Cross-scale information guide multi-task detection method based on improved YOLOv11

Abstract

The invention relates to a cross-scale information guide multi-task detection method based on improved YOLOv11, and belongs to the field of computer vision. The method comprises the steps of obtaining a multi-task data set including traffic target detection, drivable area segmentation and lane line detection, and performing preprocessing; yOLOv11 is adopted as a backbone network, an SPPF module is introduced at the tail of the network, and meanwhile, a target detection and semantic segmentation structure of the YOLOv11 is improved; a multi-scale context sensing up-sampling module is used to adaptively capture local and global context information through dynamic learning of space offset, and a differentiable bilinear interpolation is adopted to realize feature map amplification; inputting the processed data for network training; according to the detection performance of the model, evaluation is carried out in a real scene, an optimal model is stored, and tasks of traffic target detection, drivable area segmentation and lane line detection of multi-task panoramic driving perception are completed.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a cross-scale information guided multi-task detection method based on an improved YOLOv11. Background Art

[0002] In recent years, multi-task learning for autonomous driving has garnered widespread attention in both academia and industry within the field of artificial intelligence, significantly impacting areas such as intelligent transportation, assisted driving systems, and autonomous driving technology. Multi-task learning for autonomous driving aims to improve the overall performance of perception systems by simultaneously completing core tasks such as traffic object detection, drivable area segmentation, and lane detection using a unified deep learning model. Compared to single-task models, multi-task learning optimizes resource utilization, improves computational efficiency, and enhances synergy between different tasks by leveraging shared features. This enables autonomous driving systems to generalize more effectively in complex traffic environments.

[0003] Benefiting from the powerful perception capabilities of deep neural networks, multi-task learning frameworks have made significant progress in autonomous driving perception tasks in recent years. Multi-task architectures based on convolutional neural networks (CNNs) and visual transformers (ViTs) enable models to simultaneously learn global and local features, improving the accuracy of detection, segmentation, and lane prediction. Furthermore, the introduction of methods such as feature fusion and attention mechanisms effectively alleviates conflicts between tasks and improves the overall performance of multi-task perception models. Therefore, designing effective multi-task learning frameworks, optimizing feature sharing and allocation mechanisms between tasks, and properly balancing the loss functions of each task play a key role. Summary of the Invention

[0004] The purpose of the present invention is to provide a cross-scale information-guided multi-task detection method based on an improved YOLOv11, which can accurately and effectively perform multi-task detection under autonomous driving.

[0005] To achieve the above objectives, the technical solution of the present invention is: a cross-scale information-guided multi-task detection method based on improved YOLOv11, comprising: obtaining a multi-task dataset including traffic target detection, drivable area segmentation and lane line detection, and performing preprocessing; using YOLOv11 as the backbone network, introducing the SPPF module at the end of the network, and improving the target detection and semantic segmentation structure of YOLOv11; using a multi-scale context-aware upsampling module to adaptively capture local and global context information by dynamically learning spatial offsets, and using differentiable bilinear interpolation to achieve feature map amplification; inputting the processed data for network training, using a multi-task loss function to optimize model parameters, using an SGD optimizer and a back-propagation algorithm to update the parameters to obtain a trained network model; based on the detection performance of the model, evaluating it in a real scene, and saving the best model to complete the traffic target detection, drivable area segmentation and lane line detection tasks of multi-task panoramic driving perception.

[0006] The present invention also provides a cross-scale information-guided multi-task detection system based on improved YOLOv11, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement any of the method steps described above.

[0007] The present invention also provides a computer-readable storage medium on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, any of the method steps described above can be implemented.

[0008] Compared with the prior art, the present invention has the following beneficial effects:

[0009] 1. The multi-task panoramic driving perception method constructed by the present invention based on the improved YOLOv11 can accurately perform tasks such as traffic target detection, drivable area segmentation and lane line detection under the framework of multi-task joint learning, achieving high driving perception accuracy and robustness.

[0010] 2. The present invention only uses category annotations, without the need for a series of additional manual annotations such as annotation boxes.

