Track returning navigation method based on rut line recognition of deep learning

Abstract

The application discloses a kind of based on rut line identification of deep learning line navigation method of patrolling, with the center point of rut line and the sine and cosine value of rut line direction angle to represent rut line;According to the coordinate value of the default coordinate system of image, the endpoint of rut line is labeled;Depth learning identification model is constructed, including coding-decoding module and prediction module, three branches respectively predict the gaussian distribution kernel of rut line center point, the offset of rut line center point relative to true value and the direction of rut line center point;Model finds the area with the characteristics of gaussian kernel distribution from the feature map of rut line center point prediction branch, and selects its center;According to threshold, the pixel value corresponding to the rut line center point of candidate is screened, and the center point of the rut line preliminarily predicted is obtained, and after clustering, it guides the autonomous navigation of unmanned vehicle.The application can realize end-to-end identification of rut line center coordinate and direction angle, has higher efficiency, and is convenient for completing unmanned vehicle inspection return navigation.

Description

Technical Field

[0001] This invention relates to the field of unmanned vehicle patrol return navigation, and in particular to an unmanned vehicle patrol return navigation method based on deep learning image rut line recognition. Background Technology

[0002] Relative positioning information in the field plays a crucial role in enabling automatic navigation, positioning, and remote operation of unmanned exploration vehicles during exploration. The desert surface is covered by a loose layer of sandy loam, primarily composed of loose, powdery fine particles. This layer varies in thickness across almost the entire desert surface, averaging 4–5 meters thick, and reaching 10–20 meters in higher areas. It is irregularly distributed with pits, rocks, and slopes of different sizes and shapes, resulting in poor traversability. Vehicle wheels easily become stuck in the sand, causing severe slippage, and ordinary wheeled odometers become severely ineffective in the desert environment. Therefore, a method is urgently needed to ensure accurate relative positioning of vehicles traveling on the desert surface. Visual SLAM point features are the most widely used, and feature extraction and matching are relatively mature. However, single-point features are highly dependent on the environment and perform poorly in scenarios with missing textures. In contrast, the "tracks" left by vehicles on the desert surface provide valuable linear features. Therefore, combining point and linear features can achieve better feature complementarity. By mounting a binocular camera at the rear of a desert surface vehicle, the vehicle's tracks can be detected, allowing for relatively accurate calculation of the vehicle's mileage through visual detection, thus enabling relative positioning of the vehicle.

[0003] After completing the inspection task, the unmanned vehicle can also return along the tracks. However, due to the complex actual distribution of tracks, in some areas, track information is not accurately mapped into navigation maps, making accurate navigation difficult using GPS information. It's also challenging to guarantee that the camera can capture clear images of the tracks and their surroundings. Furthermore, GPS signal strength is weak in some areas, and signal loss is sometimes possible. Therefore, GPS navigation alone is not very reliable. Vision-based track-following navigation, on the other hand, relies on accurate track identification, making it more reliable and accurate than GPS navigation.

[0004] Therefore, identifying tire tracks is an essential task in image-based autonomous navigation for unmanned vehicles (UAVs). If an UAV can identify the location and orientation of tire tracks, it can adjust its direction and position in real time based on this information to maintain a path consistent with the tire tracks. Numerous studies have been conducted on tire track identification in UAV inspections. For example, Burns et al. (1986) proposed a method for identifying tire tracks using edge detection and pixel clustering; Akinlar et al. (2011) proposed an edge contour detection method called EDlines; and Ceron et al. (2014) proposed a tire track detection model for UAVs based on Circle Search (CBS). Based on CBS, Ceron et al. (2018) proposed a tire track identification method based on Histogram of Line Segments (HOS). However, these methods are not stable and reliable enough when dealing with UAV images taken against complex backgrounds. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the purpose of this invention is to provide a deep learning-based rut line recognition method for track-following return navigation, which can achieve end-to-end identification of the center coordinates and direction angle of rut lines, has high efficiency, and facilitates unmanned vehicle inspection and return navigation.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0007] A line-following return navigation method based on deep learning-based wheel rut recognition includes:

[0008] Step S1: Mathematically abstract the wheel tracks from the top view of the unmanned vehicle, and use the center point of the wheel tracks and the sine and cosine values ​​of the direction angle of the wheel tracks to represent the wheel tracks.

