Lane line detection method with row-column anchor adaptive selection capability
By introducing the FSAM module and kurtosis feedback mechanism into the UFLDv2 network, adaptive selection of row and column anchors is achieved, which solves the problems of insufficient generalization ability and insufficient feature extraction caused by fixed anchor points, and improves the accuracy and adaptability of lane line detection.
Patent Information
- Application Number
- CN202411008530.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-12
AI Technical Summary
Existing lane line detection methods based on row classification rely on fixed anchor points, resulting in insufficient model generalization ability, difficulty in adapting to complex scenarios, and insufficient feature extraction capabilities, leading to low detection accuracy.
We introduce the FSAM module, which integrates frequency and spatial domain attention mechanisms, and achieve adaptive selection of row and column anchors through a kurtosis feedback mechanism to optimize the UFLDv2 network model. By combining Dice Loss and ordered classification loss function, we improve the model's generalization ability and feature extraction quality.
It improves the accuracy and generalization ability of lane line detection, reduces localization errors, and enhances the model's adaptability in complex scenarios.
Smart Images

Figure CN118942067A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of computer vision target detection based on deep learning, and relates to a lane line detection method based on UFLDv2 with row and column anchor adaptive selection capability. Background Art
[0002] Lane detection is an important function in advanced driver assistance systems and one of the key technologies for realizing autonomous driving. Lane detection provides positioning and navigation information for vehicles by identifying and tracking lane lines on the road, ensuring that vehicles drive in the correct lane, assisting cars to follow traffic rules and safe driving principles, and providing safety protection for car driving.
[0003] However, lane detection faces the following problems: traditional methods based on feature extraction are difficult to apply in complex scenarios (such as bad weather, occlusion, etc.), and the large number of parameters in the network model based on deep learning makes edge deployment difficult. In order to solve the problem that traditional methods are difficult to apply in complex scenarios, more and more researchers tend to use data-driven deep learning to solve it. In addition to using common operations such as pruning, quantization, and depthwise separable convolution to solve the problem of large number of network model parameters, some researchers have redefined the lane detection task and proposed a lane detection method based on row classification, which greatly reduces the parameters of the network model. Typical models include UFLD and UFLDv2.
[0004] However, the lane line detection method based on row classification currently still has the problem that the arrangement of anchor points is artificially fixed, which cannot fully meet the needs of actual application scenarios, that is, the generalization ability of the model is insufficient. In order to solve the problem of adaptive selection of row and column anchors, the present invention feeds back the peak information in the row anchor or column anchor to the network, so that the network model can adaptively select row and column anchors for specific lane lines, making the model more in line with the actual situation, while improving the accuracy and generalization ability of the model. Summary of the invention
[0005] The present invention provides a lane line detection method with row and column anchor adaptive selection capability, which improves the generalization capability of the original network while ensuring the lane line detection accuracy.
[0006] The present invention provides a lane line detection solution with row and column anchor adaptive selection capability, comprising the following steps:
[0007] (1) Preprocess the images containing lane lines to complete the dataset;
[0008] (2) The FSAM module that integrates the frequency domain attention mechanism and the spatial domain attention mechanism is introduced behind the Backbone of the UFLDv2 network model;
[0009] (3) Introducing a row-column anchor adaptive selection mechanism based on peak feedback to improve the generalization ability of UFLDv2;
[0010] (4) The obtained feature map is input into the MLP, and the lane line position coordinates are obtained through the existence branch and the positioning branch respectively;
[0011] (5) Substitute the obtained lane line position coordinates and the true value into the loss function that integrates Dice Loss, and perform back propagation after calculating the results.
[0012] Furthermore, the implementation process of step (1) is as follows:
[0013] The lane line images are cropped into a uniform size for lane line annotation, and the resulting images are divided into training and test sets in a ratio of 8:2.
