Object positioning method, system and equipment for dynamic environment and medium

Abstract

The invention discloses an object positioning method, system and device for a dynamic environment and a medium, and relates to the technical field of visual SLAM (Simultaneous Localization and Mapping), and the method comprises the steps: obtaining a continuous image data set of object movement in the dynamic environment; processing two adjacent frames of image pairs through a FlowNet2 optical flow method to generate optical flow images of the image pairs; obtaining a complete contour of the dynamic object according to the optical flow of the image pair; calculating geometric center coordinates of each contour, and taking the obtained coordinates as dynamic feature points; segmenting the image pair into a plurality of masks through FastSAM, and splicing the masks containing the dynamic feature points to obtain a complete mask of a dynamic object in the image pair; positioning the dynamic object according to the complete mask of the dynamic object; according to the method, FastSAM and Flownet2 optical flow motion detection methods are combined, so that the accuracy of positioning a dynamic object in a dynamic scene is improved.

Description

Technical Field

[0001] The present invention relates to the field of visual SLAM technology, and in particular to a method, system, device and medium for positioning an object in a dynamic environment. Background Art

[0002] With the development of science and technology in the field of robotics, SLAM (Simultaneous Localization and Mapping), as a key technology in the field of robotics, has been widely used in fields such as autonomous driving, drones, and AR / VR. The current SLAM system is mainly divided into visual SLAM and laser SLAM according to the sensors used to obtain environmental information. Since visual sensors can provide SLAM systems with richer image information, visual SLAM has gradually become a popular research direction for researchers in the fields of robotics, autonomous driving, and AR / VR. Among them, visual sensors are divided into monocular cameras, binocular cameras, and RGB-D cameras. Visual sensors provide rich RGB information for SLAM systems as well as depth map information, and are often used in visual SLAM systems.

[0003] At present, an ORB_SLAM algorithm based on ORB features has been proposed. With the good robustness and fast processing speed of ORB features, it has high robustness and accuracy in static scenes and is widely used in intelligent mobile robots, autonomous driving and other fields. The ORB-SLAM2 algorithm has also been proposed. Based on the ORB-SLAM algorithm, the algorithm adds support for binocular and RGB-D cameras. Subsequently, ORB_SLAM3 was proposed, which added support for fisheye cameras and IMUs, and introduced the atlas multi-map mechanism.

[0004] However, the current positioning methods based on ORB_SLAM, ORB-SLAM2 and ORB_SLAM3 are usually applied to static scenes and cannot achieve the expected accuracy in highly dynamic environments, thereby reducing the accuracy of positioning dynamic objects in dynamic scenes. Summary of the invention

[0005] In view of the fact that the prior art cannot achieve the expected accuracy in a high dynamic environment, thereby reducing the accuracy of positioning dynamic objects in dynamic scenes, the present invention proposes an object positioning method, system, device and medium for a dynamic environment. Through the motion detection method of Flownet2 optical flow, the optical flow can output the specific contour of the dynamic object, and based on the object contour, the dynamic feature points are obtained, FastSAM is introduced and optimized, and combined with the optical flow method to output a complete mask of the dynamic object; thereby solving the problems existing in the prior art.

[0006] A method for positioning an object in a dynamic environment comprises the following steps: Acquire a continuous image data set of a moving object in a dynamic environment to be tested; The FlowNet2 optical flow method is used to process two adjacent frames of images in the continuous image data set of moving objects to generate an optical flow image of the image pair; the complete outline of each moving object in the image pair is obtained based on the optical flow image; the geometric center coordinates of each outline are calculated, and the obtained coordinates are used as dynamic feature points; the current frame image is divided into multiple masks by FastSAM, and the masks containing dynamic feature points are spliced ​​to obtain the complete masks of all moving objects in the image; The moving objects in the dynamic environment to be measured are located based on the complete masks of all moving objects.

[0007] Furthermore, the method of processing two adjacent image pairs in a continuous image data set of a moving object by using the FlowNet2 optical flow method to generate an optical flow image of the image pair specifically includes the following steps: The optical flow data between two adjacent frames is obtained by FlowNet2 optical flow method [D x ,D y ]; where each pixel in the image is traversed, with the traversed pixel as the origin of the coordinate, the horizontal axis is the x-axis, the vertical axis is the y-axis, D x is the optical flow data in the x-axis direction, D y is the optical flow data in the y-axis direction; Will Determine the optical flow at the point (x, y) and generate an optical flow image for the image pair.

