A Road Vanishing Point Detection Method Based on Pyramid Dual-Compressed Attention HRNet

Abstract

This invention discloses a road vanishing point detection method based on a pyramid dual-compressed attention HRNet, relating to the fields of computer vision and autonomous driving technology. The method first acquires an input image of an unstructured road, which is then processed through the first layer of an HRNet network to generate a feature map. Next, using an improved high-resolution HRNet network, the feature map is processed in each basic block through pyramid convolution and dual-pooling channel activation to generate channel attention vectors. These channel attention vectors are normalized and then element-wise multiplied with multi-scale feature maps to achieve weighted feature fusion, generating the final output feature map. Finally, the vanishing point coordinates of the road are predicted based on the output feature map of subsequent HRNet networks. This method utilizes multi-scale pyramid convolution and dual-pooling channel activation to achieve dynamic recalibration of spatial and channel dimension features, improving detection accuracy and speed, and is suitable for scenarios such as autonomous driving.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and autonomous driving technology, and in particular to a road vanishing point detection method based on pyramid dual-compressed attention HRNet. Background Technology

[0002] The vanishing point is the geometric point where parallel lines converge at infinity in perspective projection. It reflects the spatial structure of the road and the direction of vehicle travel and is a core element of autonomous driving, robot navigation, and 3D scene understanding. Vision-based road detection in unstructured environments (such as desert or rural roads) is extremely challenging, as these environments lack sufficient and distinctive fixed features to describe the road or its boundaries, making accurate and rapid vanishing point detection a persistent problem.

[0003] Traditional methods rely on manually designed features, such as region, edge, and texture features. Edge-based methods extract road contour features for geometric deduction, such as spline models, cascaded Hough transform frameworks, and B-snake models, leveraging the advantages of morphological modeling to adapt to various road structures. However, these methods are prone to failure in unstructured roads due to blurred lane boundaries and road surface interference.

[0004] Texture-based methods achieve general scene adaptation through feature extraction and voting mechanisms, with typical schemes including a combination of Gabor filtering and Hough voting. Subsequent research has built upon this foundation, using improved Gabor filters combined with soft voting to optimize detection accuracy. Further work has improved efficiency by reducing the number of filters, introducing gLoG filters to enhance feature representation, and improving Weber descriptors to strengthen texture direction information, among other methods, to enhance the robustness of the algorithms. Furthermore, the proposal of effective block extraction based on random forests and a block-weighted soft voting strategy has further improved the effectiveness of this type of algorithm in complex environments. Although these methods have some adaptability to both structured and unstructured roads, they suffer from significant bottlenecks: firstly, the quality of feature extraction is constrained by uneven lighting or image blurring, affecting detection accuracy; secondly, the voting mechanism has a high computational load, making it difficult to meet the demands of real-time scenarios such as autonomous driving.

[0005] Convolutional neural networks (CNNs), as the core architecture of deep learning, have achieved revolutionary breakthroughs in computer vision, particularly excelling in tasks such as image classification and object detection. To address the vanishing point detection problem in unstructured roads, a research team proposed an advanced algorithm using the high-resolution HRHet network as its backbone, combined with multi-scale supervised learning and heatmap super-resolution techniques. However, HRHet's basic blocks use fixed-size 3×3 convolutional kernels, resulting in a statically limited receptive field. This makes it difficult to fully extract multi-scale key information from unstructured roads, such as narrow sections, distant roads, or sudden curves, where feature sizes vary significantly. A fixed receptive field can easily lead to insufficient or lost features, affecting the accuracy of vanishing point detection. Furthermore, traditional convolutional operations assume equal importance for all channels when processing multi-channel features, failing to dynamically differentiate channel contributions based on scene content. In complex real-world environments, shadows, occlusions, reflections, and blurring often cause some channels to contain noisy or redundant features. Failure to effectively identify and suppress these low-quality channels weakens the model's discriminative ability. Therefore, the traditional HRNet structure suffers from limited receptive field and insufficient channel representation capabilities when dealing with unstructured road scenes. Summary of the Invention

[0006] The purpose of this invention is to propose a road vanishing point detection method based on Pyramid Dual Compression Attention HRNet. By designing a novel Pyramid Dual Compression Attention mechanism module (PDSA) to replace the 3×3 convolution kernel in the HRHet basic block, and using multi-scale pyramid convolution and dual pooling channel excitation, dynamic recalibration of spatial and channel dimension features is achieved.