[0009] Step S2: For the images in the training and test sets, label the coordinates of the endpoints of each rut line according to the default coordinate system of the images;

[0010] Step S3: Construct a deep learning recognition model. The model structure includes an encoder-decoder module and a prediction module, which contains three fully convolutional network branches to predict the Gaussian distribution kernel of the rut center point, the offset of the rut center point relative to the true value, and the direction of the rut center point, respectively.

[0011] Step S4: The model predicts the branch from the center point of the rut line, finds the region with Gaussian kernel distribution characteristics on the feature map, and selects its center;

[0012] For multiple Gaussian distribution kernels on the feature map, the model ranks them according to the value of their center points and selects the top N points as candidate rut center points.

[0013] The model filters the pixel values ​​corresponding to the candidate rut center points and uses those values ​​greater than or equal to the threshold as the center points of the ruts initially predicted by the model.

[0014] The model uses the coordinates of the center point of the rut line predicted in the preliminary calculation to find the offset of the corresponding coordinate value on the prediction map generated by the corresponding rut line center offset prediction branch, and then corrects the offset to obtain the final predicted value of the rut line center point coordinates.

[0015] Optionally, in step S1, the wheel track in the top view of the unmanned vehicle is approximately a straight line. The average of its endpoints is taken to obtain the approximate center point of the wheel track. The wheel track is represented by the center point of the wheel track and the sine and cosine values ​​of the direction angle of the wheel track, which serve as the navigation basis for the unmanned vehicle to travel along the wheel track.

[0016] Optionally, in step S2, the dataset may be augmented using one or more of the following data augmentation methods based on the endpoints of the rut lines: translation in any direction, scaling down by any ratio, rotation at any angle, and shear deformation.

[0017] Optionally, the model adopts a compact single-step target recognition model structure based on an anchorless frame, which is used to directly predict the center point and direction of the wheel track, while canceling the recognition of the object target box.

[0018] Optionally, in step S3, the input of the model is a color image of size 512×512×3, which generates a feature map of size 128×128×5 after passing through the encoding-decoding module.

[0019] The feature map is divided into three branches. The first prediction branch consists of a single-channel feature map, which generates a Gaussian distribution kernel for fitting the center of the object. A heat map consistent with the true value distribution is generated through the first prediction branch. The center of the ruts on the heat map has a Gaussian kernel distribution.

[0020] The second prediction branch of the model, which is 128×128×1, is used to predict the offset between the center coordinates of the Gaussian kernel in the heat map and the actual rut center.

[0021] The model's third prediction branch is used to predict the sine and cosine values ​​of the rut direction angle; the sine and cosine values ​​range from [-1, 1].

[0022] Optionally, the parameters of the model can be optimized using an appropriate loss function;

[0023] Specifically, the focal loss function used for classification tasks is employed for the first prediction branch; the L1 loss function is employed for the second and third prediction branches; and the final loss function of the model is a weighted sum of the loss functions of the three prediction branches.

[0024] Optionally, in step S4, the third prediction branch is also used to directly predict the sine and cosine values ​​of the direction angle of each rut line, and corresponds to the center point of the rut line predicted by the other two branches.

[0025] Optionally, in step S4, for multiple Gaussian distribution kernels on the feature map, the model ranks them according to the value of their center points and selects up to 100 points as candidate rut center points.

[0026] Optionally, step S5 is further included after step S4, which clusters the center points and directions of all rut lines predicted by the model; wherein, the minimum bounding rectangle of the center points of all rut lines predicted by the model is calculated, and a Gaussian kernel is constructed based on the center position of the minimum bounding rectangle to describe the weights corresponding to the predicted rut lines. The Gaussian kernel is used to strengthen the influence of rut lines located in the central region of the rut line bundle on the final decision, while also taking into account the position and direction of other rut lines.

[0027] This invention primarily proposes a method for rut line recognition in autonomous vehicles (RVs) performing rut line following and return navigation. Since RVs operate in complex environments, automatic rut line recognition and following is a challenging task. To address this issue, this invention proposes a novel method based on deep convolutional neural networks. Unlike traditional methods, this method can automatically extract rut line features from complex backgrounds and effectively extract rut line features, predicting the center and direction of each rut line in real time. Based on this, this invention further proposes a clustering method for the center position and orientation of rut lines, aggregating the positions and orientations of all rut lines into a total position and orientation, which is used to guide the RV's rut ​​line following and return navigation.