[0008] Furthermore, it also includes introducing Gaussian blur to process the optical flow image of the image pair to generate an optical flow image after Gaussian blur; the processing process specifically includes the following steps: According to the two-dimensional Gaussian distribution, a weight matrix is ​​generated, the size of which is σ×σ; each element in the matrix represents the weight of the pixel at the corresponding position; The weight matrix is ​​convolved with the original image; specifically, the weight matrix is ​​covered on a certain area of ​​the optical flow image, and the weighted sum of the weight matrix and the pixel value of the covered area is calculated; The weighted sum is used as the pixel value of the corresponding position in the output image; The convolution operation is repeated by sliding the weight matrix, traversing the entire image and generating a Gaussian blurred optical flow image.

[0009] Furthermore, the two-dimensional Gaussian distribution is expressed as: , Here, σ is the standard deviation, which is the same in both the x-direction and the y-direction.

[0010] Furthermore, before positioning the moving objects in the dynamic environment to be measured according to the complete masks of all moving objects, the dynamic feature points covered by the complete masks of the dynamic objects are removed.

[0011] The present invention also includes an object positioning system for a dynamic environment, comprising: An acquisition module is used to acquire a continuous image data set of a moving object in a dynamic environment to be tested; The calculation module is used to process two adjacent frames of images in the continuous image data set of moving objects through the FlowNet2 optical flow method to generate an optical flow image of the image pair; obtain the complete outline of each moving object in the image pair based on the optical flow image; calculate the geometric center coordinates of each outline and use the obtained coordinates as dynamic feature points; divide the current frame image into multiple masks through FastSAM, and splice the masks containing dynamic feature points to obtain the complete masks of all moving objects in the image; The positioning module is used to locate the moving objects in the dynamic environment to be measured based on the complete mask of all moving objects.

[0012] The present invention also includes a computer device for positioning objects in dynamic environments, including: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the method for positioning objects in dynamic environments when executing the computer program.

[0013] The present invention also includes a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the object positioning method for a dynamic environment.

[0014] The present invention provides a method, system, device and medium for positioning an object in a dynamic environment, which has the following beneficial effects: The present invention adopts the motion detection method of Flownet2 optical flow in the detection of moving objects in a dynamic environment, so that the optical flow can output the specific contour of the moving object, and obtain dynamic feature points based on the object contour; in the segmentation of the moving object image, FastSAM is introduced to optimize the motion detection method of Flownet2 optical flow, and the speed is increased by nearly 50 times while ensuring the accuracy, thereby ensuring the real-time performance of DFS-SLAM and reducing the problem of GPU resource expansion in the process of continuous image processing; the method combines FastSAM with the motion detection method of Flownet2 optical flow, improves the accuracy of dynamic object positioning in dynamic scenes, and has higher accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of a method for positioning an object in a dynamic environment according to an embodiment of the present invention; Figure 2 It is a structural diagram of DFS-SLAM in an embodiment of the present invention; Figure 3 is a flow chart of an adaptive optical flow method in an embodiment of the present invention; Figure 3 (a) in the figure is the previous frame image. Figure 3 (b) in the figure is the current frame image. Figure 3 (c) in the figure is the optical flow graph calculated by FlowNet2. Figure 3 (d) in the figure is the optical flow map after Gaussian blur. Figure 3 (e) in is the optical flow grayscale image, Figure 3 (f) in the figure is the outline of a dynamic object. Figure 3 (g) is a schematic diagram of the final result; Figure 4 is a schematic diagram of segmentation of multiple masks in an embodiment of the present invention; Figure 4 (a) in the figure is the current frame image. Figure 4 (b) in the figure is the mask image of the current frame after SAM segmentation. Figure 4 (c) in the figure is a mask image containing dynamic cue points. Figure 4 (d) in the figure is the final mask image obtained by adding the masks in (c); Figure 5 A schematic diagram of segmenting dynamic objects in an environment according to an embodiment of the present invention; Figure 5 (a) is the previous frame image, Figure 5 (b) is the current frame image, Figure 5 (c) is the final mask image; Figure 6 ATE diagrams of DFS-SLAM and ORB-SLAM3 in eight dynamic scene sequences in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0017] The present invention proposes a method for locating objects in a dynamic environment. Figure 1 As shown, the specific steps include: S1. Obtain a continuous image data set of the motion of a dynamic object in a dynamic environment to be tested.

