A target detection method fusing edge information protection and geometric perception feature extraction
By improving the YOLOv11n framework and introducing Haar wavelet downsampling, variable kernel convolution, and cross-dimensional interactive attention modules, the problems of edge information loss, irregular target representation, and semantic focusing in complex scenes of YOLOv11n are solved, and high-precision and robust road obstacle detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAIYIN INSTITUTE OF TECHNOLOGY
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-21
AI Technical Summary
The existing YOLOv11n object detection framework suffers from problems such as loss of high-frequency edge information, insufficient ability of fixed-shape convolution kernels to represent irregular targets, and insufficient semantic focusing ability in complex backgrounds when dealing with complex scenes, resulting in insufficient detection accuracy and robustness.
By introducing a Haar wavelet downsampling module, a geometrically adaptive variable kernel convolution module, a cross-dimensional interactive attention module, and an attention-guided subpixel reconstruction module, YOLOv11n is improved to construct an enhanced road obstacle detection model, achieving edge information protection, geometric perception feature extraction, and deep semantic focusing.
It significantly improves the accuracy and robustness of detecting multi-scale and irregularly shaped targets in complex environments, solves the problem of sluggish perception of fine targets by lightweight detectors, and meets the real-time inference requirements of vehicle-mounted equipment.
Smart Images
Figure CN122435239A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and deep learning technology, specifically to a target detection method that integrates edge information protection and geometric perception feature extraction. Background Technology
[0002] Object detection, as a core technology in computer vision and deep learning, directly impacts the reliability of intelligent perception systems through the stability of its detection accuracy and the robustness of its algorithms. In practical applications, targets often exhibit characteristics such as large scale spans, diverse geometric shapes, and complex environmental backgrounds, placing extremely high demands on the underlying feature processing mechanisms of algorithms.
[0003] In recent years, lightweight one-stage object detection algorithms, represented by YOLOv11n, have demonstrated significant technical advantages in the field of real-time perception due to their excellent inference speed and end-to-end detection architecture. YOLOv11n introduces features such as the C3k2 feature extraction module and the C2PSA module based on the attention mechanism, aiming to improve the ability to capture complex semantics.
[0004] However, in-depth research and practice have revealed that even when using the representative framework YOLOv11n, its original architecture still faces the following deep technical challenges when dealing with specific complex scenarios (such as occluded obstacles, distant small targets, and irregularly shaped targets):
[0005] 1. Feature information loss in downsampling and upsampling paths: The original YOLOv11n still uses traditional convolutions with a stride of 2 for downsampling in the backbone network and relies on interpolation algorithms for upsampling in the neck network. This sampling mechanism leads to irreversible pixel-level loss and texture blurring of high-frequency edge information of targets (especially small targets at a distance) during feature map scaling, making it difficult for the model to capture accurate boundary features in deep networks.
[0006] 2. Standard convolution operators have limited geometric perception capabilities for irregularly shaped targets: The geometric topology of a target (such as elongated or flattened forms) is strongly correlated with its response region in image space. YOLOv11n's C3k2 and other modules often use fixed square convolution kernels to extract features, failing to adaptively adjust the receptive field according to the target's actual geometry. This "operator-shape mismatch" leads to the introduction of a large amount of irrelevant background noise during feature extraction, resulting in low efficiency in representing irregularly shaped targets and hindering further improvements in localization accuracy.
[0007] 3. Insufficient background suppression and focusing during deep semantic interaction: Although YOLOv11n introduces an attention mechanism, its semantic focusing ability for deep features remains insufficient in extreme scenarios with drastic changes in lighting, road shadows, and complex texture interference. Due to the lack of an efficient cross-dimensional feature interaction mechanism, the model struggles to accurately focus on key semantic information globally and effectively suppress background interference, making it prone to false positives and false negatives.
[0008] In summary, the original YOLOv11n architecture lacks specific optimizations in areas such as edge information protection, geometric shape adaptation, and cross-dimensional semantic focus. Therefore, its overall detection performance and robustness still have considerable room for improvement when dealing with complex and ever-changing real-world application scenarios. Summary of the Invention
[0009] To address the technical challenges of existing target detection frameworks, such as loss of high-frequency edge information during downsampling and upsampling, insufficient representation ability of fixed-shape convolution kernels for irregular and irregular targets, and lack of semantic focus under complex background interference, this technical solution provides a target detection method that integrates edge information protection and geometric perception feature extraction. By introducing Haar wavelet downsampling, designing a geometrically adaptive variable kernel convolution feature extraction operator, a cross-dimensional interactive attention mechanism, and an attention-guided sub-pixel reconstruction strategy, the YOLOv11n architecture is deeply reconstructed, constructing a feature perception model with high accuracy and strong robustness, enabling accurate capture and recognition of multi-scale and irregular targets in complex environments; effectively solving the technical problems.