[0014] Furthermore, the FSAM module introduced in step (2) that integrates the frequency domain attention mechanism and the spatial domain attention mechanism is mainly divided into two parts: a frequency domain module and a spatial domain module, wherein the implementation process of the frequency domain module is as follows:
[0015] The spatial features are converted to the frequency domain through two-dimensional discrete Fourier transform, and the formula is as follows:
[0016] Where: x[m,n] is a two-dimensional signal, which represents the two-dimensional coordinates of the pixel point on the image before Fourier transform, where 0≤m≤M-1,0≤n≤N-1; u in X[u,v] is the frequency component after row transform, and v is the frequency component after column transform;
[0017] Multiply the frequency domain features by the learnable filter K pixel by pixel:
[0018]
[0019] Where: K is a global filter with the same dimension as X, which is continuously learned during the back-propagation process;
[0020] The frequency domain features are mapped back to the spatial domain through the two-dimensional inverse Fourier transform. The formula is as follows:
[0021]
[0022] Furthermore, the implementation process of the spatial domain module in the FSAM module described in step (2) is as follows:
[0023] The input feature map is subjected to global maximum pooling and global average pooling, the channel size is compressed, the two results are channel-joined, and then a 7×7 convolution operation is performed on the concatenated result. After the Sigmoid activation function, the attention weight matrix is obtained:
[0024] M[m,n]=σ(f 7x7 ([AvgPool(m,n);MaxPool(m,n)]))
[0025] Where: M[m,n] is the spatial attention weight matrix, σ is the Sigmoid activation function, AvgPool and MaxPool are global average pooling and global maximum pooling respectively, f 7x7 It is a 7×7 convolution operation.
[0026] Furthermore, the implementation process of step (3) is as follows:
[0027] Assume that the feature map is divided by lane anchors before entering the multi-layer perceptron, and the probability distribution of the corresponding lane line is obtained after the positioning branch. The kurtosis of the probability distribution of each lane line is calculated separately:
[0028]
[0029] Where: E[(X-μ) 4 ] is the fourth-order center distance, E[(X-μ) 2 ] is the variance, Kurt(X) is the kurtosis of a row anchor, which reflects the discrete degree of the probability distribution of the existence of lane lines for each anchor point in the current row anchor.
[0030] After calculating the kurtosis for each row anchor, take the mean:
[0031]
[0032] Where: h is the number of row anchors divided artificially.
[0033] After calculating the kurtosis for each lane line, compare it with the specified prior threshold. If it does not reach the specified threshold, it means that the probability distribution of the located lane line is relatively scattered, and the row anchor detection effect is not good, so it should be changed to column anchor.
[0034] After determining that the anchor type needs to be changed for the specified lane line, return to the multi-layer perceptron, re-divide the anchor points, and re-perform branch prediction.
[0035] Furthermore, the implementation process of step (4) is as follows:
[0036] After passing through the multi-layer perceptron, two one-dimensional vectors are obtained. One is used as the input of the existence branch to predict whether each anchor has a lane line; the other is used as the input of the positioning branch to predict the probability of the existence of a lane line for each anchor point. After obtaining the probability distribution, the positioning branch is still a one-dimensional vector and needs to be reshaped to a size of h×w, where h is the number of rows or columns of anchors and w is the number of anchor points in an anchor. The output result of the positioning branch takes the position coordinates of the expected value of each anchor, and multiplies them with the corresponding results of the existence branch to obtain the final position information of each lane line.
[0037] Furthermore, the implementation process of step (5) is as follows:
[0038] The loss function design in UFLDv2 is different from that of traditional classification networks. Its loss function is called ordered classification loss because its classification network defines adjacent categories as closely ordered relationships. In order to better utilize the prior of the order relationship, the basic classification loss and the expected loss are used.
[0039] The basic classification loss is defined as follows:
[0040]
[0041] Where L CE (·) is the cross entropy loss, For the prediction of the row anchor assigned to the i-th lane and the j-th row anchor, for The corresponding classification label, Assignment of the mth lane to the row anchor and the prediction of the nth row anchor, for The corresponding classification label, onehot(·) is the one-hot encoding function.
[0042] The expected loss is defined as follows:
[0043]
[0044] where L1(·) is the smoothed L1 loss.
[0045] The existence branch loss is defined as follows:
[0046]
[0047] The cross entropy loss value only considers the loss in a microscopic sense, rather than the global one. The Dice Loss loss can solve this problem well. This loss can be used to measure the similarity between two samples, and its mathematical definition is:
[0048]
[0049] The above formula shows the dice coefficient equation, where p i and g i They represent the predicted pixel value and the true value, which are 0 or 1. In the lane detection scenario, it indicates whether the pixel is a boundary. If it is, the value is 1, and if it is not, the value is 0. The denominator is the total boundary pixel sum of the predicted value and the true value, and the numerator is the total boundary pixel sum of the correct prediction. i and g i The loss function increases only when the values (all 1) match.