[0018] S2. Input the continuous image data set of the dynamic object motion into the dynamic object positioning system DFS-SLAM to obtain the complete mask of the dynamic object.

[0019] S2.1. Send the continuous image data set of dynamic object motion into DFS-SLAM, perform image segmentation through the SAM model (segmentanything model), and process the two adjacent frames of the continuous image data set of dynamic object motion through the FlowNet2 optical flow method to generate the optical flow image of the image pair. Among them, DFS-SLAM is a SLAM system based on ORB-SLAM3. ORB-SLAM3 has three major threads: tracking, mapping, and loop detection. After being tested by most public data sets, ORB-SLAM3 is the most accurate algorithm at present. Therefore, ORB-SLAM3 is used as the basic framework to provide global positioning and mapping functions. The flowchart of DFS-SLAM is as follows: Figure 2 As shown in the figure, ORB-SLAM3 is an open source visual SLAM system that supports monocular, binocular and RGB-D cameras. It is known for its high efficiency, robustness and modular design, and can perform real-time positioning and map construction. ORB-SLAM3 uses ORB feature points for matching and tracking, and optimizes the camera pose and map point position through graph optimization algorithms. It supports closed-loop detection to eliminate drift and generate dense maps. ORB-SLAM3 is widely used in drones, autonomous driving, robots and other fields, providing precise positioning and navigation support. DFS-SLAM obtains the mask of dynamic objects in the scene when the image is input by introducing the large model FastSAM and FlowNet2, and removes dynamic feature points in the tracking thread to improve positioning accuracy.

[0020] S2.2. Obtain the complete outline of the dynamic object based on the optical flow image; calculate the geometric center coordinates of each outline, use the obtained coordinates as dynamic feature points, divide the image pair into multiple masks through FastSAM, and splice the masks containing dynamic feature points to obtain the complete mask of the dynamic object in the image pair. Specifically, traverse each pixel point in the image, take the traversed pixel point as the coordinate origin, the horizontal direction is the x-axis, and the vertical direction is the y-axis; obtain the optical flow data between two adjacent frames through the FlowNet2 optical flow method [D x ,D y ], D x is the optical flow data in the x-axis direction, D y is the optical flow data in the y-axis direction. Determined as the optical flow at the point (x,y).

[0021] In order to remove the optical flow noise caused by camera motion, Gaussian blur is introduced, and its two-dimensional Gaussian function is as follows: ; Among them, σ is the same standard deviation in the x-direction and the y-direction, and (x, y) represents the pixel point with the selected pixel point as the origin of the coordinate axis at its coordinate (x, y).

[0022] According to the two-dimensional Gaussian distribution, a weight matrix (also called Gaussian kernel) of size σ×σ can be generated. In DFS-SLAM, σ is set to 9. Each element in this matrix represents the weight of the pixel at the corresponding position. The weight matrix is ​​convolved with the original image. Specifically, the weight matrix is ​​covered on a certain area of ​​the optical flow image, the weighted sum of the weight matrix and the pixel value of the covered area is calculated, and then the weighted sum is used as the pixel value of the corresponding position in the output image. By sliding the weight matrix and repeating the above convolution operation, the entire image can be traversed and a Gaussian blurred optical flow image can be generated. The blurred optical flow image is grayed and mapped to [0,255]. The weight matrix calculated above is used to calculate the Gaussian weighted average for each pixel, and the values ​​greater than the mean are set to 255 to obtain the complete outline of the dynamic object. The geometric center of each outline is then calculated and used as a dynamic feature point, which is combined with the large model FastSAM to obtain a complete dynamic object mask. The operation sequence of the optical flow driven method is as follows: Figure 3 As shown, Figure 3 (a) is the previous frame image. Figure 3 (b) in the figure is the current frame image. Figure 3 (c) in the figure is the optical flow graph calculated by FlowNet2. Figure 3 (d) in the figure is the optical flow map after Gaussian blur. Figure 3 (e) in is the optical flow grayscale image, Figure 3 (f) in the figure is the outline of a dynamic object. Figure 3 (g) in the figure is the final result.