[0010] This invention is achieved through the following technical solution:
[0011] A target detection method integrating edge information protection and geometric perception feature extraction includes the following steps: improving the YOLOv11n model to construct an enhanced road obstacle detection model; the improvement includes:
[0012] (1) In the downsampling layer of the backbone network, the Haar wavelet downsampling module (HWD) is introduced to replace the traditional convolutional downsampling;
[0013] (2) In the feature extraction units of the backbone network and the neck network, the C3k2 module of the integrated variable kernel convolution module AKConv is used to perform geometric adaptive feature extraction on the irregular target;
[0014] (3) The cross-dimensional interactive attention module C2EMA is used to reconstruct the feature representation in the deep layer of the backbone network to enhance the model’s ability to suppress background and perceive context under complex road conditions;
[0015] (4) In the upsampling layer of the Neck network, the attention-guided subpixel reconstruction module EMA-PixelShuffle is used to replace the original traditional upsampling module to achieve lossless restoration of multi-level features.
[0016] Furthermore, the Haar wavelet downsampling module HWD described in step (1) is set in the starting layer of the backbone network and replaces the downsampling layer before each feature extraction layer C3k2 in the backbone network; the HWD module uses the Haar wavelet basis function to decompose the feature map into four sub-band components LL, LH, HL and HH, and achieves half resolution through channel splicing; and maps spatial information to the channel dimension without loss.
[0017] Furthermore, the Haar wavelet downsampling module HWD, when performing downsampling, includes the following steps for edge-fidelity downsampling: The input feature map X is processed by the Discrete Wavelet Transform (DWT) operator, and decomposed in the horizontal and vertical directions using Haar wavelet basis functions to generate four complementary sub-band components: a low-frequency approximation component (LL), a horizontal high-frequency component (LH), a vertical high-frequency component (HL), and a diagonal high-frequency component (HH); the four sub-band components are then concatenated along the channel dimension to obtain the number of channels as the input feature. Figure Four Enhanced feature maps with dimensions reduced by half and width and height; this method maps edge transition information in the original space to the channel dimension, avoiding the diffusion of obstacle contour information caused by pixel discarding in traditional stride convolution.
[0018] Furthermore, the C3k2 module integrating AKConv described in step (2) replaces the 3×3 convolution in the standard bottleneck structure with a variable kernel convolution AKConv; it generates the deformation deviation value of the branch prediction sampling point coordinates through offset, induces the convolution kernel to produce geometric deformation to fit the contour of the irregular target, and enhances the feature representation ability of irregular targets.
[0019] Furthermore, the AKConv method for achieving geometrically adaptive feature extraction includes the following steps: the module contains a parallel offset generation branch and a sampling feature extraction branch, which predicts the deviation vector of the sampling coordinates based on the semantics of the currently input obstacle; for example, when the detected target is a slender pedestrian, the convolution sampling points are automatically stretched vertically; when the target is a flat vehicle, the sampling points are expanded horizontally; through this geometrically adaptive deformation capability, the model achieves accurate extraction of features of irregular obstacles; the offset generation branch learns a set of parameterized offsets Δpn based on the semantic content of the input features, which is used to define the irregular deformation trajectory of the sampling grid points of the convolution kernel; the sampling feature extraction branch performs bilinear interpolation sampling on the input features based on the offsets, so that the convolution receptive field can dynamically fit the vertical shape of pedestrians, the horizontal shape of vehicles, or the irregular geometric shape of roadblocks, thereby achieving accurate coverage of features of irregular obstacles.
[0020] Furthermore, the cross-dimensional interactive attention module C2EMA described in step (3) replaces the internal PSABlock with an efficient multi-scale attention EMA structure. It adopts the EMA mechanism to establish cross-dimensional feature dependencies by performing one-dimensional global average pooling in the horizontal and vertical directions, so that the model can achieve deep interaction between space and channel without channel reduction, thereby achieving accurate semantic focus in complex scenes.
