A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection

Abstract

The present invention relates to the technical field of computer vision and deep learning, and provides a SLAM dynamic disturbance suppression method based on blur processing and target detection, including: obtaining a blur remover; using the blur remover to perform blur processing on an image and extracting ORB feature points; obtaining a dynamic target detector; detecting a clear image through the dynamic target detector to obtain a target detection frame; classifying the motion state of the object in the clear image according to the target detection frame and the ORB feature points, and obtaining a static feature point set using an overlapping frame strategy and epipolar geometry constraints; and performing system pose estimation and trajectory prediction according to the static feature point set. The present invention reduces the adverse effects of dynamic targets on other parts of the system through deblurring, and the target detection function can accurately identify dynamic targets. Combined with the overlapping frame strategy and epipolar geometry constraints, the dynamic feature points can be fully eliminated, so that the overall performance of the system is well improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and deep learning, and in particular to a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection. Background Art

[0002] Simultaneous Localization and Mapping (SLAM) refers to the process by which mobile devices such as robots determine their own position and posture through sensors and complete the construction of maps of the surrounding environment. Visual SLAM uses cameras as the main sensor. The rich image information gives it super scene recognition capabilities, making it one of the key technologies for solving the navigation problem of unmanned systems in unknown environments. Its application areas have covered multiple dimensions such as sea, land, and air.

[0003] Most visual SLAM methods rely on the assumption of a static environment, which poses a huge challenge to the applicability of unmanned systems in real-world environments. The real environment is often more complex. On the one hand, dynamic objects will cause motion blur, which will deteriorate the quality of the collected images and make it difficult for the system to identify dynamic targets. On the other hand, dynamic objects will interfere with the system, affecting the accurate acquisition of positioning information and even causing the system to crash. Therefore, suppressing the adverse disturbances of dynamic objects is crucial to improving the robustness, accuracy and applicability of visual SLAM systems.

[0004] Existing dynamic visual SLAM methods can be divided into classical methods and intelligent methods. Classic methods include geometric information based on camera motion models and other sensor information. Dynamic visual SLAM methods based on geometric information have low reliability and are not suitable for environments with complex dynamic types; methods based on sensor information will increase the difficulty of system information fusion. The intelligent method is to integrate deep learning into the visual SLAM architecture, including semantic segmentation and target detection. The model of the method based on semantic segmentation is large and it is difficult to meet the real-time requirements of the system. The method based on target detection cannot determine the actual motion state of the object, which will cause a large number of static feature points to be wasted. Dynamic targets in real environments bring huge challenges to visual SLAM. The dynamic objects themselves and the fuzzy errors caused by motion seriously affect the feature extraction quality and positioning accuracy of the system. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection, which reduces the adverse effects of dynamic targets on other parts of the system through defuzzification, and the target detection function can accurately identify dynamic targets. Combined with the overlapping frame strategy and epipolar geometry constraints, the dynamic feature points are completely eliminated, so that the overall performance of the system is well improved.

[0006] The present invention provides a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection, comprising:

[0007] S1: Improve the DeblurGANv2 model through the MobileNetV2 module to obtain the blur remover;

[0008] S2: Use the blur remover to blur the image, obtain a clear image, and extract the ORB (ORiented Brief, referred to as ORB) feature points of the clear image;

[0009] S3: Improve the YOLOv5 model through GSConv and RepNCSPELAN4 to obtain a dynamic target detector; detect the clear image through the dynamic target detector to obtain the target detection frame;

[0010] S4: classifying the motion state of the object in the clear image according to the target detection frame and the ORB feature points to obtain a first dynamic feature point set, a first static feature point set, and a first uncertainty feature point set;

[0011] S5: Using the overlapping frame strategy to screen the first dynamic feature point set and the first uncertainty feature point set to obtain a second dynamic feature point set, a second static feature point set, and a second uncertainty feature point set;

[0012] S6: Using the epipolar geometry constraint to screen the second dynamic feature point set and the second uncertainty feature point set to obtain a third static feature point set;

[0013] S7: performing system pose estimation and trajectory prediction according to the first static feature point set, the second static feature point set and the third static feature point set.