[0023] S3. Input the complete mask of the dynamic object into the tracking thread, remove the dynamic feature points covered by the mask, and realize the positioning of the dynamic object.

[0024] The present invention proposes an object positioning method for dynamic environments. First, in the detection of dynamic objects, the motion detection method based on Flownet2 optical flow is improved and used for moving objects in dynamic environments, so that the optical flow can output the specific contour of the dynamic object, and based on the object contour, the dynamic feature points are obtained; in the segmentation of moving objects, FastSAM is introduced and optimized, which ensures real-time performance while reducing the use of GPU resources, and it is combined with the improved optical flow method to output the complete mask of the dynamic object; based on ORB-SLAM3 as the framework, a complete RGB-D SLAM system (SLAM system using RGB-D camera as input) is proposed, which has higher accuracy and robustness in dynamic scenes compared with the original ORB-SLAM3. When it is uncertain which things in the environment are dynamic, it is a good choice to segment all elements in the environment. The segment anything model (SAM) has a powerful segmentation function. When a picture is input into SAM, it can accurately identify different elements in the picture and segment them into multiple masks. Although the original SAM has good segmentation accuracy, it is not real-time capable. It usually takes more than one second to completely segment an image, which is obviously not suitable for SLAM systems. In consideration of real-time performance, DFS-SLAM introduced FastSAM, which increases the speed by nearly 50 times while ensuring accuracy, ensuring the real-time performance of DFS-SLAM, and optimizing FastSAM, reducing the problem of GPU resource expansion during continuous image processing, making the entire SLAM system occupy fewer resources.

[0025] like Figure 4 As shown, Figure 4 (a) is the current frame, Figure 4 (b) is the mask of the current frame after SAM segmentation. Figure 4 (c) is the mask containing the dynamic cue points. Figure 4 (d) is the final mask obtained by adding the masks in (c); FastSAM divides the image into multiple masks, and splices the masks containing dynamic prompt points to obtain the masks of all dynamic objects in the image. The spliced ​​masks reveal the segmentation strategy: objects containing dynamic objects will also be segmented. For example, when only the hands of a person are moving, since the mask of the entire person includes the mask of his hands, the complete mask of the person will also be counted. Subsequent experiments have also proved that this strategy does not have a significant impact on SLAM positioning.

[0026] like Figure 5As shown in the figure, in order to reflect the feature of no prior knowledge, a scene of soy sauce being poured into water is selected from the d2-nerf dataset. The segmentation algorithm proposed in the present invention successfully segments the dynamic objects in the environment. It can be concluded that under the premise of no prior knowledge, the segmentation algorithm proposed in the present invention can still segment the dynamic objects, among which, Figure 5 (a) is the previous frame, Figure 5 (b) is the current frame, Figure 5 (c) is the final mask.

[0027] Experimental Proof: DFS-SLAM will be experimentally evaluated and demonstrated. First, the tracking performance is evaluated using two public datasets. Second, it is demonstrated that good accuracy can be achieved without prior knowledge, and the advantages of the fusion algorithm over the individual algorithms are analyzed. The experiments are mainly conducted on TITAN Xp, with Ubuntu 18.04 as the system environment.

[0028] Performance evaluation of TUM RGB-D dataset: TUM RGB-D dataset is a large dataset provided by the Computer Vision Group of Technical University of Munich to create a new benchmark for visual odometry and SLAM systems. In order to evaluate the accuracy and robustness of DFS-SLAM system in dynamic scenes, the experiment mainly uses eight sequences under the dynamic object category in the dataset, of which the first four are high dynamic scene sequences and the last four are low dynamic scene sequences as supplement.