[0028] This invention has at least one of the following technical effects:

[0029] (1) The method proposed in this invention can achieve end-to-end identification of the center coordinates and direction angle of the rut line, with high efficiency. Compared with traditional straight line detection and edge detection, the proposed method for identifying the center point and direction of the rut line has higher accuracy and robustness.

[0030] (2) When designing a deep learning-based model for identifying the center point and direction of ruts, the sine and cosine values ​​of the direction angle of the ruts were innovatively used as the output of the deep learning model and activated by the tanh function. This extended the deep learning-based target detection method from the identification of object boxes to the identification of the center point and direction of ruts, enabling the deep learning-based target detection model to realize the new function of detecting the center point and direction of ruts.

[0031] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0032] Figure 1 The center point and direction diagram of the wheel tracks in a top view provided in an embodiment of the present invention;

[0033] Figure 2 This is an arbitrary horizontal displacement transformation diagram provided in an embodiment of the present invention;

[0034] Figure 3 This is a scaled-down diagram provided in an embodiment of the present invention;

[0035] Figure 4 This is an arbitrary angle rotation diagram provided in an embodiment of the present invention;

[0036] Figure 5 This is an arbitrary shear deformation diagram provided in one embodiment of the present invention;

[0037] Figure 6 A model structure diagram for identifying the center point and direction of wheel tracks provided in an embodiment of the present invention;

[0038] Figure 7 A Gaussian distribution diagram of the wheel track weights provided in an embodiment of the present invention. Detailed Implementation

[0039] The following describes this embodiment in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0040] The present invention provides a line-following return navigation method based on deep learning rut line recognition, which includes the following steps 1 to 5.

[0041] Step 1: During the autonomous vehicle's line-following journey, the wheel tracks are usually in the vehicle's top-down view. Therefore, for guiding the autonomous vehicle's navigation task, the wheel tracks can be approximated as straight lines. Based on the endpoints of the straight lines, the approximate center point and direction of the wheel tracks can be obtained.

[0042] like Figure 1 As shown, P 31 (x 31 ,y 31 ) and P 32 (x 32 ,y 32 ) are the two endpoints of a rut line, and the rut line can be approximated by a line segment formed by these two points; assuming that C3 is the midpoint of the approximate line segment and θ is the direction angle of this line, then the coordinates and direction of the center point of this rut ​​line can be expressed by equations (1), (2), (3) and (4):

[0043]

[0044]

[0045]

[0046]

[0047] That is, the average of the endpoints of the wheel tracks in the top view of the unmanned vehicle is taken to obtain the approximate center point of the wheel tracks. The center point of the wheel tracks and the sine and cosine values ​​of the direction angle of the wheel tracks are used to represent the wheel tracks, which serve as the navigation basis for the unmanned vehicle to travel along the wheel tracks.

[0048] Step 2: For the images in the training and test sets, the coordinate values ​​of the endpoints of each rut line were labeled according to the default coordinate system of the images and saved as a JSON file in a format similar to MS COCO.

[0049] In this example, 2818 photographs were obtained from the video published in Ceron et al.'s paper as research data, of which 2297 and 521 images were randomly selected as the training and test sets, respectively. Due to the small dataset size, this invention designs a series of data augmentation methods based on the endpoints of wheel tracks to fully utilize existing data resources, reduce overfitting during deep learning model training, and improve the model's accuracy and robustness.

[0050] 1. Translation in any direction

[0051] The translation of a wheel track in any direction can be decomposed into translation in any horizontal direction and translation in any vertical direction. The final translation result is the linear superposition of the two translations. According to the knowledge of homogeneous coordinate transformation, when point P(x,y) is translated to P'(x',y') in any direction, its corresponding coordinate change can be determined by equation (5):

[0052]

[0053] For any horizontal translation, the value is 0; such as Figure 2 As shown in (a), a schematic diagram of an untransformed wheel rut with dimensions (w×h) is presented, where the two endpoints and the center point of the wheel rut are P1(x1,y1), P2(x2,y2), and C(x1,y1), respectively. c ,y c When the wheel tracks are translated, the two endpoints remain within the horizontal boundaries of the image, as shown below. Figure 2 As shown in (b), the transformed wheel track endpoints P'1(x'1,y'1) and P'2(x'2,y'2) can be obtained from equations (6), (7), (8) and (9), respectively.