[0014] Furthermore, the DeblurGANv2 model includes a generator and a discriminator. The generator gradually restores image details through a feature extraction module, a convolutional layer, and a residual block module to generate a clear image. The discriminator uses adversarial learning to improve the deblurring effect of the generator, extracts image features through a MobileNetV2 module, and obtains a blur remover.

[0015] Furthermore, in step S2, blur removal is used to blur the image, and obtaining a clear image includes:

[0016] The image clarity is quantitatively evaluated using the Laplace gradient function to obtain the image clarity of each image in the image sequence;

[0017] The calculation expression of image clarity is:

[0018]

[0019] in, Pixel The convolution of the Laplacian operator, For images clarity;

[0020] Calculate the clarity mean of the image sequence, compare the image clarity of each image with the clarity mean, and obtain a blurred image set;

[0021] The blur remover is used to blur the images in the blur image set to obtain a clear image.

[0022] Furthermore, in step S2, extracting ORB feature points of a clear image includes:

[0023] Construct an image pyramid and calculate the number of feature points assigned to each layer of the image. The calculation expression for the number of feature points assigned to each layer of the image is:

[0024]

[0025] in, For the The number of feature points assigned to the layer image, is the total number of feature points, is the scaling factor, is the number of pyramid layers;

[0026] Divide each layer of the pyramid image into grids and perform FAST feature point extraction. The number of feature points extracted from each layer of the image is ,

[0027] Based on the quadtree idea Get evenly distributed ORB feature points from feature points.

[0028] Furthermore, the YOLOv5 model includes a backbone network, a neck network and a head network. The backbone network, the neck network and the head network all contain convolutional modules and residual modules. The residual module is replaced by RepNCSPELAN4, and the convolutional module of the neck network is replaced by GSConv. The effective channel attention mechanism is added to the backbone network to obtain a dynamic target detector.

[0029] Furthermore, in step S4, the motion state of the object includes: a deterministic dynamic object, a deterministic static object, and an uncertain object; the ORB feature points in the deterministic dynamic object detected by the target detection frame are placed into a first dynamic feature point set, the ORB feature points in the uncertain dynamic object are placed into a first uncertain feature point set, and the ORB feature points in the deterministic static object are placed into a first static feature point set.

[0030] Furthermore, in step S5, the target detection frame is a rectangular frame, and there are overlaps between dynamic objects and static objects.

[0031] If the ORB feature point only exists in the first dynamic feature point set, it is put into the second dynamic feature point set;

[0032] If the ORB feature point exists in both the first dynamic feature point set and the first uncertainty feature point set, it is placed in the second uncertainty feature point set.

[0033] If the ORB feature point is only in the first uncertain feature point set, the ORB feature point is not affected by the dynamic object and is placed in the second static feature point set.

[0034] Furthermore, in step S6, the epipolar geometry constraint screens the second dynamic feature point set and the second uncertain feature point set by determining the distance from the feature point to the epipolar line.

[0035] The calculation expression of the distance between ORB feature points and epipolar lines is:

[0036]

[0037] in, is the distance between the ORB feature point and the epipolar line, is the magnitude of the epipolar direction vector, is the basic matrix, is the homogeneous coordinate of the first feature point in the pixel coordinate system, is the homogeneous coordinate of the second feature point in the pixel coordinate system, is transposed;

[0038] The ORB feature points in the second dynamic feature point set and the second uncertain feature point set are judged using the epipolar geometry constraint.

[0039] like , put the ORB feature points into the third static feature point set;

[0040] like , put the ORB feature points into the third dynamic feature point set.