[0029] There are currently two main error evaluation indicators. One is the absolute trajectory error (ATE), which is used directly to measure the difference between the basic trajectory and the estimated trajectory. The other is the relative pose error (RPE), which is mainly used to measure rotational drift and translational drift. The present invention selects ATE as the evaluation indicator. In order to evaluate the performance improvement relative to the original system, the experimental results of DFS-SLAM are compared with ORB-SLAM3. The evaluation comparison results in eight dynamic scene sequences are shown in Table 1. Experiments show that after introducing the optical flow method and semantic segmentation algorithm into the traditional SLAM algorithm, the positioning accuracy is significantly improved for high dynamic environments, but the effect is almost the same in low dynamic environments. Figure 6 As shown in the figure, the line with * is the real trajectory, the solid line is the trajectory obtained by the present invention, and the dotted line is the difference between the estimated trajectory and the real trajectory. It shows the ATE diagram of DFS-SLAM and ORB-SLAM3. From the figure, we can more intuitively feel the improvement of the accuracy of DFS-SLAM in dynamic environments.

[0030] Table 1 Evaluation comparison results in eight dynamic scene sequences

[0031] In order to further evaluate the effectiveness of the proposed algorithm, we continue to compare it with Dyna-SLAM and Refusion, and the results are shown in Table 2. The scene captured by the w-static dataset is two people sitting, then standing up and swapping positions. In this scene, when both people are sitting, the movement amplitude is not large, and the optical flow method cannot determine the complete outline. Therefore, the feature points on the person are not completely eliminated, which affects the positioning accuracy. The decrease in accuracy in the s-static dataset is also due to the above reasons. From the experimental data of other datasets, it can be found that the positioning accuracy of DFS-SLAM in dynamic environments is slightly better than Dyna-SLAM and Refusion.

[0032] Table 2 Comparison results of DFS-SLAM, Dyna-SLAM and Refusion

[0033] Performance evaluation of Bonn dataset: The Bonn RGB-D dynamic dataset is used to evaluate the dataset of 24 dynamic sequences of RGB-D SLAM. In order to verify the generalization performance of DFS-SLAM, eight dynamic environment datasets in the dataset are selected for another experimental evaluation. In the eight selected scenes, the "person_tracking" sequence and the "person_tracking2" sequence show that the camera tracks a person from left to right. The "placing_no_box" sequence and the "placing_no_box2" sequence show that a person moves a box in and puts it on the ground. The "removing_no_box" sequence and the "removing_no_box2" sequence show that a person moves a box from the ground. The "synchronous" sequence and the "synchronous2" sequence present a scene where several people jump together in the same direction over and over again. In order to evaluate the accuracy of the algorithm, DFS-SLAM is compared with the original ORB-SLAM3 and Dyna-SLAM, Refusion. The experimental results are shown in Table 3 below.

[0034] Table 3 Comparison results of DFS-SLAM, ORB-SLAM3, Dyna-SLAM, and Refusion

[0035] In the "person_tracking" sequence and the "synchronous" sequence, the effect is not as good as Dyna-SLAM. The main reasons are as follows: In the "person_tracking" dataset, the movement of the person and the movement of the camera are not much different, which makes Flownet2 unable to accurately identify the moving person in the scene, so it cannot completely eliminate the feature points on the person, resulting in a decrease in accuracy. In the "synchronous" dataset, because the moving person occupies most of the space, and the person moves in translation, the static feature points also appear in the form of "translation" at the beginning, resulting in a significant deviation in the positioning of ORB-SLAM3, so the accuracy will differ so much.

[0036] Table 4 shows the running time of each part of DFS-SLAM, where all_time represents the time required to run one frame. Compared with other modules, optical flow processing takes up a lot of time because it is necessary to traverse each pixel in this part to achieve Gaussian blur and calculate Gaussian weighted average. Another reason for the speed limitation is that the experiment was conducted on TITAN XpGPU. If it is replaced with an experimental platform with higher computing power, the running time may be greatly reduced.

[0037] Table 4 Running time of each part of DFS-SLAM

[0038] The present invention proposes a SLAM method for dynamic environments without prior knowledge. First, FastSAM is used to segment the image, and the dynamic area in the image is determined using a dense optical flow method with an adaptive threshold. Since DFS-SLAM does not have any prior knowledge, it does not need to be trained by a pre-trained deep learning network model, and still has good accuracy even in a highly dynamic environment.

[0039] Based on the same inventive concept, the present invention also proposes an object positioning system for a dynamic environment, comprising: The acquisition module is used to acquire a continuous image data set of a moving object in a dynamic environment to be tested.