[0054] x1'=x1+δx (6)

[0055] x'2=x2+δx (7)

[0056] y1'=y1 (8)

[0057] y'2=y2 (9)

[0058] Therefore, the coordinates of the center point of the transformed line can be calculated using equations (10) and (11):

[0059]

[0060]

[0061] 2. Reduced to any scale

[0062] Because the autonomous vehicle's camera altitude varies, the scale of the tire tracks in the acquired images will also differ significantly. To generate data at different scales, this invention employs a data augmentation method with arbitrary scaling.

[0063] Figure 3 (a) is the original image. When the image is scaled by a scaling factor s (s∈[0,1]), the scaled rut lines are as follows: Figure 3As shown in (b), after scaling, both ends of the rut lines are located within the boundaries of the image. w and h represent the length and width of the original image, and δh and δw are the vertical and horizontal distances between the border of the scaled-down image and the border of the original image, respectively. Specifically, they can be calculated using equations (12) and (13):

[0064]

[0065]

[0066] Because the coordinates of the endpoints of the transformed wheel ruts in the original image coordinate system can be calculated using equations (14), (15), (16), and (17):

[0067]

[0068]

[0069]

[0070]

[0071] 3. Rotate at any angle

[0072] Since the autonomous vehicle has a yaw angle relative to the direction of the tracks during the photography process, and the value of the yaw angle may fluctuate within a certain range, this invention introduces an arbitrary angle rotation transformation to reduce the fluctuation of model prediction accuracy caused by changes in the yaw angle. Figure 4 It demonstrates the process of rotating and transforming wheel tracks at any angle.

[0073] Generally speaking, when a point P(x,y) on an image is rotated by an angle δθ, the corresponding point is P'(x',y'). This correspondence can be expressed by equation (18):

[0074]

[0075] Figure 4 (a) is the original image. Figure 4 (b) shows the rotated wheel ruts. Figure 4 (c) shows the wheel tracks whose dimensions have changed after rotation; since the size of the figure has changed after rotation, when the rotated figure and the original figure are aligned to the same coordinate system, the center of the rotated figure is displaced relative to the original figure, and the displacement values ​​δx and δy can be calculated using equations (19) and (20):

[0076]

[0077]

[0078] Therefore, the endpoint coordinates of the wheel track after rotation transformation can be calculated according to equations (21), (22), (23) and (24):

[0079]

[0080]

[0081]

[0082]

[0083] 4. Arbitrary shear deformation

[0084] Shearing deformation is a commonly used data augmentation method in image recognition, mainly used to improve the robustness of the model to different viewpoints. Figure 5 This demonstrates a typical shear deformation process.

[0085] Figure 5 (a) is the original image. Figure 5 (b) is the image after shearing deformation. For a point on the image, the expression for its shearing transformation is shown in equation (25), where m is the shearing coefficient:

[0086]

[0087] Therefore, the coordinate values ​​of the end points of the rut line after shear deformation can be obtained according to equations (26), (27), (28) and (29):

[0088]

[0089]

[0090] y1'=y1 (28)

[0091] y'2=y2 (29)

[0092] Step 3: For autonomous navigation along rut lines, this invention proposes a deep learning-based model for identifying the center point and direction of rut lines. This model adopts a compact, single-step target recognition model structure similar to CenterNet, based on an anchorless bounding box architecture. Unlike general target detection models, this model can directly predict the center point and direction of rut lines, while eliminating the need for object bounding box recognition. The model structure mainly includes an encoder-decoder module, which extracts image features and generates the feature map required by the prediction module; and a prediction module, whose main structure includes three fully convolutional network branches, which predict the Gaussian kernel of the rut line center point, the offset of the rut line center point relative to the true value, and the direction of the rut line center point, respectively. Figure 6The structure of the wheel track center point and direction recognition model proposed in this invention is shown.

[0093] The model's input is an RGB image of size (512×512×3), which, after passing through an encoder-decoder module, generates a feature map of size (128×128×3). This feature map is divided into three branches. The first branch consists of a single-channel feature map, primarily generating a Gaussian kernel for fitting the object's center. Therefore, the pixel values ​​of this feature map are projected into a numerical space with a range of [0,1] using the sigmoid function. Assuming Y is the aforementioned Gaussian kernel, These are the coordinates of the center point of this Gaussian kernel, σ. p If is the standard deviation related to the size of the object's bounding box, then any point (x, y) in this Gaussian kernel can be determined by equation (30):

[0094]

[0095] Therefore, the task of the first branch of the model is to generate a heatmap that is consistent with the distribution of the true values, where there is a Gaussian kernel distribution at the center of the ruts, while the values ​​in the parts without ruts are all 0.

