 1, wherein before performing the step of "determining the current coordinates based on the current image, the previous image and the corresponding previous coordinates", It includes the following steps: extracting feature points of the current image and the previous image respectively to generate a plurality of feature points; matching the feature points in the current image and the feature points in the previous image to calculate the relative position of the current image to the previous image. A transformation matrix of the image; and when executing the step of "determining the current coordinates based on the current image, the previous image and the corresponding previous coordinates", it is based on the previous coordinates and the transformation of the current image relative to the previous image The matrix calculates the current coordinates. The UAV visual navigation method for positioning and correcting ground feature images as described in request 1 further includes: Store the navigation map, the landmark images, landmark coordinates, the surrounding images, and the surrounding image information; wherein the flight mission information includes the sequence number corresponding to each of the landmark coordinates. A UAV visual navigation system for positioning and correcting ground feature images, including: a camera that captures a current image; a memory that is electrically connected to the camera and stores a plurality of previous images and corresponding previous coordinates, a navigation map, and a Flight mission information, a plurality of landmark images and a plurality of corresponding landmark coordinates, a plurality of surrounding images corresponding to each landmark image, and a plurality of surrounding coordinates corresponding to the surrounding images; a processor electrically connected to the storage to execute the requested item UAV visual navigation method for ground feature image positioning and correction described in any one of 1 to 8.

Images