[0040] The calculation module is used to process two adjacent frames of image pairs in the continuous image data set of moving objects through the FlowNet2 optical flow method to generate an optical flow image of the image pair; obtain the complete outline of each moving object in the image pair based on the optical flow image; calculate the geometric center coordinates of each outline and use the obtained coordinates as dynamic feature points; divide the current frame image into multiple masks through FastSAM, splice the masks containing dynamic feature points, and obtain the complete masks of all moving objects in the image.

[0041] The positioning module is used to locate the moving objects in the dynamic environment to be measured based on the complete mask of all moving objects.

[0042] The present invention also provides a computer device for positioning an object in a dynamic environment, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the method for positioning an object in a dynamic environment when executing the computer program.

[0043] The present invention further provides a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the object positioning method for a dynamic environment.

[0044] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for locating an object in a dynamic environment, characterized in that: The following steps are involved: Acquire a continuous image data set of a moving object in a dynamic environment to be tested; The FlowNet2 optical flow method is used to process two adjacent frames of image pairs in the continuous image data set of moving objects to generate an optical flow image of the image pair; the complete outline of each moving object in the image pair is obtained based on the optical flow image; the geometric center coordinates of each outline are calculated, and the obtained coordinates are used as dynamic feature points; The current frame image is divided into multiple masks through FastSAM, and the masks containing dynamic feature points are spliced ​​to obtain the complete masks of all moving objects in the image; The moving objects in the dynamic environment to be measured are located based on the complete masks of all moving objects.

2. The object positioning method for a dynamic environment according to claim 1, characterized in that: The method of processing two adjacent image pairs in a continuous image data set of a moving object by using the FlowNet2 optical flow method to generate an optical flow image of the image pair specifically includes the following steps: The optical flow data between two adjacent frames is obtained by FlowNet2 optical flow method [D x ,D y ]; where each pixel in the image is traversed, with the traversed pixel as the origin of the coordinate, the horizontal axis is the x-axis, the vertical axis is the y-axis, D x is the optical flow data in the x-axis direction, D y is the optical flow data in the y-axis direction; Will Determine the optical flow at the point (x, y) and generate an optical flow image for the image pair.

3. The object positioning method for a dynamic environment according to claim 2, characterized in that: The process also includes introducing Gaussian blur to process the optical flow image of the image pair, thereby generating an optical flow image after Gaussian blur; the processing process specifically includes the following steps: According to the two-dimensional Gaussian distribution, a weight matrix is ​​generated, the size of which is σ×σ; each element in the matrix represents the weight of the pixel at the corresponding position; The weight matrix is ​​convolved with the original image; specifically, the weight matrix is ​​covered on a certain area of ​​the optical flow image, and the weighted sum of the weight matrix and the pixel value of the covered area is calculated; The weighted sum is used as the pixel value of the corresponding position in the output image; The convolution operation is repeated by sliding the weight matrix, traversing the entire image and generating a Gaussian blurred optical flow image.

4. The object positioning method for a dynamic environment according to claim 3, characterized in that: The two-dimensional Gaussian distribution is expressed as: ， Here, σ is the standard deviation, which is the same in both the x-direction and the y-direction.

5. The object positioning method for a dynamic environment according to claim 1, characterized in that: Before positioning the moving objects in the dynamic environment to be measured according to the complete masks of all moving objects, the dynamic feature points covered by the complete masks of the dynamic objects are removed.

6. An object positioning system for a dynamic environment, characterized in that: include: An acquisition module is used to acquire a continuous image data set of a moving object in a dynamic environment to be tested; The calculation module is used to process two adjacent frames of image pairs in the continuous image data set of the moving object by using the FlowNet2 optical flow method to generate an optical flow image of the image pair; obtain the complete outline of each moving object in the image pair according to the optical flow image; calculate the geometric center coordinates of each outline, and use the obtained coordinates as dynamic feature points; The current frame image is divided into multiple masks through FastSAM, and the masks containing dynamic feature points are spliced ​​to obtain the complete masks of all moving objects in the image; The positioning module is used to locate the moving objects in the dynamic environment to be measured based on the complete mask of all moving objects.

7. A computer device for positioning objects in a dynamic environment, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein the processor implements the steps of the object positioning method for a dynamic environment as described in any one of claims 1 to 5 when executing the computer program.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the object positioning method for a dynamic environment as described in any one of claims 1 to 5.

Images