[0096] Since the heatmap output by the prediction module is 128×128×1 in size, which is 1 / 4 the size of the original image, an upsampling error will occur when the center coordinates of the corresponding Gaussian kernel on the heatmap are restored to the original image. To accurately predict the location of the rut center, this model has a second prediction branch, which is 128×128×2 in size. Its main function is to predict the offset between the center coordinates of the Gaussian kernel in the heatmap and the actual rut center.

[0097] Simultaneously, this model also has a third prediction branch, mainly used to predict the sine and cosine values ​​of the rut direction angle. Since the range of sine and cosine values ​​is between [-1, 1], this invention uses the tanh function to project the value of each pixel on the feature map output by the third branch of the model into the [-1, 1] region.

[0098] In training the rut center point and direction recognition model, this invention uses a corresponding loss function to optimize the model parameters, thereby obtaining a high-precision rut recognition model. For the heatmap branch output by the model, this invention uses the focal loss function used for classification tasks, as shown in equation (31):

[0099]

[0100] Where Y xy It is the true value, and These are the model's predicted values. α and β are hyperparameters in the focol loss, and the same settings as CornerNet are used, with α and β set to 2 and 4 respectively. For the offset of the rut center and the rut direction angle prediction branch, this invention uses the L1 loss function, and the specific calculation formulas are shown in equations (32) and (33):

[0101]

[0102]

[0103] Among them o x and o y It is the actual center point of the rut line. and

[0104] These are the coordinates of the center point of the rut predicted by the model; It is the true direction angle of the rut line, and This is the direction angle of the wheel track predicted by the model.

[0105] Therefore, the final loss function of the model is the weighted sum of the loss functions of the three prediction branches, as shown in equation (34):

[0106]

[0107] Where λ and γ are the weights of the rut center point offset loss function and the rut direction loss function, respectively. In this invention, the value of λ is set to 0.1 and the value of γ is set to 1.

[0108] Step 4: Prediction of the center point and direction angle of the wheel track

[0109] Since the model used in this invention has three branches, when predicting the center point and orientation angle of the rut line, the model first searches for regions with Gaussian kernel distribution characteristics on the feature map of the rut line center point prediction branch and selects their centers. For multiple Gaussian kernels on the feature map, the model ranks them according to the value of their center points and selects up to 100 points as candidate rut line center points; at the same time, the model filters the pixel values ​​corresponding to the candidate rut line center points, selecting those greater than or equal to a threshold T. det The point is used as the center point of the rut line initially predicted by the model. In this invention, T det The value is set to 0.33 based on commonly used empirical values ​​for object detection models.

[0110] Meanwhile, the model finds the offset of the corresponding coordinate value on the prediction map generated by the corresponding rut center offset prediction branch based on the coordinates of the center point of the rut line in the preliminary prediction, and obtains the final predicted value of the coordinates of the center point of the rut line after offset correction.

[0111] In addition to predicting the center point of the rut lines, the rut line direction prediction branch proposed in this invention can also directly predict the sine and cosine values ​​of the direction angle of each rut line, and correspond them to the center points of the rut lines predicted by the other two branches. Therefore, the model proposed in this invention can directly predict all rut line center points and their corresponding rut line directions.

[0112] Since the rutline tracking and navigation task requires a general overview of the overall position and orientation of the rutline bundle, the center points and orientations of all rutlines predicted by the model still need to be clustered to guide the autonomous vehicle's navigation. This invention addresses the characteristics of rutline recognition in autonomous vehicle rutline tracking by proposing a customized clustering method for rutline center points and orientations. First, the minimum bounding rectangle is calculated for the center points of all rutlines predicted by the model. Then, based on the center position of the minimum bounding rectangle, a Gaussian kernel is constructed to describe the weights corresponding to the predicted rutlines. If (x... i ,y i Let (x) be the coordinates of the center point of the i-th rut line. c ,y c Let be the center coordinates of the Gaussian kernel. Then, the weight value w corresponding to the i-th rut line can be calculated according to equation (35):

[0113]

[0114] The distribution of wheel track weights is as follows Figure 7 As shown, this Gaussian kernel is designed to enhance the influence of the rut lines located in the central region of the rut bundle on the final decision, while also taking into account the position and orientation of other rut lines.