[0041] Furthermore, in step S7, the feature points in the first static feature point set, the second static feature point set, and the third static feature point set are passed to the back-end optimization thread as stable features in the environment. The back-end optimization thread uses the correspondence between the feature points and the known feature points in the map to minimize the error through the optimization algorithm to obtain an accurate system pose estimation. The continuous pose estimation results constitute the motion trajectory of the system.

[0042] Furthermore, the optimization algorithm includes nonlinear least squares method and bundle adjustment, and the nonlinear least squares method includes Gauss-Newton method and Levenberg-Marquardt method.

[0043] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0044] The present invention integrates the fuzzy processing function into the front end of visual SLAM, reduces the noise caused by the jitter of moving objects or unmanned platforms, and enables the system to extract high-quality features in dynamic fuzzy scenes and accurately identify dynamic targets; the present invention improves the dynamic target detection network, while enhancing the ability to recognize dynamic objects, reducing memory usage, making the system more advantageous in mobile devices or under resource-constrained conditions; the present invention uses an overlapping box strategy in the dynamic feature point removal thread, reduces the workload of the system for judging epipolar geometric constraints, improves the running speed of the algorithm, and retains useful static feature points to the greatest extent, which is helpful for the system's posture estimation and positioning.

[0045] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a flow chart of a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection provided by the present invention.

[0048] Figure 2 It is a schematic diagram of feature point classification of a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection provided by the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme in the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0050] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0051] Combine the following Figure 1 to Figure 2 The present invention describes a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection.

[0052] like Figure 1 As shown, a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection includes:

[0053] A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection, comprising:

[0054] S1: Improve the DeblurGANv2 model through the MobileNetV2 module to obtain the blur remover;

[0055] The DeblurGANv2 model includes a generator and a discriminator. The generator gradually restores image details and generates a clear image through a feature extraction module, a convolutional layer, and a residual block module. The discriminator uses adversarial learning to improve the deblurring effect of the generator. The improved DeblurGANv2 model extracts image features through the MobileNetV2 module.

[0056] The GOPRO (sports camera) dataset is a series of continuous clear video images taken in different scenes and accumulated high-speed video frames to approximate the motion blur process in real life. In order to achieve the effect of real-time deblurring, the DeblurGANv2 model with fast operation speed and high ease of use is selected, and the Backbone part of the DeblurGANv2 model is replaced with MobileNetV2 to obtain a blur remover, which makes the model lighter while maintaining performance and improves the overall running speed. The blur remover receives the blurred image, extracts features through MobileNetV2, and then uses the feature pyramid network to perform feature fusion and enhancement, and finally restores a clear image through a series of upsampling and convolution operations. In addition, this embodiment uses the GOPRO dataset to train the improved DeblurGANv2 model, and the optimizer is replaced with RMSProp in network training, which can simplify parameter adjustment, improve the training speed of the model, and thus help improve performance indicators.

[0057] S2: Use the blur remover to blur the image, obtain a clear image, and extract the ORB feature points of the clear image;

[0058] The trained blur remover is used to blur the image. According to the blur degree, the image can be divided into a clear image, a low-degree blur image and a high-degree blur image. On the one hand, there will always be a clear image in the image sequence that does not need to be blurred; on the other hand, the adverse effect of a high-degree blur image on the system is more significant, and if all blur images are blurred in the visual SLAM system, the system's operating speed will be reduced. In this embodiment, a high-degree blur image is deblurred and the blur remover is used to blur the image. Obtaining a clear image includes:

[0059] The image clarity is quantitatively evaluated using the Laplace gradient function to obtain the image clarity of each image in the image sequence;

[0060] The calculation expression of image clarity is:

[0061]

[0062] in, Pixel The convolution of the Laplacian operator, For images clarity;

[0063] Calculate the clarity mean of the image sequence, compare the image clarity of each image with the clarity mean, and obtain a blurred image set;

[0064] Using a blur remover to blur the images in the blur image set to obtain a clear image;

[0065] The image is blurred using a blur remover to improve the accuracy of subsequent feature point extraction and the recognition rate of dynamic objects in the system, so as to compensate for the errors caused by moving objects to the image and reduce the subsequent system's missed detection rate of dynamic objects.