[0021] Furthermore, the cross-dimensional interactive attention module C2EMA establishes cross-dimensional associations by dividing the input features into multiple subgroups along the channel dimension and inputting the spatial feature encoding branch and the channel feature exchange branch in parallel. The spatial feature encoding branch captures the long-range spatial dependencies of obstacles by performing one-dimensional global average pooling in the horizontal (X) and vertical (Y) directions respectively. The channel feature exchange branch realizes feature interaction through cross-dimensional vector dot product operations. The outputs of the two branches generate multi-scale attention weights through the Sigmoid activation function and multiply them element-wise with the original features. This module achieves deep fusion of spatial and channel attention without dimensionality reduction, enhancing the model's target focusing ability in complex backgrounds.
[0022] Furthermore, the attention-guided subpixel reconstruction module described in step (4) first uses the internal EMA submodule to perform semantic weight calibration on low-resolution features to suppress redundant background noise; then, it performs physical recombination by "trading channels for space" through subpixel convolutional layers, periodically rearranging high-dimensional channel features to spatial dimensions, solving the texture blurring problem caused by traditional interpolation amplification; it uses subpixel convolution to recombine high-dimensional channel features into high-resolution spatial features, and with multi-scale attention guidance, ensures the boundary clarity and pixel-level alignment during the feature fusion process.
[0023] Furthermore, the sub-pixel reconstruction module described in step (4) processes the feature map in the following ways: First, the low-resolution feature map is input into the EMA attention submodule, and the semantic weight of the key regions of the obstacle is strengthened by the cross-dimensional interaction mechanism to suppress background noise; Second, the strengthened feature map is input into the sub-pixel convolution PixelShuffle layer, and the high-frequency information in the channel dimension is mapped to the spatial dimension by the periodic rearrangement operator, and the width and height of the feature map are enlarged to r times the original size, where r is the upsampling factor; Finally, the lossless restoration method of "trading channels for space" is executed; the edge blurring problem caused by traditional interpolation upsampling is solved, and the feature map output by the Neck layer is highly aligned with the shallow feature map extracted by the backbone network in terms of spatial resolution and semantic distribution.
[0024] Furthermore, the method also includes the following steps:
[0025] Before improving the YOLOv11n model, the following steps were taken: preprocessing and data augmentation were performed on the pre-acquired road environment obstacle image dataset, and the processed dataset was divided into training set, validation set and test set;
[0026] And the setup for improving the YOLOv11n model: using the training set to iteratively train the improved model, optimize the model parameters, and obtain a trained obstacle detection model;
[0027] The trained model is used to identify and locate obstacles in the road environment image to be detected, and outputs the bounding box and category information of the target object.
[0028] Beneficial effects
[0029] The target detection method proposed in this invention, which integrates edge information protection and geometric perception feature extraction, has the following advantages compared with existing technologies:
[0030] (1) This invention achieves full-path edge fidelity and pixel-level reconstruction of the feature flow. Addressing the information loss issue caused by the original YOLOv11n architecture during sampling, a closed-loop feature fidelity link is constructed by embedding Haar wavelet transform (HWD) in the downsampling stage and introducing attention-guided sub-pixel reconstruction (EMA-PixelShuffle) in the upsampling stage. This design utilizes frequency mapping and spatial rearrangement techniques to change the traditional algorithm's mechanical pixel discarding and fuzzy interpolation, achieving pixel-level restoration of high-frequency edge information of targets (especially distant and small targets). Compared to existing technologies, this method significantly improves the alignment accuracy in the feature fusion stage, fundamentally solving the problem of lightweight detectors being slow to perceive subtle targets.
[0031] (2) This invention breaks through the geometric limitations of standard convolutional kernels, achieving adaptive alignment between feature extraction and target morphology. By reconstructing the core components of feature extraction at the operator level, the variable kernel convolution (AKConv) is deeply integrated into the residual learning path of YOLOv11n. This improvement frees the model from the rigid constraints of fixed square convolutional kernels on feature representation, giving the network the ability to dynamically induce geometric deformation of the convolutional kernel according to the target semantics. This "morphological adaptation" design enables the operator to accurately match the physical contours of irregular targets such as slender shapes (pedestrians) and flat shapes (motor vehicles), significantly improving the regression accuracy (IoU) of the bounding box and effectively eliminating the interference of redundant background information, thus greatly improving the recognition efficiency of complex geometric targets.