[0011] 3. In view of the particularity of multi-task driving perception tasks, the present invention makes full use of the contextual information between different tasks and the multi-scale feature fusion strategy, improves the efficiency of multi-task collaborative learning, and enhances the model's ability to represent different tasks.

[0012] 4. In autonomous driving tasks, traffic targets of the same type usually have similar appearance and structure. The present invention effectively captures the fine-grained features between targets by introducing a multi-scale context perception module and feature fusion mechanism, thereby improving the recognition accuracy of similar targets and reducing interference between different tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0014] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0016] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0017] The present invention provides a cross-scale information-guided multi-task detection method based on an improved YOLOv11, comprising: obtaining a multi-task dataset including traffic target detection, drivable area segmentation, and lane line detection, and performing preprocessing; adopting YOLOv11 as a backbone network, introducing an SPPF module at the end of the network, and improving the target detection and semantic segmentation structures of YOLOv11; using a multi-scale context-aware upsampling module to adaptively capture local and global context information by dynamically learning spatial offsets, and adopting differentiable bilinear interpolation to achieve feature map amplification; inputting processed data for network training, adopting a multi-task loss function to optimize model parameters, and using an SGD optimizer and a back-propagation algorithm to update the parameters to obtain a trained network model; evaluating the model's detection performance in real scenarios, saving the best model, and completing the multi-task panoramic driving perception traffic target detection, drivable area segmentation, and lane line detection tasks.

[0018] The following is a specific implementation process of the present invention.

[0019] like Figure 1 As shown, this embodiment provides a cross-scale information-guided multi-task detection method based on improved YOLOv11, which specifically includes the following steps:

[0020] Step S1: Collect open source datasets to obtain multi-task dataset information including traffic target detection, drivable area segmentation, and lane line detection; use basic data enhancement techniques such as geometric distortion; and use data enhancement based on environmental perception to enhance adaptability to complex environments;

[0021] Step S2: YOLOv11 is used as the backbone network; the SPPF module is introduced at the end of the backbone network to enhance feature extraction capabilities; the single-task object detection and semantic segmentation structure of YOLOv11 is improved to enable it to support multi-task joint learning of traffic object detection, drivable area segmentation, and lane detection; a multi-scale context-aware upsampling module is introduced to improve segmentation performance;

[0022] Step S3: Calculate the dynamic offset based on the input features; implement the geometric transformation by adding the offset to the base coordinate grid and normalizing it; perform bilinear sampling on the input feature map to achieve efficient upsampling operation;

[0023] Step S4: Input the processed data into the improved multi-task network for training; use the multi-task loss function for optimization to ensure balanced learning of each task; use the backpropagation algorithm and SGD optimizer to update the model parameters; evaluate the detection performance of the model in real scenarios and derive the final performance indicators.

[0024] In this embodiment, step S1 specifically includes the following steps:

[0025] Step S11: Collect open source datasets, including data information for various tasks such as traffic target detection, drivable area segmentation, and lane detection;

[0026] Step S12: Use basic data augmentation techniques such as geometric distortion to enrich the diversity of training samples and improve the robustness of the model. These techniques effectively enhance the richness of the dataset by changing the geometric structure of the image, such as rotation, scaling, cropping, and translation, thereby helping the model better cope with various challenges in real-world applications.

[0027] Step S13: Data augmentation based on environmental perception is used to simulate different lighting conditions, such as dawn and dusk, strong sunlight, cloudy days, and nighttime, to improve the model's adaptability to lighting changes. Generative Adversarial Networks (GANs) are used to generate extreme weather scenarios, such as rain, snow, and fog, to enhance the model's stability in adverse environments. Occlusions, such as vehicles, pedestrians, and traffic signs, are artificially added to enable the model to better cope with occlusions in real-world driving environments.