[0050] The original UFLDv2 network uses a second-order differential equation to constrain the shape of lane lines, which works well for straight line detection, but not so well for lanes with large curvatures. For this reason, quadratic curves are used for shape constraints. The loss function after the curve constraint is defined as follows:
[0051]
[0052] Therefore, the final loss function is composed as follows:
[0053] L=L cls +αL exp +β(μL ext +γ(εL Dice +θL curve ))
[0054] Among them, μ+γ=1 and ε+θ=1.
[0055] Compared with the original network model UFLDv2, the present invention has the following beneficial effects:
[0056] (1) In order to solve the problem that the row and column anchor allocation in the original UFLDv2 network does not fully meet the actual needs (specifically: in the original UFLDv2 network, it is artificially stipulated that the inner two lanes use row anchors and the outer two lanes use column anchors, which does not meet the situation of lane changes during the actual driving process of the car, thereby amplifying the positioning error), the present invention provides a row and column anchor adaptive selection mechanism based on peak feedback, which improves the generalization ability of the network.
[0057] (2) Aiming at the problem that the original UFLDv2 network has insufficient feature extraction capability and poor performance in capturing local details, resulting in low accuracy, the present invention provides a FSAM module. This module enables the network to capture global and local information simultaneously by integrating the frequency domain attention mechanism and the spatial domain attention mechanism, thereby improving the feature extraction quality of the network.
[0058] (3) In order to solve the problem that the two fully connected layers in the original UFLDv2 network not only integrate global features but also extract and propagate local features, making both insufficient, the present invention designs a LocalBranch on the neck structure, which is specifically used for extracting and propagating local features. This branch can better capture the local information around the lane line, integrate it with the global features, and assist the fully connected layer in understanding the global context information, thereby improving the accuracy of detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 is a flow chart of the present invention;
[0060] Figure 2 This is the FSAM module structure diagram;
[0061] Figure 3 Figure 2 is a structural diagram of a network model with a kurtosis feedback mechanism. DETAILED DESCRIPTION
[0062] The present invention will be further described below in conjunction with the accompanying drawings.
[0063] The present invention provides a lane line detection method with row and column anchor adaptive selection capability. Aiming at the problem of insufficient local feature extraction capability in the original UFLDv2 network and the problem of amplified positioning error caused by artificially assigning different row and column anchors to different lane lines, the present invention introduces a FSAM module that integrates the frequency domain attention mechanism and the spatial domain attention mechanism on the basis of the original UFLDv2 network model. At the same time, a row and column anchor adaptive selection mechanism based on peak feedback is creatively provided to improve the accuracy and generalization capability of the original UFLDv2 network model. The overall flow chart is as follows: Figure 1 The specific steps are as follows:
[0064] Step 1: Preprocess the obtained images containing lane lines to complete the preparation of the data set.
[0065] This paper uses the public datasets CULane and TuSimple as examples. The former is a dataset of lane lines in different scenes in Beijing (including cities, rural areas, highways, etc.) (including normal, crowded, night, blocked, shadowed, etc.), and the latter is a dataset of lane lines on American highways. The two datasets have completed the annotation of lane lines and the division of training and test sets.
[0066] Step 2: Introduce the FSAM module that integrates the frequency domain attention mechanism and the spatial domain attention mechanism behind the Backbone of the UFLDv2 network model.
[0067] Lane lines have some inherent characteristics, such as lane lines usually occupy a small proportion of the entire image, the location area is fixed, the shape is slender and smooth, and the color is fixed. Therefore, frequency domain processing is a suitable choice for lane line detection. For the frequency domain attention mechanism, the global filtering module is the main one, which mainly includes three operations: 2D discrete Fourier transform to convert the input spatial features to the frequency domain, pixel-by-pixel product between the frequency domain features and the global filter, and 2D inverse Fourier transform to map the features back to the spatial domain:
[0068] (1) Convert spatial features to frequency domain through two-dimensional discrete Fourier transform:
[0069] Where: x[m,n] is a two-dimensional signal, representing the two-dimensional coordinates of the pixel point on the image before Fourier transform, where 0≤m≤M-1,0≤n≤N-1; u in X[u,v] is the frequency component after row transform, and v is the frequency component after column transform.
[0070] (2) Pixel-by-pixel product between frequency domain features and learnable filter K:
[0071]
[0072] Where: K is a global filter with the same dimension as X, which is continuously learned during the back-propagation process.
[0073] (3) Map the features back to the spatial domain through two-dimensional inverse Fourier transform:
[0074]
[0075] For the spatial domain attention mechanism, a general spatial attention structure is adopted. After the feature map is subjected to maximum pooling and flat pooling, two two-dimensional vectors are obtained. Then the two vectors are concatenated and convolution operations are performed to obtain two-dimensional spatial attention, which is then multiplied pixel by pixel with the original input feature map.