[0007] To achieve the above objectives, this invention proposes a road vanishing point detection method based on pyramid dual-compressed attention HRNet, the specific steps of which are as follows:

[0008] Step S1: Obtain the input image of the unstructured road and generate a feature map through the first layer of HRNet network;

[0009] Step S2: Extract multi-scale feature maps based on the improved high-resolution network HRNet;

[0010] Step S3: Perform dual pooling activation processing on the multi-scale feature map to generate channel attention vectors;

[0011] Step S4: Normalize the channel attention vector and multiply it element-wise with the multi-scale feature map to achieve weighted feature fusion and generate the final output feature map.

[0012] Step S5: Predict the vanishing point coordinates of the road based on the output feature map of the subsequent HRNet network.

[0013] Preferably, in step S2, the improved HRNet uses the Pyramid Dual Compression Attention (PDSA) module to replace the 3×3 convolution kernel in the HRHet basic block. Through multi-scale pyramid convolution and dual pooling channel activation, it achieves dynamic recalibration of the importance of spatial and channel dimension features, including split pyramid convolution, dual pooling activation, and feature weighted fusion.

[0014] Preferably, multi-scale road features are extracted using an improved HRNet, with the following steps:

[0015] Step S21: Divide the input feature map into four groups of sub-feature maps evenly along the channel dimension;

[0016] Step S22: Perform depth-separable convolution operations on each group of sub-feature maps using convolution kernels of different sizes to generate multi-scale feature maps.

[0017] Preferably, in step S21, given the input feature map The feature maps are uniformly divided into four sub-groups, as shown in the following formula:

[0018]

[0019] Where X is the input feature map, X i Let i be a sub-feature map, i = 0, 1, 2, 3; C is the number of channels, G is the image height, and W is the image width.

[0020] Preferably, in step S22, for each group of sub-feature maps X i Using a depthwise convolution kernel with a size of k i ×k i Convolution is performed to generate feature maps of different scales, and then these feature maps are concatenated to form a multi-scale feature map. The formula is as follows:

[0021] k i =2(i+1)+1;

[0022]

[0023] Among them, F i This is a feature map with one-quarter channel, where Conv(·) is the convolution operation and Stack(·) is the concatenation operation.

[0024] Preferably, in step S3, for each Performing global average pooling and global max pooling operations respectively yields the following pooling results: The formula is as follows:

[0025]

[0026] in, This is the result of global average pooling. The result is the global max pooling function, AvgPool(·) is the global average pooling function, and MaxPool(·) is the global max pooling function.

[0027] The corresponding activation vector is obtained by using a multilayer perceptron (MLP) composed of convolution, activation functions, and convolution again, as shown in the following formula:

[0028]

[0029] in, This is the activation vector of the global average pooling result. This is the activation vector of the global max pooling result;

[0030] right and Summation is performed, and channel attention vectors are generated through a gating mechanism implemented using the sigmoid function. The formula is as follows:

[0031]

[0032] Where σ is the sigmoid function.

[0033] Preferably, in step S4, the channel attention vector is normalized using a softmax operation to obtain the recalibrated weights for each scale channel, and the normalized weights are then... With the corresponding multi-scale feature map F c Element-wise multiplication yields a refined output that incorporates multi-scale feature information, as shown in the following formula:

[0034]

[0035] Where Softmax(·) is the normalization function. These are the channel attention vectors for the first, second, third, and fourth groups, respectively, F. out This is for outputting feature maps.

[0036] Therefore, this invention proposes a road vanishing point detection method based on pyramid dual-compressed attention HRNet, which has the following advantages:

[0037] (1) Adaptive recalibration of channel importance: Through dual-path excitation of global average pooling and global max pooling, and the generation of complementary attention weights by multilayer perceptron, noise channels can be effectively identified and suppressed, and key feature channels can be enhanced. For scenes with disturbances such as uneven lighting, shadows and occlusion, dual pooling attention effectively suppresses noise, ensures the quality of feature extraction, and still performs stably in complex unstructured environments.

[0038] (2) Strong multi-scale dynamic perception capability: By utilizing the split pyramid deep convolution group, it can simultaneously extract multi-level features from local details to global context. Compared with the traditional fixed 3×3 convolution, PDSA has a wider receptive field and is more stable and accurate in detecting unstructured roads with significant scale changes.

[0039] (3) Excellent performance: The detection results of the method of the present invention are closest to the actual ground location and have higher accuracy in challenging unstructured environments.