[0066] The ORB feature points for extracting clear images include:

[0067] Construct an image pyramid and calculate the number of feature points assigned to each layer of the image. The calculation expression for the number of feature points assigned to each layer of the image is:

[0068]

[0069] in, For the The number of feature points assigned to the layer image, is the total number of feature points, is the scaling factor, is the number of pyramid layers;

[0070] Divide each layer of the pyramid image into grids and perform FAST feature point extraction. The number of feature points extracted from each layer of the image is , based on the quadtree idea Get evenly distributed ORB feature points from feature points.

[0071] By dividing each layer of the pyramid image into grids, the extraction speed of FAST feature points is accelerated;

[0072] After feature point extraction is completed, feature points tend to cluster and distribute in places with rich textures, which brings redundant information to the system. Therefore, the image is divided into nodes based on the quadtree idea, and the feature point with the largest response value in each node is retained to increase the uniformity of feature extraction.

[0073] S3: Improve the YOLOv5 model through lightweight convolution (GSConv) and RepNCSPELAN4 to obtain a dynamic target detector; detect the clear image through the dynamic target detector to obtain the target detection frame;

[0074] The YOLOv5 model includes a backbone network (Backbone), a neck network (Neck) and a head network (Neck). When an image is input, the Backbone extracts features and generates a feature layer; the Neck performs multi-scale feature fusion on the feature layer; the Head performs classification and regression prediction on the feature layer and outputs the location and category information of the target. The backbone network, the neck network and the head network all contain convolutional modules (Conv) and residual modules (C3). Through the coordinated cooperation of each structural layer, efficient recognition of the target is achieved.

[0075] In order to further improve the recognition accuracy of YOLOv5 and make the detection network more lightweight, RepNCSPELAN4 is used instead of C3, GSConv is used instead of the ordinary Conv in the Neck part, and the Effective Channel Attention mechanism (ECA) is added to the Backbone part; a dynamic target detector is obtained.

[0076] Taking into account the real-world environment, people are dynamic objects that are likely to affect posture estimation. Therefore, people are mainly used as detection targets. When people move, they are far or near relative to the camera, that is, the targets are large or small relative to the viewing angle. RepNCSPELAN4 helps to better aggregate multi-scale features, enhance the dynamic target detector's ability to recognize targets of different sizes, and improve detection accuracy. GSConv is used instead of the original ordinary Conv structure in the Neck part. It effectively reduces the computational complexity of the model through the combination of grouped convolution and spatial convolution. At the same time, GSConv located in the Neck can enhance the feature fusion capability of the model, which is of great significance to improving detection accuracy. The lightweight ECA is added to the Backbone part of the network, which helps the model extract the key features of the target and improve computational efficiency.

[0077] S4: classifying the motion state of the object in the clear image according to the target detection frame and the ORB feature points to obtain a first dynamic feature point set, a first static feature point set, and a first uncertainty feature point set;

[0078] The motion state of an object includes deterministic dynamic objects, deterministic static objects, and uncertain objects;

[0079] Deterministic dynamic objects refer to objects that often move, such as people and animals;

[0080] Deterministic static objects refer to objects that are always stationary under normal circumstances, such as a clock on the wall;

[0081] Uncertain objects; refers to objects that may move, such as books, computers, and other objects that move due to external forces;

[0082] The ORB feature points in the deterministic dynamic objects detected by the target detection box are put into the first dynamic feature point set, the feature points in the uncertain dynamic objects are put into the first uncertain feature point set, and the feature points in the deterministic static objects are put into the first static feature point set.

[0083] S5: Using the overlapping frame strategy to screen the first dynamic feature point set and the first uncertainty feature point set to obtain a second dynamic feature point set, a second static feature point set, and a second uncertainty feature point set;

[0084] The target detection box is a rectangular box, and there are cases where dynamic objects and static objects overlap.