[0028] In this embodiment, step S2 specifically includes the following steps:

[0029] Step S21: YOLOv11 is used as the backbone network, integrating multiple innovative designs, including an improved feature pyramid structure, an efficient channel attention mechanism, and an optimized gradient propagation path, providing powerful feature extraction capabilities for multi-task learning. The network consists of five downsampling stages (P1-P5), each equipped with an improved C3k2 feature enhancement module, which can effectively capture multi-scale features from local details to global semantics. The specific calculation method is as follows:

[0030] Input feature map: Among them, B represents the batch size, C in represents the number of input channels, H and W are the spatial sizes of the feature map.

[0031] The input channel is split into two parts through 1×1 convolution:

[0032]

[0033] in Indicates the number of hidden layer channels, C out Represents the number of output channels, e represents the expansion coefficient, and split means doubling the number of feature map channels through 1×1 convolution and then dividing it into two independent parts for different processing.

[0034] The main branch handles X2 via n Bottleneck or C3k modules:

[0035]

[0036] Bottleneck(X)=Conv k×k (Conv 1×1 (X))+X(if shortcut))

[0037] Where X2' represents the output result of feature map X2 after processing by the main branch, Bottleneck is an efficient residual module, shortcut is a Boolean parameter used to control whether to add residual connection, Sequential refers to the sequence structure of Bottleneck or C3k2 modules stacked in sequence, g represents the number of group convolutions, k represents the custom convolution kernel size when using C3k, and n BN Indicates the number of Bottleneck stacks.

[0038] The processed branch is concatenated with the original branch and then the dimension is reduced by 1×1 convolution to obtain the output feature map Y C3k2 :

[0039]

[0040] C3k2, the core operator, uses a channel expansion ratio of 0.25 and lightweight mode in the P3 / P4 stages. In the deeper P5 stages, it uses a standard configuration and defaults to the integrated channel attention mechanism. This dynamic design maintains high computational efficiency in the shallow layers while enabling fine-grained semantic modeling in the deeper P5 stages. The number of network channels follows an exponential growth pattern, complementing the spatial downsampling rate to ensure stable pixel information density in feature maps at each stage.

[0041] Step S22: The SPPF module is introduced at the end of the backbone network to enhance the core module of the receptive field at the end of the backbone network, and to capture contextual information of different granularities through multi-scale pooling fusion. The specific calculation method is as follows:

[0042] Input feature map:

[0043] Compress the channels by 1×1 convolution:

[0044] in Indicates the number of hidden layer channels, X SPPF ' represents the input feature map X SPPF Output feature map after channel compression by 1×1 convolution.

[0045] For X SPPF 'Perform k×k maximum pooling and achieve an equivalent large kernel receptive field by repeating pooling three times:

[0046] Y pool1 =MaxPool k×k (X SPPF ')

[0047] Y pool2 =MaxPool k×k (Y pool1 )

[0048] Y pool3 =MaxPool k×k (Y pool2 )

[0049] Among them, Y pool1 ,Y pool2 ,Y pool3 Represents the intermediate feature maps generated by serial maximum pooling, which represent multi-scale contextual information under different receptive fields.

[0050] The original features are concatenated with the pooling results of each stage along the channel dimension, and then the number of channels is restored through 1×1 convolution to obtain the output feature map Y SPPF :

[0051]

[0052] Step S23: Improve the single-task target detection and semantic segmentation structure of YOLOv11 to support multi-task joint learning of traffic target detection, drivable area segmentation, and lane line detection. In response to the multi-task requirements of traffic target detection, drivable area segmentation, and lane line detection in autonomous driving scenarios, the neck structure of YOLOv11 has been systematically improved, the single detection structure of the original YOLOv11 has been decoupled, and a three-branch parallel processing flow has been constructed. In order to further enhance the feature extraction capability, the C2PSA module is introduced into the neck structure, and the self-attention mechanism is used to improve the feature expression capability and strengthen the global information interaction, thereby enhancing the model's adaptability to complex traffic scenarios. The specific calculation process is as follows:

[0053] For the feature map input to C2PSA:

[0054] C2PSA uses the pyramid squeeze attention module PSABlock for self-attention calculation, which is mainly based on dot product attention:

[0055]

[0056] Among them, Q represents the query vector, K represents the key vector, V represents the value vector, and N head Indicates the number of attention heads, N pixel Indicates the number of pixels, d k Indicates the key / query dimension, d v Represents the value dimension.