[0032] (3) This invention enhances the deep semantic focusing capability and achieves robust detection in complex dynamic backgrounds. Addressing the issue of YOLOv11n being easily confused by background noise during deep feature abstraction, a cross-dimensional attention interaction mechanism is introduced through the C2EMA module. This design enables deep fusion and collaborative weighting of spatial and channel features without reducing channel dimensions, actively filtering shadows, strong light, and complex texture interference. Through precise self-focusing on the core semantics of the target, this invention constructs a highly robust feature flow, giving it excellent anti-interference capabilities in complex and dynamic detection scenarios. Simultaneously, all improved modules adhere to lightweight design principles, ensuring that while significantly improving accuracy, the model still meets the real-time inference requirements of terminal devices such as vehicles. Attached Figure Description
[0033] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0034] Figure 2 This is a schematic diagram of the overall architecture of the improved yolov11n model in this invention.
[0035] Figure 3 This is a schematic diagram of the workflow of the Haar wavelet downsampling module in this invention.
[0036] Figure 4 This is a schematic diagram of the internal architecture of the C3k2 module that integrates AKConv in this invention.
[0037] Figure 5 This is a schematic diagram of the internal architecture of the cross-dimensional interactive attention module in this invention.
[0038] Figure 6 This is a schematic diagram of the internal architecture of the sub-pixel reconstruction module in this invention. Detailed Implementation
[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. The described embodiments are merely some embodiments of the present invention, and not all embodiments. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention should fall within the protection scope of the present invention.
[0040] Example 1:
[0041] like Figure 1 As shown, a target detection method integrating edge information protection and geometric perception feature extraction includes the following steps:
[0042] Step 1: Acquire road environment obstacle image data in advance, perform preprocessing and data augmentation, and divide the processed dataset into training, validation, and test sets; the specific operation method is as follows:
[0043] Step 1.1: Data Acquisition and Labeling: Using a vehicle-mounted high-definition camera, road scene images were acquired under different weather conditions (sunny, rainy, cloudy), different lighting conditions (daytime, dusk, nighttime), and different road types (urban main roads, highways, rural roads). The acquired raw images were subjected to quality screening and data cleaning to remove blurry and invalid data. Labeling tools were used to annotate the cleaned images with bounding boxes to construct a basic dataset containing labels for four typical obstacle categories: pedestrians, vehicles, non-motorized vehicles, and traffic barriers.
[0044] Step 1.2: Introduce negative and interference samples: To improve the model's generalization ability in complex road scenes, a certain proportion of complete unobstructed road surface images are added to the dataset as negative samples, as well as images containing strong shadow interference, ground reflection, rain and fog obstruction, complex lane line textures, and other interference.
[0045] Step 1.3: Data Augmentation: The training set is augmented using a combination of multi-dimensional spatial transformation and color space perturbation strategies. Specifically, this includes: using random horizontal flipping, random rotation (-15° to 15°), and multi-scale random scaling (0.8 to 1.2 times) to simulate different installation angles and target distances of the vehicle-mounted camera; simulating varying weather, lighting, and shadow occlusion conditions by randomly adjusting brightness and contrast (±30%); randomly adding Gaussian noise and salt-and-pepper noise; and finally, the Mosaic 9-patch augmentation algorithm is introduced as a key feature. By randomly stitching together 9 images at a certain ratio, a complex sample with dense targets is artificially constructed. This aims to increase the sampling density of small obstacles at long distances during training, significantly improving the model's ability to extract and represent features of small targets.
[0046] Step 1.4: Dataset partitioning: All processed image data are randomly divided into training set, validation set and test set in a ratio of 8:1:1, which are used for model training, hyperparameter tuning and final performance evaluation, respectively.
[0047] Step 2: Construct the improved YOLOv11n model;
[0048] Step 2.1: Improve the backbone network:
[0049] Step 2.11: Introduce the Haar wavelet downsampling module (HWD) in the downsampling layer to completely replace the original strided convolution module. The Haar wavelet downsampling module (HWD) reconstructs and optimizes the downsampling convolution layer (StridedConvolution) with a strided size of 2 in the original model to preserve the high-frequency features of the obstacle edge.
[0050] The HWD module uses Haar wavelet basis functions to decompose the feature map into four sub-band components: LL, LH, HL, and HH. The resolution is halved by channel stitching. Spatial information is mapped to the channel dimension without loss, effectively protecting the high-frequency edge features of obstacles.