[0085] If the ORB feature point only exists in the first dynamic feature point set, it is put into the second dynamic feature point set;

[0086] If the ORB feature point exists in both the first dynamic feature point set and the first uncertainty feature point set, it is placed in the second uncertainty feature point set.

[0087] If the ORB feature point is only in the first uncertain feature point set, the ORB feature point is not affected by the dynamic object and is placed in the second static feature point set.

[0088] S6: Use the epipolar geometry constraint to screen the second dynamic feature point set and the second uncertainty feature point set to obtain a third static feature point set.

[0089] The epipolar geometry constraint screens the second dynamic feature point set and the second uncertainty feature point set by judging the distance from the feature point to the epipolar line.

[0090] If it is a static feature point, the static feature point is on the polar line, and the distance between the static feature point and the polar line is 0;

[0091] If it is a dynamic feature point, the distance between the dynamic feature point and the polar line is greater than 0;

[0092] The pyramid-based Lucas-Kanade optical flow method is used to calculate the matching feature points of the first image and the second image. The calculation expression is:

[0093]

[0094]

[0095] in, is the homogeneous coordinate of the first feature point in the pixel coordinate system, is the homogeneous coordinate of the second feature point in the pixel coordinate system, is the pixel coordinate of the first matched feature point of the first image, The pixel coordinates of the second feature point matched to the second image;

[0096] The calculation expression of the polar line is:

[0097]

[0098] in, is the polar line, is the basic matrix;

[0099] The epipolar constraints are:

[0100]

[0101]

[0102] in, is transposed;

[0103] The calculation expression of the distance between the feature point and the epipolar line is:

[0104]

[0105] in, is the distance between the dynamic feature point and the epipolar line, is the magnitude of the epipolar direction vector, is the basic matrix, is the homogeneous coordinate of the first feature point in the pixel coordinate system, is the homogeneous coordinate of the second feature point in the pixel coordinate system, is transposed;

[0106] The feature points in the second dynamic feature point set and the second uncertain feature point set are judged using the epipolar geometry constraint.

[0107] like , put the feature points into the third static feature point set;

[0108] like , put the feature points into the third dynamic feature point set.

[0109] S7: performing system pose estimation and trajectory prediction according to the first static feature point set, the second static feature point set and the third static feature point set;

[0110] The feature points in the first static feature point set, the second static feature point set, and the third static feature point set are passed to the backend optimization thread as stable features in the environment. The backend optimization thread uses the correspondence between the feature points and the known feature points in the map to minimize the error through the optimization algorithm to obtain an accurate camera pose estimation. The continuous pose estimation results constitute the motion trajectory of the camera.

[0111] The optimization algorithms include nonlinear least squares method and bundle adjustment. The nonlinear least squares method includes Gauss-Newton method and Levenberg-Marquardt method.

[0112] Simulation experiment:

[0113] The simulation environment for the SLAM dynamic disturbance suppression method experiment based on fuzzy processing and target detection is: GPUGeForce RTX 4070, CPU Intel i9, Ubuntu 20.04.

[0114] Blur remover simulation experiment:

[0115] The present invention selects peak signal-to-noise ratio (PSNR), structural similarity (SSIM), model size and average time as evaluation indicators to evaluate the effect of the blur remover. PSNR and SSIM are used to measure the quality of the deblurred image. After training, the performance of the optimized blur remover and the original version on the GOPRO dataset is shown in Table 1.

[0116] Table 1 Performance of the optimized blur remover and the original version on the GOPRO dataset

[0117]

[0118] Compared with the original deblurring network A, the network B after replacing the backbone has a reduced model size of 13M, and the PSNR and SSIM are significantly reduced, but the processing speed is faster. The network C after replacing the backbone and optimizer has a reduced model size of 13M and a faster processing speed. While sacrificing a small part of the accuracy, the running speed is improved and the space resources required for the hardware device are reduced.