[0040] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0041] Figure 1 This is a flowchart of a road vanishing point detection method based on pyramid dual-compressed attention HRNet according to the present invention;

[0042] Figure 2 This is a schematic diagram illustrating different stages of the Pyramid Dual Compression Attention Mechanism (PDSA) module in this invention; wherein, Figure 2 (a) Schematic diagram of the overall structure of the pyramid dual-compression attention mechanism module; Figure 2 (b) in the diagram is a schematic diagram of the structure of the segmented pyramid convolution stage; Figure 2 (c) in the diagram is a schematic diagram of the dual-pooling excitation stage;

[0043] Figure 3 This is a comparison diagram of the structure of the basic module and the PDSA module used in HRNet in this invention;

[0044] Figure 4 This is a comparison chart of the detection results of different detection methods in the embodiments of the present invention;

[0045] Figure 5 This is a performance comparison chart of different detection methods on the Kong dataset in this embodiment of the invention. Detailed Implementation

[0046] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0047] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0048] Example

[0049] like Figure 1 The diagram shows a flowchart of a road vanishing point detection method based on pyramid dual-compressed attention HRNet according to the present invention. The specific steps are as follows:

[0050] S1. Obtain the input image of the unstructured road and generate a feature map through the first layer of HRNet network;

[0051] S2. Extracting multi-scale feature maps based on the improved high-resolution network HRNet;

[0052] like Figure 2-3 As shown, the improved HRNet uses the Pyramid Dual Compression Attention (PDSA) module to replace the 3×3 convolution kernel in the HRHet basic block. Through multi-scale pyramid convolution and dual pooling channel activation, it achieves dynamic recalibration of the importance of spatial and channel dimension features, including split pyramid convolution, dual pooling activation, and feature weighted fusion.

[0053] Multi-scale road features are extracted using an improved HRNet, and the steps are as follows:

[0054] S21. Divide the input feature map into four groups of sub-feature maps evenly along the channel dimension, as shown in the following formula:

[0055]

[0056] Where X is the input feature map, X i Let i be a sub-feature map, i = 0, 1, 2, 3; C is the number of channels, H is the image height, and W is the image width.

[0057] S22. For each group of sub-feature maps X i Using a depthwise convolution kernel with a size of k i ×k i Convolution is performed to generate feature maps of different scales, and then these feature maps are concatenated to form a multi-scale feature map. The formula is as follows:

[0058] k i =2(i+1)+1;

[0059]

[0060] Among them, F i This is a feature map with one-quarter channel, where Conv(·) is the convolution operation and Stack(·) is the concatenation operation.

[0061] Step S3: Perform dual pooling activation processing on the multi-scale feature map to generate channel attention vectors, specifically:

[0062] For each Performing global average pooling and global max pooling operations respectively yields the following pooling results: The formula is as follows:

[0063]

[0064] in, This is the result of global average pooling. The result is the global max pooling function, AvgPool(·) is the global average pooling function, and MaxPool(·) is the global max pooling function.

[0065] The corresponding activation vector is obtained by using a multilayer perceptron (MLP) composed of convolution, activation functions, and convolution again, as shown in the following formula:

[0066]

[0067] in, This is the activation vector of the global average pooling result. This is the activation vector of the global max pooling result;

[0068] right and Summation is performed, and channel attention vectors are generated through a gating mechanism implemented using the sigmoid function. The formula is as follows:

[0069]

[0070] Where σ is the sigmoid function.

[0071] Step S4: Normalize the channel attention vectors and perform element-wise multiplication with the multi-scale feature maps to achieve dynamic recalibration, generating the final output feature map. Specifically, the channel attention vectors are normalized using a softmax operation to obtain the recalibration weights for each scale channel. The normalized weights are then... With the corresponding multi-scale feature map F c Element-wise multiplication yields a refined output that incorporates multi-scale feature information, as shown in the following formula:

[0072]

[0073]

[0074] Where Softmax(·) is the normalization function. These are the channel attention vectors for the first, second, third, and fourth groups, respectively, F. outThis is for outputting feature maps.

[0075] Step S5: Predict the vanishing point coordinates of the road based on the output feature map of the subsequent HRNet network.

[0076] The invention will be further illustrated by specific experiments below.

[0077] like Figure 4 As shown in the figure, the detection performance of this invention is compared with three other methods. Red markers represent ground-based virtual reality (VPs); blue markers represent the results of the URVP method; yellow and white markers correspond to the results of Kong (Gabor) and Kong (gLoG), respectively; and green markers represent the method proposed in this invention. The comparison shows that the green point of this invention is consistently closest to the red ground-based VP location, indicating that the method of this invention has higher accuracy in challenging unstructured environments.