[0076] M[m,n]=σ(f 7x7 ([AvgPool(m,n);MaxPool(m,n)]))
[0077] Where: M[m,n] is the spatial attention weight matrix, σ is the Sigmoid activation function, AvgPool and MaxPool are global average pooling and global maximum pooling respectively, f 7x7 It is a 7×7 convolution operation.
[0078] The two attention modules are connected in series. On the basis of using the frequency domain attention mechanism to reduce redundant information, the spatial attention mechanism is used to make the lane position detection more accurate. The overall block diagram of the FSAM module is shown in the figure. Figure 2 shown.
[0079] Step 3: Introduce a row and column anchor adaptive selection mechanism based on kurtosis feedback to improve the generalization ability of UFLDv2.
[0080] The original UFLDv2 network artificially fixed the allocation of row anchors and column anchors: for the case of four lanes, the inner two lanes use row anchors and the outer two lanes use column anchors. However, such a choice does not always meet the needs of the actual situation. This is because in some cases, the vehicle may be located at the far right of the road, and there are three lanes on the left side of the vehicle. According to the working principle of UFLDv2, this fixed choice may lead to the problem of amplifying the positioning error. The root cause of this problem is that the angle between the lane line and the horizontal direction may affect the suitability of the anchor point selection. For example, when the vehicle is located at the far right of the road, the two leftmost lanes are actually more suitable for column anchors rather than row anchors. Taking the row anchor as an example, when the angle between the lane line and the horizontal direction is small (that is, the perspective angle of the lane line in the image is relatively horizontal), the grids predicted to have lane lines in the row anchor should be more and dispersed, that is, from the perspective of probability distribution, its peak state is relatively flat; conversely, when the angle between the lane line and the horizontal direction is large, the grids predicted to have lane lines in the row anchor will be more concentrated, that is, the peak state is more sharp. Therefore, the peak state of its probability distribution can be expressed by kurtosis. Kurtosis is also called kurtosis coefficient. It is a characteristic number that characterizes the height of the peak of the probability density distribution curve at the mean value.
[0081] Assume that the feature map is divided by lane anchors before entering the multi-layer perceptron, and the probability distribution of the corresponding lane line is obtained after the positioning branch. The kurtosis of the probability distribution of each lane line is calculated separately:
[0082]
[0083] Where: E[(X-μ) 4 ] is the fourth-order center distance, E[(X-μ) 2 ] is the variance, Kurt(X) is the kurtosis of a row anchor, which reflects the discrete degree of the probability distribution of the existence of lane lines for each anchor point in the current row anchor.
[0084] After calculating the kurtosis for each row anchor, take the mean:
[0085]
[0086] Where: h is the number of row anchors divided artificially.
[0087] After calculating the kurtosis for each lane line, compare it with the specified prior threshold. If it does not reach the specified threshold, it means that the probability distribution of the located lane line is relatively scattered, and the row anchor detection effect is not good, so it should be changed to column anchor.
[0088] After determining that the anchor type needs to be changed for the specified lane line, return to the multi-layer perceptron, re-divide the anchor points, and re-perform branch prediction.
[0089] Step 4: Input the obtained feature map into MLP, and obtain the lane line position coordinates through the existence branch and the positioning branch respectively.
[0090] After passing through the multi-layer perceptron, two one-dimensional vectors are obtained, one as the input of the existence branch to predict whether each anchor has a lane line; the other as the input of the positioning branch to predict the probability of the existence of a lane line for each anchor point. After obtaining the probability distribution, the positioning branch is still a one-dimensional vector and needs to be reshaped to a size of h×w, where h is the number of rows or columns of anchors and w is the number of anchor points in an anchor. The output result of the positioning branch takes the coordinates of the expected value of each anchor, and multiplies them with the corresponding results of the existence branch to obtain the predicted information of the position of each lane line.
[0091] Step 5: Substitute the obtained lane line position coordinates and the true value into the loss function that incorporates Dice Loss, and perform backpropagation after calculating the results.
[0092] The loss function design in UFLDv2 is different from that of traditional classification networks. Its loss function is called ordered classification loss because its classification network defines adjacent categories as closely ordered relationships. In order to better utilize the prior of the order relationship, the basic classification loss and the expected loss are used.