[0051] like Figure 3 As shown, the Haar wavelet downsampling module (HWD) performs edge-fidelity downsampling in the following way: the input image size is uniformly adjusted to 640×640. In the backbone network, this embodiment eliminates all traditional downsampling convolutional layers with a stride of 2 and replaces them all with the Haar wavelet downsampling module (HWD). When performing downsampling, this module passes the input feature map through the Discrete Wavelet Transform (DWT) operator, decomposes it in the horizontal and vertical directions using the Haar wavelet basis function, and decomposes the feature map into four sub-band components: LL (low-frequency approximation), LH (horizontal high-frequency), HL (vertical high-frequency), and HH (diagonal high-frequency) through two-dimensional discrete wavelet transform. The four sub-band components are then concatenated in the channel dimension to obtain the number of channels as the input feature map.Figure Four Enhanced feature maps with dimensions reduced by half and width and height; by stitching the above components in the channel dimension, the model achieves half the resolution while fully mapping the edge transition information and high-frequency edge information reflecting the outline of obstacles in the original space to the channel space, thereby fundamentally solving the problem of the disappearance of features of small distant target obstacles during the downsampling process.
[0052] Step 2.12: In the feature extraction stage, the original C3k2 module is replaced with the C3k2 module of the integrated variable kernel convolution module AKConv to perform geometrically adaptive feature extraction for irregular targets. The C3k2 module of the integrated variable kernel convolution module AKConv internally replaces the 3×3 convolution in the standard bottleneck structure with a variable kernel convolution (AKConv). AKConv learns the sampling offset by itself, enabling the convolution receptive field to dynamically fit the irregular geometric shape of the obstacle, thereby improving the model's feature extraction accuracy for irregular targets.
[0053] The AKConv is the core of feature extraction, containing a parallel offset generation branch used to predict the offset vector of the sampling coordinates based on the semantics of the obstacle in the current input. For example, when detecting a pedestrian with a slender shape, the convolutional sampling points are automatically stretched vertically; when the target is a vehicle with a flat shape, the sampling points are expanded horizontally. Through this geometrically adaptive deformation capability, the model achieves accurate extraction of features from irregularly shaped obstacles.
[0054] like Figure 4 As shown, AKConv achieves geometric adaptive feature extraction in the following way: the module includes a parallel offset generation branch and a sampling feature extraction branch; the offset generation branch learns a set of parameterized offsets Δpn based on the semantic content of the input features, which are used to define the irregular deformation trajectory of the sampling grid points of the convolution kernel; the sampling feature extraction branch performs bilinear interpolation sampling on the input features based on the offsets, so that the convolution receptive field can dynamically fit the vertical shape of pedestrians, the horizontal shape of vehicles, or the irregular geometric shape of road obstacles, thereby achieving accurate coverage of the features of irregular obstacles.
[0055] Step 2.13: In the deepest layer of the backbone network, the original C2PSA module is replaced with a Cross-Dimensional Interactive Attention Module (C2EMA). The C2EMA is an improvement on the original C2PSA module, replacing the PSABlock structure with an efficient multi-scale attention block (EMA_Block). The C2EMA module replaces the internal PSABlock with an efficient multi-scale attention (EMA) structure. The EMA structure establishes cross-dimensional feature dependencies by performing one-dimensional global average pooling in the X and Y directions, enabling deep spatial and channel interaction without channel reduction. By stacking multiple C2EMA modules, the model achieves progressive abstraction and enhanced focusing of the core semantic information of obstacles under complex background interference (such as road shadows and roadside buildings).
[0056] like Figure 5 As shown, the EMA_Block establishes cross-dimensional associations in the following way: the input features are divided into multiple subgroups along the channel dimension, and the spatial feature encoding branch and the channel feature exchange branch are input in parallel; the spatial feature encoding branch captures the long-range spatial dependencies of obstacles by performing one-dimensional global average pooling in the horizontal (X) and vertical (Y) directions respectively; the channel feature exchange branch realizes feature interaction through cross-dimensional vector dot product operation; the outputs of the two branches generate multi-scale attention weights through the Sigmoid activation function and multiply them element-wise with the original features; this module achieves deep fusion of spatial and channel attention without dimensionality reduction, enhancing the model's target focusing ability in complex backgrounds.