[0057] Enter X C2PSA After Conv 1×1 After dimensionality reduction, it is split into X C2PSA1 、X C2PSA2 Two parts:

[0058] X C2PSA1 ,X C2PSA2 =split(Conv 1×1 (X C2PSA ))

[0059] X C2PSA2 After PSABlock calculates attention, it is combined with X C2PSA1 Perform channel splicing to obtain the final output Y C2PSA :

[0060] Y C2PSA =Conv 1×1 (Concat(X C2PSA1 ,PSABlock(X C2PSA2 )))

[0061] For traffic object detection, the network gradually constructs feature maps of three different scales through multiple downsampling, feature concatenation, and processing with the C3k2 module, corresponding to the detection requirements of small, medium, and large objects. This design not only retains the efficiency of the original YOLOv11 but also further improves the detection accuracy of multi-scale objects.

[0062] For the lane segmentation task, the network adds a refinement module for high-resolution feature maps. Through dynamic sampling and feature concatenation, the model is able to capture the fine structural information of lane lines in low-level features, significantly improving the accuracy of lane segmentation. Similarly, in the drivable area segmentation branch, the network also utilizes dynamic sampling and multi-scale feature fusion strategies to ensure accurate delineation of the drivable area boundary. In addition, each branch maintains a certain degree of independence in network design to avoid mutual interference between tasks, while reducing computational overhead by sharing some basic feature extraction modules.

[0063] Step S24: Introduce a multi-scale context-aware upsampling module to improve segmentation performance. This module effectively solves the problems of detail loss and edge blurring in traditional methods through multi-level feature fusion and intelligent upsampling mechanism.

[0064] In this embodiment, step S3 specifically includes the following steps:

[0065] Step S31: Calculate the dynamic offset based on the input features. The calculation method is as follows:

[0066] For the input feature map: Map the input features to the offset space Δ through 1×1 convolution final :

[0067]

[0068] Among them, P init Represents a predefined fixed coordinate offset, G is the number of groups, and S is the upsampling multiple.

[0069] Step S32: geometric transformation is achieved by adding an offset to the base coordinate grid and normalizing it. The specific calculation method is as follows:

[0070] Given an input feature map of size H×W, we first define the standard grid coordinates, C h Represents the standard grid coordinate in the height direction, h is the index in the height direction, C w Represents the standard grid coordinate in the width direction, where w is the index in the width direction:

[0071]

[0072] Constructing the base mesh: C base =(C h ,C w )∈R 2×H×W , and then apply a dynamic offset to it:

[0073] C base =C base +Δ final

[0074] Normalize to the range [-1, 1] to meet the requirements of grid sampling:

[0075]

[0076] Step S33: Perform bilinear sampling on the input feature map to achieve efficient upsampling operation. norm Next, from the input feature map X dy Perform interpolation calculation to generate a new upsampled feature map Y dy .

[0077] Get the floating point coordinates (x, y) of the target pixel:

[0078]

[0079] Since (x, y) may not be an integer pixel, we need to use the surrounding integer pixels for interpolation. For any floating point coordinate (x, y), the nearest 4 integer pixels are:

[0080] (x0,y0),(x1,y0),(x0,y1),(x1,y1)

[0081] x1=x0+1

[0082] y1=y0+1

[0083] Calculate the relative offset of the target floating point coordinate to the nearest integer coordinate:

[0084] Δx=x-x0, Δy=y-y0

[0085] According to the bilinear interpolation formula, calculate the weights of the four integer pixels:

[0086] w 00 =(1-Δx)(1-Δy)

[0087] w 01 =Δx(1-Δy)

[0088] w 10 =(1-Δx)Δy

[0089] w 11 =ΔxΔy

[0090] Using the weights calculated above, we perform a weighted sum of the four pixel values:

[0091]

[0092] Apply this interpolation calculation to the entire new coordinate grid in parallel to generate the entire output feature map Y dy . Finally, the shape of the output feature map is:

[0093] In this embodiment, step S4 specifically includes the following steps:

[0094] Step S41: The processed data is fed into the improved multi-task network for training. Leveraging the enhanced feature representation and optimized network structure, the rich feature information required for complex tasks such as traffic object detection, drivable area segmentation, and lane detection can be more effectively learned.