[0093] The basic classification loss is defined as follows:
[0094]
[0095] Where L CE (·) is the cross entropy loss, For the prediction of the row anchor assigned to the i-th lane and the j-th row anchor, for The corresponding classification label, Assignment of the mth lane to the row anchor and the prediction of the nth row anchor, for The corresponding classification label, onehot(·) is the one-hot encoding function. Since the categories are ordered, the expected prediction can be regarded as the average voting result. For convenience, the expectation is expressed as:
[0096]
[0097] Where [·] represents the index operator. Prob is defined as:
[0098]
[0099] In this way, the expected value of the prediction can be constrained to be closer to the true value. Therefore, the expected loss is defined as follows:
[0100]
[0101] where L1(·) is the smoothed L1 loss.
[0102] The existence branch loss is defined as follows:
[0103]
[0104] The cross entropy loss value only considers the loss in a microscopic sense, rather than the global one. The Dice Loss loss can solve this problem well. This loss can be used to measure the similarity between two samples, and its mathematical definition is:
[0105]
[0106] The above formula shows the dice coefficient equation, where p i and g i They represent the pixel prediction value and the true value, which are 0 or 1 respectively.
[0107] The original UFLDv2 network uses a second-order differential equation to constrain the shape of lane lines, which works well for straight line detection, but not so well for large bends. Therefore, a quadratic curve is used for shape constraints. The lower left corner of the image is used as the origin of the coordinate axis, the right direction of the bottom edge is the x-axis, and the upper left direction is the y-axis. Therefore, the location of the lane line can be expressed using a two-variable linear equation y=Ax 2 +Bx+c, and the derivative is:
[0108]
[0109] For adjacent points a and b on the lane line, The same is true for points b and c. Taking the second-order derivative, we get:
[0110]
[0111] Because for row anchors, the number of grids is considered to be evenly divided, that is, Δ 2 x a =Δ 2 x b , indicating that the quadratic difference of a quadratic curve at two adjacent points is equal. a 、x b For Loc i,j 、Loc i,j+1 , the loss function after adding the curve constraint is:
[0112]
[0113] Therefore, the final loss function is composed as follows:
[0114] L=L cls +αL exp +β(μL ext +γ(εL Dice +θL curve ))
[0115] Where α and β are hyperparameters, μ+γ=1, ε+θ=1.
Claims
1. A lane detection method with row and column anchor adaptive selection capability, characterized in that: The following steps are involved: (1) Preprocess the images containing lane lines to complete the preparation of the data set; (2) The FSAM module that integrates the frequency domain attention mechanism and the spatial domain attention mechanism is used behind the Backbone of the UFLDv2 network model; (3) Use a row-column anchor adaptive selection mechanism based on peak feedback to improve the generalization ability of UFLDv2; (4) The obtained feature map is input into the MLP, and the lane line position coordinates are obtained through the existence branch and the positioning branch respectively; (5) Substitute the obtained lane line position coordinates and the true value into the loss function that integrates Dice Loss, and perform back propagation after calculating the results.
2. The lane line detection method with row and column anchor adaptive selection capability according to claim 1, characterized in that: The implementation process of step (1) is as follows: The lane line images are cropped into a uniform size for lane line annotation, and the resulting images are divided into training and test sets in a ratio of 8:
2.
3. The lane line detection method with row and column anchor adaptive selection capability according to claim 1, characterized in that: The implementation process of step (2) introducing the FSAM module that integrates the frequency domain attention mechanism and the spatial domain attention mechanism into the backbone network of the UFLDv2 network model is as follows: (1) Convert spatial features to frequency domain through two-dimensional discrete Fourier transform: Where: x[m,n] is a two-dimensional signal, which represents the two-dimensional coordinates of the pixel point on the image before Fourier transform, where 0≤m≤M-1,0≤n≤N-1; u in X[u,v] is the frequency component after row transform, and v is the frequency component after column transform; (2) Pixel-by-pixel product between frequency domain features and learnable filter K: Where: K is a global filter with the same dimension as X, which is continuously learned during the back-propagation process; (3) Map the features back to the spatial domain through two-dimensional inverse Fourier transform: For the spatial domain attention mechanism, a general spatial attention structure is used. After the feature map is subjected to maximum pooling and flat pooling, two two-dimensional vectors are obtained. Then the two vectors are concatenated and convolved to obtain a two-dimensional spatial attention, which is then multiplied pixel by pixel with the original input feature map. M[m,n]=σ(f 7x7 ([AvgPool(m,n);MaxPool(m,n)])) Where: M[m,n] is the spatial attention weight matrix, σ is the Sigmoid activation function, AvgPool and MaxPool are global average pooling and global maximum pooling respectively, f 7x7 It is a 7×7 convolution operation.