[0057] like Figure 2 As shown, the improved backbone network in this embodiment includes an 11-layer structure: a 5-layer HWD module, a 4-layer C3k2 module integrating AKConv, a 1-layer SPPF module, and a 1-layer C2EMA module. The input features sequentially pass through the 0th layer HWD module to obtain P1 / 2 features, the 1st layer HWD module to obtain P2 / 4 features, then are input to the 2nd layer C3k2 module, then to the 3rd layer HWD module to obtain P3 / 8 features, then to the 4th layer C3k2 module, then to the 5th layer HWD module to obtain P4 / 16 features, then to the 6th layer C3k2 module, then to the 7th layer HWD module to obtain P5 / 32 features, then to the 8th layer C3k2 module, then to the 9th layer SPPF module, and finally to the 10th layer C2EMA. The attention enhancement module outputs backbone features.
[0058] Step 2.2: Improvements to the Neck Network: In this embodiment, an enhanced multi-scale feature pyramid is constructed in the neck network. During feature map upsampling (enhancing), an attention-guided subpixel reconstruction module (EMA-PixelShuffle) is used instead of the linear interpolation upsampling module in the original design.
[0059] like Figure 6 As shown, the attention-guided subpixel reconstruction module (EMA-PixelShuffle) is a reconstruction and optimization of the linear interpolation upsampling layer in the original model's neck network. It processes the feature map as follows: First, the low-resolution feature map is input into the EMA attention submodule, where a cross-dimensional interaction mechanism is used to enhance the semantic weights of key obstacle regions and suppress background noise. Second, a 1×1 convolutional layer is used to expand the channel dimension of the enhanced feature map, mapping its channel number from C to C×r² (where r is the upsampling factor), thereby constructing a thick feature map with high-frequency information redundancy, providing a physical basis for subsequent spatial reorganization. Next, the expanded feature map is input into the subpixel convolution (PixelShuffle) layer, where a periodic rearrangement operator maps the high-frequency information in the channel dimension to the spatial dimension, enlarging the feature map's width and height to r times its original size. Finally, this lossless reconstruction method of "trading channels for space" solves the edge blurring problem caused by traditional interpolation upsampling, ensuring the neck remains intact. The layer output feature map and the shallow feature map extracted by the backbone network are highly aligned in terms of spatial resolution and semantic distribution.
[0060] The subpixel reconstruction module first uses its internal EMA submodule to perform semantic weight calibration on low-resolution features, suppressing redundant background noise. Then, it performs a "channel-for-space" physical recombination through subpixel convolutional layers, periodically rearranging high-dimensional channel features to the spatial dimension. This approach avoids image blurring and texture loss caused by traditional interpolation methods, ensuring clear restoration of edge details when fusing features at P3, P4, and P5 scales, and significantly optimizing the alignment of feature flows in spatial resolution.
[0061] Simultaneously, the feature processing unit in the Neck layer introduces the AKConv operator to ensure that the fused multi-scale features still possess geometrically adaptive shape capture capabilities when handling obstacles at different distances and angles. This design allows the fused features to retain rich spatial details and high-level semantics.
[0062] The improved Neck network comprises a 12-layer structure, as follows: Layer 11 is an EMA-PixelShuffle module that upsamples the P5 features; Layer 12 is a Concat module that concatenates the output of Layer 11 with the P4 features; Layer 13 is a C3k2 module integrating AKConv; Layer 14 is an EMA-PixelShuffle module that upsamples the output of Layer 13; Layer 15 is a Concat module that concatenates the output of Layer 14 with the P3 features; Layer 16 is a C3k2 module integrating AKConv to output small target detection features; Layer 17 is a module that performs 3×3 convolution downsampling on the output of Layer 16; Layer 18 is a Concat module that concatenates it with the output features of Layer 13; Layer 19 is a C3k2 module integrating AKConv to output medium target detection features; Layer 20 is a module that performs 3×3 convolution downsampling on the output of Layer 19; Layer 21... The Concat module, which concatenates the output features of layer 10, and the C3k2 module, which integrates AKConv in layer 22, output large target detection features. Finally, the three-way multi-scale fusion features output from layers 16, 19, and 22 are input into the head detection module to complete road obstacle detection and localization.
[0063] Step 2.3: Establish the Detection Head: The improved neck network outputs enhanced feature maps at three scales: P3, P4, and P5, corresponding to the detection tasks of small targets (e.g., distant obstacles), medium targets (e.g., ordinary vehicles), and large targets (e.g., nearby large trucks), respectively. These feature maps are fed into the decoupled detection head of standard YOLOv11. The detection head employs a parallel branching strategy: one branch regresses the normalized bounding box center coordinates and width / height offsets (cx, cy, w, h) through convolution operations; the other branch calculates the category confidence probability distribution of obstacles. Finally, overlapping boxes are removed using non-maximum suppression (NMS) or similar post-processing mechanisms, outputting the precise location and category label of all obstacles in the road scene.