[0095] Step S42: By detecting the target L det loss and segmentation task loss L seg Combined into a total loss function L total , achieving multi-task joint optimization, ensuring balanced learning of each task, and making full use of the correlation between target detection, drivable area segmentation and lane detection. The calculation method is as follows:

[0096] The total loss function is L total =L det +L seg .

[0097] The object detection loss is composed of the bounding box loss L box , classification loss L cls and distribution focus loss L dfl composition:

[0098] L det =λ box L box +λ cls L cls +λ dfl L dfl

[0099] Both drivable area segmentation and lane line detection use segmentation loss L seg , which consists of Tversky loss and focal loss:

[0100] L seg =λ TL L TL +λ FLL FL

[0101] Among them, the hyperparameter λ is introduced box ,λ cls ,λ dfl ,λ TL ,λ FL , to adjust the weights of different subtasks and ensure balanced learning of each task.

[0102] Step S43: Update the model parameters using the backpropagation algorithm and the SGD optimizer. Backpropagation propagates the error gradient layer by layer using the chain rule, adjusting the weights backward from the output layer to the input layer. SGD, as an optimization strategy, calculates gradients based on a mini-batch of samples at each iteration and updates the parameters according to a set learning rate, gradually optimizing model performance and ensuring efficient convergence to the optimal solution.

[0103] Step S44: Evaluate the model's detection performance in real-world scenarios and derive final performance metrics. In real-world application scenarios, the model's detection performance is comprehensively evaluated to ensure its robustness and generalization capabilities in complex environments. This evaluation process typically involves running the model on different test datasets or real-world scenarios, analyzing its performance under various conditions, and calculating key performance metrics.

[0104] In particular, this embodiment only uses category annotations and does not require a series of additional manual annotations such as annotation boxes. Under the particularity of multi-task driving perception tasks, the present invention makes full use of the visual and semantic differences between various types of targets to provide key task identification information. Most existing methods focus on relying on global features or high-level semantic information, while the features of different tasks contain semantic information at different levels. By fusing features at different levels, the present invention can effectively improve the representation ability of the model in multi-task learning, while enhancing the ability to accurately recognize multi-scale and multi-task targets. Considering that target objects in the same task often have similar structures and visual features, the present invention introduces a multi-scale context-aware upsampling module and combines it with a multi-task joint learning strategy to effectively extract feature semantic information between different tasks, thereby reducing misjudgment and confusion between different task targets. Through this method, the multi-task panoramic driving perception model based on the improved YOLOv11 proposed by the present invention can efficiently perform multi-task learning and accurately perform tasks such as traffic target detection, drivable area segmentation and lane line detection, ultimately improving the overall perception accuracy and robustness of the model in complex traffic scenarios.

[0105] The present invention also provides a cross-scale information-guided multi-task detection system based on improved YOLOv11, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement any of the method steps described above.

[0106] The present invention also provides a computer-readable storage medium on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, any of the method steps described above can be implemented.

[0107] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions and effects do not exceed the scope of the technical solution of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A cross-scale information guided multi-task detection method based on improved YOLOv11, characterized in that: include: A multi-task dataset including traffic target detection, drivable area segmentation and lane line detection is obtained and preprocessed; YOLOv11 is used as the backbone network, and the SPPF module is introduced at the end of the network. At the same time, the target detection and semantic segmentation structure of YOLOv11 are improved; a multi-scale context-aware upsampling module is used to adaptively capture local and global context information by dynamically learning spatial offsets, and differentiable bilinear interpolation is used to achieve feature map amplification; the processed data is input into the network training, and the model parameters are optimized using a multi-task loss function. The SGD optimizer and back-propagation algorithm are used for parameter update to obtain the trained network model; based on the model's detection performance, it is evaluated in real scenarios and the best model is saved to complete the multi-task panoramic driving perception traffic target detection, drivable area segmentation and lane line detection tasks.

2. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 1 is characterized in that: The method comprises the following steps: Step S1: Collect open source datasets to obtain multi-task dataset information including traffic target detection, drivable area segmentation, and lane line detection; use basic data enhancement technology to enhance data; use data enhancement based on environmental perception to enhance adaptability to complex environments; Step S2: YOLOv11 is used as the backbone network; the SPPF module is introduced at the end of the backbone network to enhance feature extraction capabilities; the single-task object detection and semantic segmentation structure of YOLOv11 is improved to enable it to support multi-task joint learning of traffic object detection, drivable area segmentation, and lane detection; a multi-scale context-aware upsampling module is introduced to improve segmentation performance; Step S3: Calculate the dynamic offset based on the input features; implement the geometric transformation by adding the offset to the base coordinate grid and normalizing it; Perform bilinear sampling on the input feature map to achieve efficient upsampling operation; Step S4: Input the processed data into the improved multi-task network for training; use the multi-task loss function for optimization to ensure balanced learning of each task; use the backpropagation algorithm and SGD optimizer to update the model parameters; evaluate the detection performance of the model in real scenarios and derive the final performance indicators.

3. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 2 is characterized in that Step S1 is specifically implemented as follows: Step S11: Collect open source datasets, including data information for various tasks such as traffic target detection, drivable area segmentation, and lane detection; Step S12: Using basic data augmentation techniques to enrich the diversity of training samples and improve the robustness of the model; Step S13: Use data enhancement based on environmental perception to simulate different lighting conditions including dawn and dusk, strong light, cloudy days, and nighttime to improve the model's adaptability to lighting changes. Use GAN to generate extreme weather scenes including rainy days, snowy days, and foggy days to improve the model's stability in harsh environments. Artificially add occlusion conditions including vehicles, pedestrians, and traffic signs to enable the model to better cope with occlusion problems in real driving environments.

4. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 2 or 3, characterized in that Basic data enhancement techniques include basic data enhancement techniques of geometric distortion, that is, basic data enhancement techniques by changing the geometric structure of the image, including operations of rotation, scaling, cropping and translation.

5. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 2 is characterized in that Step S2 is specifically implemented as follows: Step S21: YOLOv11 is used as the backbone network, integrating an improved feature pyramid structure, an efficient channel attention mechanism, and an optimized gradient propagation path to provide powerful feature extraction capabilities for multi-task learning. The network consists of five downsampling stages, each equipped with an improved C3k2 feature enhancement module, which can effectively capture multi-scale features from local details to global semantics. The specific calculation method is as follows: Input feature map: Among them, B represents the batch size, C in represents the number of input channels, H and W are the spatial sizes of the feature map; Through 1×1 convolution Conv 1×1 Split the input channel into two parts X1 and X2: in Indicates the number of hidden layer channels, C out Represents the number of output channels, e represents the expansion coefficient, and split refers to doubling the number of feature map channels through 1×1 convolution and then dividing it into two independent parts for different processing. The main branch processes X2 through n Bottleneck or C3k2 feature enhancement modules: Bottleneck(X)=Conv k×k (Conv 1×1 (X))+X(if shortcut) Where X2' represents the output result of feature map X2 after processing by the main branch, Bottleneck is an efficient residual module, shortcut is a Boolean parameter used to control whether to add residual connection, Sequential refers to the sequence structure of Bottleneck or C3k2 modules stacked in sequence, g represents the number of group convolutions, k represents the custom convolution kernel size when using the C3k2 feature enhancement module, and n BN Indicates the number of times Bottleneck is stacked; The processed branch is concatenated with the original branch and then the dimension is reduced by 1×1 convolution to obtain the output feature map Y C3k2 : Concat represents the concatenation operation; Step S22: The SPPF module is introduced at the end of the backbone network to enhance the receptive field of the backbone network terminal and capture contextual information of different granularities through multi-scale pooling fusion. The specific calculation method is as follows: Input feature map: Compress the channels by 1×1 convolution: in Indicates the number of hidden layer channels, X SPPF ' represents the input feature map X SPPF Output feature map after channel compression by 1×1 convolution; For X SPPF 'Perform k×k maximum pooling and achieve an equivalent large kernel receptive field by repeating pooling three times: Y pool1 =MaxPool k×k (X SPPF ′) AND pool2 =MaxPool k×k (AND pool1 ) AND pool3 =MaxPool k×k (AND pool2 ) Among them, Y pool1 ,Y pool2 ,Y pool3 Represents the intermediate feature maps generated by the cascaded maximum pooling, which represent multi-scale contextual information under different receptive fields; The original features are concatenated with the pooling results of each stage along the channel dimension, and then the number of channels is restored through 1×1 convolution to obtain the output feature map Y SPPF : Step S23: Improve the single-task object detection and semantic segmentation structure of YOLOv11 to support multi-task joint learning of traffic object detection, drivable area segmentation, and lane detection. To meet the multi-task requirements of traffic object detection, drivable area segmentation, and lane detection in autonomous driving scenarios, the neck structure of YOLOv11 is systematically improved. The single detection structure of the original YOLOv11 is decoupled, and a three-branch parallel processing flow is constructed. The C2PSA module is introduced into the neck structure, and the self-attention mechanism is used to improve feature expression capabilities and strengthen global information interaction, thereby enhancing the model's adaptability to complex traffic scenarios. The specific calculation process is as follows: For the feature map input to the C2PSA module: The C2PSA module uses the pyramid squeeze attention module PSABlock for self-attention calculation, based on dot product attention: Among them, Q represents the query vector, K represents the key vector, V represents the value vector, and N head Indicates the number of attention heads, N pixel Indicates the number of pixels, d k Indicates the key / query dimension, d v Represents the value dimension; Enter X C2PSA After Conv 1×1 After dimensionality reduction, it is split into X C2PSA1 、X C2PSA2 Two parts: X C2PSA1 ,X C2PSA2 =split(Conv 1×1 (X C2PSA )) X C2PSA2 After PABlock calculates the attention, it is combined with X C2PSA1 Perform channel splicing to obtain the final output Y C2PSA : Y C2PSA =Conv 1×1 (Concat(X C2PSA1 ,PSABlock(X C2PSA2 ))) For the traffic target detection branch, the network gradually constructs feature maps of three different scales through multiple downsampling, feature splicing, and C3k2 feature enhancement module processing, corresponding to the detection requirements of small targets, medium targets, and large targets respectively. For the lane line segmentation task branch, the network adds a refinement processing module for high-resolution feature maps on the basis of the original. Through dynamic sampling and feature splicing operations, the model can capture the fine structural information of lane lines in low-level features, thereby significantly improving the accuracy of lane line segmentation. For the drivable area segmentation branch, the network utilizes dynamic sampling and multi-scale feature fusion strategies to ensure accurate characterization of the drivable area boundary. In addition, each branch maintains independence in network design to avoid mutual interference between tasks, while reducing computational overhead by sharing some basic feature extraction modules. Step S24: Introduce a multi-scale context-aware upsampling module to solve the problems of detail loss and edge blurring through multi-level feature fusion and intelligent upsampling mechanism.

6. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 5, characterized in that: In step S21, the C3k2 feature enhancement module adopts a channel expansion rate of 0.25 and a lightweight mode in the P3 and P4 stages, and enables the standard configuration in the deep P5 stage and integrates the channel attention mechanism by default for fine semantic modeling.

7. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 5, characterized in that: Step S3 is specifically implemented as follows: Step S31: Calculate the dynamic offset based on the input features. The calculation method is as follows: For the input feature map: Map the input features to the offset space Δ through 1×1 convolution final : Among them, P init Represents a predefined fixed coordinate offset, G is the number of groups, and S is the upsampling multiple; Step S32: geometric transformation is achieved by adding an offset to the basic coordinate grid and normalizing it. The specific calculation method is as follows: Given an input feature map of size H×W, we first define the standard grid coordinates, C h Represents the standard grid coordinate in the height direction, h is the index in the height direction, C w Represents the standard grid coordinate in the width direction, where w is the index in the width direction: Constructing the base mesh: C base =(C h ,C w )∈R 2×H×W , and then applying a dynamic offset to it, we get C base ': C base ′=C base +D final Normalize to the range [-1,1] to meet the requirements of grid sampling: C norm C base 'Normalize the results; Step S33: perform bilinear sampling on the input feature map to achieve efficient upsampling operation; in the calculated offset coordinate C norm Next, from the input feature map X dy Perform interpolation calculation to generate a new upsampled feature map Y dy ; Get the floating point coordinates (x, y) of the target pixel: Since (x, y) may not be an integer pixel, it is necessary to use the surrounding integer pixels for interpolation; for any floating-point coordinate (x, y), the nearest 4 integer pixels are: (x0,y0),(x1,y0),(x0,y1),(x1,y1) Calculate the relative offset of the target floating point coordinate to the nearest integer coordinate: Δx=x-x0, Δy=y-y0 According to the bilinear interpolation formula, calculate the weights of the four integer pixels: w 00 =(1-Δx)(1-Δy) w 01 =Δx(1-Δy) w 10 =(1-Δx)Δy w 11 =ΔxΔy Using the weights calculated above, we perform a weighted sum of the four pixel values: Apply the interpolation calculation to the entire new coordinate grid in parallel to generate the entire output feature map Y dy ; Finally, the shape of the output feature map is:

8. The cross-scale information guided multi-task detection method based on improved YOLOv11 according to claim 7 is characterized in that: Step S4 is specifically implemented as follows: Step S41: The processed data is input into the improved multi-task network for training; using the enhanced feature representation and optimized network structure, the rich feature information required for complex tasks including traffic object detection, drivable area segmentation, and lane detection is learned; Step S42: By setting the target detection loss L det and segmentation task loss L seg Combined into a total loss function L total , achieving multi-task joint optimization, ensuring balanced learning of each task, and making full use of the correlation between target detection, drivable area segmentation and lane detection; the calculation method is as follows: The total loss function is L total =L det +L seg ; The object detection loss is composed of the bounding box loss L box , classification loss L cls and distribution focus loss L dfl composition: L det =λ box L box +λ cls L cls +λ dfl L dfl Both drivable area segmentation and lane line detection use segmentation task loss L seg , by Tversky loss L TL and focal loss L FL composition: L seg =λ TL L TL +λ FL L FL Among them, the hyperparameter λ is introduced box ,λ cls ,λ dfl ,λ TL ,λ FL , to adjust the weights of different subtasks and ensure balanced learning of each task. Step S43: Update the model parameters using the backpropagation algorithm and SGD optimizer. Backpropagation propagates the error gradient layer by layer using the chain rule, adjusting the weights backward from the output layer to the input layer. SGD is used as an optimization strategy to calculate the gradient based on a small batch of samples at each iteration and update the parameters according to the set learning rate, thereby gradually optimizing the model performance and ensuring efficient convergence to the optimal solution. Step S44: Evaluate the detection performance of the model in real scenarios and derive the final performance indicators; in actual application scenarios, comprehensively evaluate the detection performance of the model to ensure its robustness and generalization ability in complex environments; the evaluation process includes running the model in different test data sets or real-world scenarios, analyzing its performance under various conditions, and calculating key performance indicators.

9. A cross-scale information guided multi-task detection system based on improved YOLOv11, characterized in that: The method comprises a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium storing computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

Images