[0119] Dynamic target detector simulation experiment:

[0120] Ablation experiments are conducted on the VOC dataset to verify the accuracy of each module improved on the object detection network. Table 2 shows the results of the ablation experiment.

[0121] Table 2 Ablation experiment results

[0122]

[0123] Among them, Rep in Table 2 represents the RepNCSPELAN4 module, mAp represents the average accuracy of target detection when IOU (intersection-over-union ratio) = 0.5, FPS is the number of images that can be processed per second, P is the precision rate, R is the recall rate, "×" represents that the target detector does not contain this module, and "√" represents that the target detector contains this module. It can be seen from Table 2 that the network parameter amount (Params) of the improved target detector of the present invention is reduced by 23.9%, the computing power (GFLOPS) requirement is reduced by 25.8%, the model size (Size) is reduced by 21.4%, the detection accuracy (mAp) is improved by 2.1%, and the detection speed (FPS) is improved by about 1 frame.

[0124] The impact of blur remover on object detection:

[0125] In order to verify the impact of the blur remover on the missed detection rate of the target detection network, the experiment counted the number of images containing dynamic objects in the four high-dynamic sequences of the TUM dataset, and used a dynamic object detector to detect the sequence images before and after deblurring. The results are shown in Table 3.

[0126] Table 3 Detection results of dynamic object detector on sequence images before and after deblurring

[0127]

[0128] Among them, in the two sequences w / static and w / xyz, in addition to the movement of dynamic targets, the camera's movement amplitude is also small, and the image blur is also light, so even before deblurring, the missed detection rate of the target detection network is also very small. For w / half and w / rpy, the complex movement trajectory of the camera aggravates the blur of the image. At this time, deblurring before target detection can more effectively improve the missed detection problem.

[0129] Overall performance evaluation and comparative experiments:

[0130] The TUM dataset was used for performance testing, and the root mean square error (RMSE), mean, and standard deviation (SD) were used to perform quantitative analysis of the absolute trajectory error (ATE). Table 4 shows the ATE results of the proposed method and the original algorithm ORB-SLAM2.

[0131] Table 4 ATE results of the proposed method and the original algorithm ORB-SLAM2

[0132]

[0133] Among them, in the five dynamic sequences in Table 4, compared with ORB-SLAM2, the present invention has an average improvement of about 82% in RMSE. Specifically, ORB-SLAM2 itself does not have a dynamic feature point removal thread, so the system trajectory error is large in high dynamic sequences. The present invention adds a deblurring function and a target detection function on the basis of ORB-SLAM2. Deblurring reduces the adverse effects of dynamic targets on other parts of the system. The target detection function can accurately identify dynamic targets. Combined with the overlapping frame strategy and epipolar geometry constraints, the comprehensive removal of dynamic feature points is achieved, so that the overall performance of the system is well improved.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection, characterized in that: include: S1: Improve the DeblurGANv2 model through the MobileNetV2 module to obtain the blur remover; S2: Use the blur remover to blur the image, obtain a clear image, and extract the ORB feature points of the clear image; S3: Improve the YOLOv5 model through GSConv and RepNCSPELAN4 to obtain a dynamic target detector; detect the clear image through the dynamic target detector to obtain the target detection frame; S4: classifying the motion state of the object in the clear image according to the target detection frame and the ORB feature points to obtain a first dynamic feature point set, a first static feature point set, and a first uncertainty feature point set; S5: Using the overlapping frame strategy to screen the first dynamic feature point set and the first uncertainty feature point set to obtain a second dynamic feature point set, a second static feature point set, and a second uncertainty feature point set; S6: Using the epipolar geometry constraint to screen the second dynamic feature point set and the second uncertainty feature point set to obtain a third static feature point set; S7: performing system pose estimation and trajectory prediction according to the first static feature point set, the second static feature point set and the third static feature point set.

2. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, characterized in that: The DeblurGANv2 model includes a generator and a discriminator. The generator gradually restores image details through a feature extraction module, a convolutional layer, and a residual block module to generate a clear image. The discriminator uses adversarial learning to improve the deblurring effect of the generator and extracts image features through a MobileNetV2 module to obtain a blur remover.

3. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, characterized in that: In step S2, the image is blurred using a blur remover to obtain a clear image including: The image clarity is quantitatively evaluated using the Laplace gradient function to obtain the image clarity of each image in the image sequence; The calculation expression of image clarity is: in, Pixel The convolution of the Laplacian operator, For images clarity; Calculate the clarity mean of the image sequence, compare the image clarity of each image with the clarity mean, and obtain a blurred image set; The blur remover is used to blur the images in the blur image set to obtain a clear image.

4. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, characterized in that: In step S2, the ORB feature points of the clear image are extracted including: Construct an image pyramid and calculate the number of feature points assigned to each layer of the image. The calculation expression for the number of feature points assigned to each layer of the image is: in, For the The number of feature points assigned to the layer image, is the total number of feature points, is the scaling factor, is the number of pyramid layers; Divide each layer of the pyramid image into grids and perform FAST feature point extraction. The number of feature points extracted from each layer of the image is , Based on the quadtree idea Get evenly distributed ORB feature points from feature points.

5. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, characterized in that: The YOLOv5 model includes a backbone network, a neck network and a head network. The backbone network, the neck network and the head network all contain convolutional modules and residual modules. The residual module is replaced by RepNCSPELAN4, and the convolutional module of the neck network is replaced by GSConv. The effective channel attention mechanism is added to the backbone network to obtain a dynamic target detector.

6. According to the SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, in step S4, the motion state of the object includes: Deterministic dynamic objects, deterministic static objects and uncertain objects; putting the ORB feature points in the deterministic dynamic objects detected by the target detection box into the first dynamic feature point set, putting the ORB feature points in the uncertain dynamic objects into the first uncertain feature point set, and putting the ORB feature points in the deterministic static objects into the first static feature point set.

7. According to a SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, in step S5, the target detection frame is a rectangular frame, and there is a situation where the dynamic object and the static object overlap, If the ORB feature point only exists in the first dynamic feature point set, it is put into the second dynamic feature point set; If the ORB feature point exists in both the first dynamic feature point set and the first uncertainty feature point set, it is placed in the second uncertainty feature point set. If the ORB feature point is only in the first uncertain feature point set, the ORB feature point is not affected by the dynamic object and is placed in the second static feature point set.

8. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, characterized in that: In step S6, the epipolar geometry constraint screens the second dynamic feature point set and the second uncertainty feature point set by determining the distance from the feature point to the epipolar line. The calculation expression of the distance between ORB feature points and epipolar lines is: in, is the distance between the ORB feature point and the epipolar line, is the magnitude of the epipolar direction vector, is the basic matrix, is the homogeneous coordinate of the first feature point in the pixel coordinate system, is the homogeneous coordinate of the second feature point in the pixel coordinate system, is transposed; The ORB feature points in the second dynamic feature point set and the second uncertain feature point set are judged using the epipolar geometry constraint. like , put the ORB feature points into the third static feature point set; like , put the ORB feature points into the third dynamic feature point set.

9. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 1, characterized in that: In step S7, the feature points in the first static feature point set, the second static feature point set, and the third static feature point set are passed to the back-end optimization thread as stable features in the environment. The back-end optimization thread uses the correspondence between the feature points and the known feature points in the map to minimize the error through the optimization algorithm to obtain an accurate system pose estimation. The continuous pose estimation results constitute the motion trajectory of the system.

10. A SLAM dynamic disturbance suppression method based on fuzzy processing and target detection according to claim 9, characterized in that: The optimization algorithms include nonlinear least squares method and bundle adjustment. The nonlinear least squares method includes Gauss-Newton method and Levenberg-Marquardt method.

Images