[0064] Step 3: Iteratively train the improved model using the training set to optimize the model parameters and obtain a trained obstacle detection model; specifically:
[0065] The prepared training set is then fed into the improved model constructed above for training.
[0066] Training Environment and Hyperparameters: This experiment was conducted on a server equipped with an NVIDIA GeForce RTX 4090 GPU. The training core hyperparameter settings are shown in the table below:
[0067] Table 1
[0068]
[0069] Training process: The model is initialized using YOLOv11n weights pre-trained on the COCO dataset to accelerate convergence. A cosine annealing learning rate scheduler is used. During training, metrics such as mean accuracy (mAP@0.5) on the validation set are monitored, and an early-stopping strategy is employed to prevent overfitting, saving the model weights that perform best on the validation set.
[0070] To further verify the effectiveness of the algorithm of this invention, this section compares it with other benchmark models. The experiments were conducted under the same parameters and environmental configuration, and the results are shown in Table 2. All data are based on the test set verification results. The improved model proposed in this study achieves optimal overall detection performance while maintaining extremely high parameter efficiency. Specifically, due to the introduction of the HWD edge fidelity module and the AKConv geometric adaptive operator, the model maintains a lightweight design and exhibits particularly outstanding recognition accuracy for small obstacles at long distances and irregularly shaped targets.
[0071] Table 2
[0072]
[0073] While improving accuracy by 2.7% and mAP50 by 2.5%, the number of model parameters was reduced by approximately 29.8% due to the effective compression of redundant parameters by AKConv. It comprehensively surpasses the original YOLO series models, verifying the effectiveness of the improved scheme in multi-scale feature representation and complex background recognition, and achieving high-precision and robust detection of road obstacles.
[0074] Step 4: Use the trained model to identify and locate obstacles in the road environment image to be detected, and output the bounding box and category information of the target object.
[0075] The trained optimal model is exported to a deployment format (such as ONNX or TensorRT) and integrated into an intelligent assisted driving system based on in-vehicle vision sensors and edge computing devices (such as the NVIDIA Orin series). The system continuously acquires images of the road scene ahead of the driving path, and the model performs real-time inference on each frame of the image, automatically labeling the location and category of all pedestrians, vehicles, roadblocks, and other obstacles. It can also trigger warning signals based on a set safe distance threshold, realizing automated, intelligent, and efficient online detection of road obstacles.
[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A target detection method integrating edge information protection and geometric perception feature extraction, characterized in that, The steps include: improving the YOLOv11n model to construct an enhanced road obstacle detection model; the improvements include: (1) In the downsampling layer of the backbone network, the Haar wavelet downsampling module (HWD) is introduced to replace the traditional convolutional downsampling; (2) In the feature extraction units of the backbone network and the neck network, the C3k2 module of the integrated variable kernel convolution module AKConv is used to perform geometric adaptive feature extraction on the irregular target; (3) The cross-dimensional interactive attention module C2EMA is used to reconstruct the feature representation in the deep layer of the backbone network to enhance the model’s ability to suppress background and perceive context under complex road conditions; (4) In the upsampling layer of the Neck network, the attention-guided subpixel reconstruction module EMA-PixelShuffle is used to replace the original traditional upsampling module to achieve lossless restoration of multi-level features.
2. The target detection method integrating edge information protection and geometric perception feature extraction according to claim 1, characterized in that: The Haar wavelet downsampling module (HWD) described in step (1) is set in the starting layer of the backbone network and replaces the downsampling layer before each feature extraction layer C3k2 in the backbone network. The HWD module uses the Haar wavelet basis function to decompose the feature map into four sub-band components: LL, LH, HL, and HH. The resolution is halved by channel splicing. The spatial information is mapped to the channel dimension without loss.
3. The target detection method integrating edge information protection and geometric perception feature extraction according to claim 1 or 2, characterized in that: The Haar wavelet downsampling module HWD performs edge-fidelity downsampling by performing the following steps: passing the input feature map X through the Discrete Wavelet Transform (DWT) operator, and decomposing it in the horizontal and vertical directions using the Haar wavelet basis function to generate four complementary sub-band components, namely the low-frequency approximation component (LL), the horizontal high-frequency component (LH), the vertical high-frequency component (HL), and the diagonal high-frequency component (HH); and concatenating the four sub-band components in the channel dimension to obtain an enhanced feature map with four times the number of channels of the input feature map and half the width and height.