[0115] It should be noted that in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0116] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for line-following return navigation based on deep learning-based wheel track recognition, characterized in that, include: Step S1: Mathematically abstract the wheel tracks from the top view of the unmanned vehicle, and use the center point of the wheel tracks and the sine and cosine values ​​of the direction angle of the wheel tracks to represent the wheel tracks. Step S2: For the images in the training and test sets, label the coordinates of the endpoints of each rut line according to the default coordinate system of the images; Step S3: Construct a deep learning recognition model. The model structure includes an encoder-decoder module and a prediction module, which contains three fully convolutional network branches to predict the Gaussian distribution kernel of the rut center point, the offset of the rut center point relative to the true value, and the direction of the rut center point, respectively. Step S4: The model predicts the branch from the center point of the rut line, finds the region with Gaussian kernel distribution characteristics on the feature map, and selects its center; For multiple Gaussian distribution kernels on the feature map, the model ranks them according to the value of their center points and selects the top N points as candidate rut center points. The model filters the pixel values ​​corresponding to the candidate rut center points and uses those values ​​greater than or equal to the threshold as the center points of the ruts initially predicted by the model. The model uses the coordinates of the center point of the rut line predicted in the preliminary calculation to find the offset of the corresponding coordinate value on the prediction map generated by the corresponding rut line center offset prediction branch, and then corrects the offset to obtain the final predicted value of the rut line center point coordinates.

2. The line-following return navigation method based on deep learning wheel track recognition as described in claim 1, characterized in that, In step S1, the wheel track in the top view of the unmanned vehicle is approximately a straight line. The average of its endpoints is taken to obtain the approximate center point of the wheel track. The wheel track is represented by the center point of the wheel track and the sine and cosine values ​​of the direction angle of the wheel track, which serve as the navigation basis for the unmanned vehicle to travel along the wheel track.

3. The line-following return navigation method based on deep learning wheel track recognition as described in claim 1, characterized in that, In step S2, the dataset is subjected to one or more of the following data augmentation methods based on the endpoints of the rut lines: translation in any direction, scaling down by any ratio, rotation at any angle, and shear deformation.

4. The line-following return navigation method based on deep learning wheel track recognition as described in claim 1, characterized in that, The model adopts a compact single-step target recognition model structure based on an anchorless frame, which is used to directly predict the center point and direction of the wheel track, while canceling the recognition of the object target box. In step S3, the input of the model is a color image of size 512×512×3, which generates a feature map of size 128×128×5 after passing through the encoding-decoding module. The feature map is divided into three branches. The first prediction branch consists of a single-channel feature map, which generates a Gaussian distribution kernel for fitting the center of the object. A heat map consistent with the true value distribution is generated through the first prediction branch. The center of the ruts on the heat map has a Gaussian kernel distribution. The second prediction branch of the model, which is 128×128×1, is used to predict the offset between the center coordinates of the Gaussian kernel in the heat map and the actual rut center. The model's third prediction branch is used to predict the sine and cosine values ​​of the rut direction angle; the sine and cosine values ​​range from [-1, 1].

5. The line-following return navigation method based on deep learning wheel track recognition as described in claim 4, characterized in that, The model parameters are optimized using appropriate loss functions. For the first prediction branch, the focal loss function used for classification tasks is adopted; for the second and third prediction branches, the L1 loss function is adopted; and the final loss function of the model is the weighted sum of the loss functions of the three prediction branches.

6. The line-following return navigation method based on deep learning wheel track recognition as described in claim 4, characterized in that, In step S4, the third prediction branch is also used to directly predict the sine and cosine values ​​of the direction angle of each rut line, and corresponds to the center point of the rut line predicted by the other two branches.

7. The line-following return navigation method based on deep learning wheel track recognition as described in claim 1, characterized in that, In step S4, for multiple Gaussian distribution kernels on the feature map, the model ranks them according to the value of their center points and selects the top 100 points as candidate rut center points.

8. The line-following return navigation method based on deep learning rut line recognition as described in claim 1, characterized in that, Step S4 is followed by step S5, which clusters the center points and directions of all rut lines predicted by the model. Specifically, the minimum bounding rectangle of the center points of all rut lines predicted by the model is calculated. Based on the center position of the minimum bounding rectangle, a Gaussian kernel is constructed to describe the weights corresponding to the predicted rut lines. This Gaussian kernel is used to strengthen the influence of rut lines located in the central region of the rut line bundle on the final decision, while also taking into account the position and direction of other rut lines.

Images