[0078] like Figure 5 As shown, on the Kong test dataset, the normalized Euclidean distance (NormDist) error of this invention is less than 0.01 for 271 images and less than 0.03 for 630 images. The average normalized Euclidean distance (NormDist) is only 0.03182. The error rate is significantly lower than that of traditional HRNet and texture-based voting methods, and the accuracy is significantly improved.

[0079] The comparison results of the present invention with the other three methods in terms of average running speed and accuracy are shown in Table 1.

[0080] Table 1. Average running speed and accuracy of different algorithms

[0081] Methods Mean error Running time (fps) Kong (gLoG) 0.04644 0.13 Kong (Gabor) 0.04631 0.15 URVP 0.03487 55 Proposed 0.03182 65

[0082] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0083] Therefore, this invention provides a road vanishing point detection method based on pyramid dual-compressed attention HRNet, which adaptively calibrates channel importance through a dual-pooling activation mechanism and uses split pyramid convolution to achieve dynamic perception of multi-scale features; while suppressing noise interference in complex unstructured environments, it improves the accuracy and stability of vanishing point detection, and the increase in computational burden is not significant.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A pyramid double compression attention HRNet-based road vanishing point detection method, characterized in that, The specific steps are as follows: Step S1, obtaining an input image of an unstructured road, generating a feature map through a first layer network of HRNet; Step S2, extracting a multi-scale feature map based on the improved high-resolution network HRNet; Step S3, performing double-pooling excitation processing on the multi-scale feature map to generate a channel attention vector; Step S4, normalizing the channel attention vector and element-wise multiplying it with the multi-scale feature map to realize feature weighted fusion and generate a final output feature map; Step S5, predicting the vanishing point coordinates of the road according to the output feature map of the subsequent network of HRNet; In step S2, the improved HRNet uses a pyramid double compression attention mechanism module PDSA to replace the 3x3 convolution kernel in the HRHet basic block, and through multi-scale pyramid convolution and double-pooling channel excitation, it realizes dynamic recalibration of spatial and channel dimension features, including split pyramid convolution, double-pooling excitation and feature weighted fusion; Through the improved HRNet, the multi-scale road features are extracted, and the steps are as follows: Step S21, uniformly dividing the input feature map into four groups of sub-feature maps along the channel dimension; Step S22, respectively performing depth separable convolution operation on each group of sub-feature maps using convolution kernels of different sizes to generate multi-scale feature maps.

2. The road vanishing point detection method based on pyramid double compression attention HRNet according to claim 1, characterized in that, In step S21, the input feature map is given The input feature map is uniformly divided into four groups of sub-feature maps, and the formula is as follows: ； wherein, is an input feature map, is a sub-feature map, i =0, 1, 2, 3; is a number of channels, is an image height, is an image width.

3. The road vanishing point detection method based on pyramid double compression attention HRNet according to claim 1, characterized in that, In step S22, for each group of sub-feature maps , a deep convolution kernel with a size of is used to generate feature maps of different scales, and then the feature maps of different scales are spliced to form multi-scale feature maps , and the formula is as follows: ； ； ； wherein, is a feature map of a quarter of a channel, is a convolution operation, is a concatenation operation.

4. The road vanishing point detection method based on pyramid double compression attention HRNet according to claim 1, characterized in that, In step S3, for each The global average pooling and global maximum pooling operations are performed respectively, and the pooling results are The formula is as follows: ； wherein, is a global average pooling result, is a global max pooling result, is a global average pooling function, is a global max pooling function; Through a multi-layer perceptron MLP composed of convolution, activation function and convolution, the corresponding excitation vector is obtained, and the formula is as follows: ； wherein, is an activation vector for the global average pooling result, is an activation vector for the global max pooling result; Summing up and generates a channel attention vector through the gating mechanism realized by the sigmoid function, as follows: ； wherein, is a sigmoid function.

5. The pyramid dual compression attention HRNet-based road vanishing point detection method according to claim 1, characterized in that, In step S4, the channel attention vector is normalized to obtain recalibration weights of each scale channel, and the normalized weights are multiplied by the corresponding multi-scale feature maps softmax to obtain a refined output fused with multi-scale feature information, as shown in the following formula: ​​ ； ； wherein, is a normalization function, , , , are the first group, the second group, the third group and the fourth group of channel attention vectors, respectively, is the output feature map.

Images