4. The target detection method integrating edge information protection and geometric perception feature extraction according to claim 1, characterized in that: The C3k2 module integrating AKConv described in step (2) replaces the 3×3 convolution in the standard bottleneck structure with a variable kernel convolution AKConv; it generates the deformation deviation value of the branch prediction sampling point coordinates through offset, and induces the convolution kernel to produce geometric deformation to fit the contour of the irregular target.
5. A target detection method integrating edge information protection and geometric perception feature extraction according to claim 1 or 4, characterized in that: The AKConv method for implementing geometric adaptive feature extraction includes the following steps: the module contains a parallel offset generation branch and a sampling feature extraction branch. The offset generation branch learns a set of parameterized offsets Δpn based on the semantic content of the input features, which are used to define the irregular deformation trajectory of the sampling grid points of the convolution kernel. The sampling feature extraction branch performs bilinear interpolation sampling on the input features based on the offsets, so that the convolution receptive field can dynamically fit the vertical shape of pedestrians, the horizontal shape of vehicles, or the irregular geometric shape of road obstacles.
6. The target detection method integrating edge information protection and geometric perception feature extraction according to claim 1, characterized in that: The cross-dimensional interactive attention module C2EMA described in step (3) replaces the internal PSABlock with an efficient multi-scale attention EMA structure. It adopts the EMA mechanism to establish cross-dimensional feature dependencies by performing one-dimensional global average pooling in the horizontal and vertical directions, so that the model can achieve deep interaction between space and channel without channel reduction.
7. A target detection method integrating edge information protection and geometric perception feature extraction according to claim 1 or 6, characterized in that: The operation method of establishing cross-dimensional association by the cross-dimensional interactive attention module C2EMA is as follows: the input features are divided into multiple subgroups along the channel dimension, and the input space feature encoding branch and the channel feature exchange branch are carried out in parallel. The spatial feature encoding branch captures the long-range spatial dependencies of obstacles by performing one-dimensional global average pooling in the horizontal (X) and vertical (Y) directions respectively. The channel feature exchange branch achieves feature interaction through cross-dimensional vector dot product operations; the outputs of the two branches generate multi-scale attention weights through the Sigmoid activation function and are multiplied element-wise with the original features.
8. The target detection method integrating edge information protection and geometric perception feature extraction according to claim 1, characterized in that: The attention-guided subpixel reconstruction module described in step (4) first uses the internal EMA submodule to perform semantic weight calibration on low-resolution features and suppress redundant background noise. Subsequently, a 1×1 convolutional layer is used to expand the channel dimension; finally, a subpixel convolutional layer is used to perform a physical recombination of "channel for space", periodically rearranging the high-dimensional channel features to the spatial dimension; it uses subpixel convolution to recombine high-dimensional channel features into high-resolution spatial features, and with multi-scale attention guidance, it ensures the boundary clarity and pixel-level alignment during the feature fusion process.
9. A target detection method integrating edge information protection and geometric perception feature extraction according to claim 1 or 8, characterized in that: The operation method of the sub-pixel reconstruction module in step (4) for processing feature maps includes the following steps: First, input the low-resolution feature map into the EMA attention submodule, and use the cross-dimensional interaction mechanism to strengthen the semantic weight of the key regions of the obstacle and suppress background noise; Secondly, the enhanced feature map is input into the sub-pixel convolution PixelShuffle layer. The high-frequency information in the channel dimension is mapped to the spatial dimension through the periodic rearrangement operator, and the width and height of the feature map are enlarged to r times the original size, where r is the upsampling factor. Finally, a lossless restoration method of "trading channels for space" is performed.
10. The target detection method integrating edge information protection and geometric perception feature extraction according to claim 1, characterized in that: The method further includes the steps of: Before improving the YOLOv11n model, the following steps were taken: preprocessing and data augmentation were performed on the pre-acquired road environment obstacle image dataset, and the processed dataset was divided into training set, validation set and test set; And the setup for improving the YOLOv11n model: using the training set to iteratively train the improved model, optimize the model parameters, and obtain a trained obstacle detection model; The trained model is used to identify and locate obstacles in the road environment image to be detected, and outputs the bounding box and category information of the target object.