4. The lane line detection method with row and column anchor adaptive selection capability according to claim 1, characterized in that: The implementation process of introducing the row and column anchor adaptive selection mechanism based on kurtosis feedback in step (3) is as follows: Assume that the feature map is divided by lane anchors before entering the multi-layer perceptron, and the probability distribution of the corresponding lane line is obtained after the positioning branch. The kurtosis of the probability distribution of each lane line is calculated separately: Where: E[(X-μ) 4 ] is the fourth-order center distance, E[(X-μ) 2 ] is the variance, Kurt(X) is the kurtosis of a row anchor, which reflects the discrete degree of the probability distribution of the existence of lane lines for each anchor point in the current row anchor; After calculating the kurtosis for each row anchor, take the mean: Where: h is the number of row anchors divided artificially; After calculating the kurtosis for each lane line, it is compared with the specified prior threshold; If the specified threshold is not reached, it means that the probability distribution of the located lane line is relatively scattered, and the row anchor detection effect is poor, so it should be changed to column anchor; After determining that the anchor type needs to be changed for the specified lane line, return to the multi-layer perceptron, re-divide the anchor points, and re-perform branch prediction.
5. The lane line detection method with row and column anchor adaptive selection capability according to claim 1, characterized in that: The implementation process of step (4) is as follows: After passing through the multi-layer perceptron, two one-dimensional vectors are obtained respectively. One is used as the input of the existence branch to predict whether there is a lane line for each anchor; the other is used as the input of the positioning branch to predict the probability of the existence of a lane line for each anchor point. The positioning branch is still a one-dimensional vector after obtaining the probability distribution and needs to be reshaped to a size of h×w, where h is the number of rows or columns of anchors and w is the number of anchor points in an anchor. The output result of the positioning branch takes the coordinates of the expected value of each anchor, and multiplies them with the corresponding results of the existence branch to obtain the predicted information of the position of each lane line.
6. The lane line detection method with row and column anchor adaptive selection capability according to claim 1, characterized in that: The implementation process of step (5) is as follows: The loss function design in UFLDv2 is different from that of the traditional classification network. Its loss function is called ordered classification loss because its classification network defines adjacent categories as closely ordered relationships. In order to better utilize the prior of the order relationship, the basic classification loss and expected loss are used. The basic classification loss is defined as follows: Where L CE (·) is the cross entropy loss, For the prediction of the row anchor assigned to the i-th lane and the j-th row anchor, for The corresponding classification label, Assignment of the mth lane to the row anchor and the prediction of the nth row anchor, for The corresponding classification label, onehot(·) is the one-hot encoding function; The expected loss is defined as follows: where L1(·) is the smoothed L1 loss; The existence branch loss is defined as follows: The cross entropy loss value only considers the loss in a microscopic sense, rather than the global one. The Dice Loss loss can solve this problem well. This loss can be used to measure the similarity between two samples, and its mathematical definition is: The above formula shows the dice coefficient equation, where p i and g i They represent the predicted pixel value and the true value, which are 0 or 1 respectively. In the lane detection scenario, it indicates whether the pixel is a boundary, which is 1 if it is and 0 if it is not. The denominator is the total boundary pixel sum of the predicted value and the true value, and the numerator is the total boundary pixel sum of the correct prediction. i and g i The loss function increases only when the values (all 1) match; The loss function after the curve constraint is defined as follows: Therefore, the final loss function is composed as follows: L=L cls +αL exp +β(μL ext +γ(εL Dice +θL curve )) Among them, μ+γ=1 and ε+θ=1.
Citation Information
Patent Citations
Improved wavelet threshold denoising-based fault diagnosis method and improved wavelet threshold denoising-based fault diagnosis system
CN110717472A
Low-illumination image enhancement method based on multi-scale stacked attention network
CN114972107A
Face counterfeit video detection method based on Fourier domain adaptation
CN116563957A
PIV speed measurement method and device based on Fourier feature enhancement and global matching
CN117825743A
Single-image rain removal model construction method based on frequency domain contrast regularization and rain removal method
CN118014892A
Cited By
Lane line detection and lane keeping method based on machine vision
CN119723493A
Lane line segmentation method and device, electronic equipment and storage medium
CN119942128A
Lane segmentation method, device, electronic device and storage medium
CN119942128B
