Road target detection method based on geometric density compensation and spectrum refinement
Patent Information
- Application Number
- CN202611428152.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-10-16
AI Technical Summary
最近的频率感知检测研究表明，保留信息丰富的高频细节有利于小物体的表示，而对频率分量的不恰当增强也可能引入背景干扰
[0014]本发明的优点和积极效果是：
Smart Images
Figure CN122929520A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and target detection technology, and in particular to a road target detection method based on geometric density compensation and spectral refinement. Background Technology
[0002] Object detection is a fundamental perception task in intelligent transportation systems and autonomous driving, providing essential environmental information for scene understanding, decision-making, trajectory planning, and driving safety. In complex traffic environments, perception systems need to accurately identify and locate various road users, including vehicles, pedestrians, cyclists, and other traffic-related objects, while achieving a good balance between detection accuracy and computational efficiency. With the rapid development of deep learning, single-stage object detectors have attracted much attention due to their efficient end-to-end prediction paradigm. Among them, the YOLO series of algorithms are widely used in real-time perception due to their compact architecture and efficient inference process. However, robust road object detection still faces challenges due to significant differences in traffic objects in terms of scale, appearance, spatial distribution, occlusion degree, and visibility distance. Recent research on vehicle-pedestrian detection and multi-scale pedestrian detection shows that complex backgrounds, scale variations, weak visual cues, and occlusion remain the main factors leading to missed detections and false detections.
[0003] Compared to large and visually prominent vehicles, distant vehicles, pedestrians, and cyclists typically occupy smaller spatial areas and contain weaker discriminative information. During hierarchical feature extraction, repetitive feature transformations and spatial downsampling can further weaken their limited structural responses, while complex backgrounds and mutual occlusion increase the ambiguity between foreground objects and their surroundings. Recent studies have shown that small object size, weak semantic representation, and insufficient structural information significantly affect the detection of small and occluded objects. Furthermore, effective multi-scale representation is particularly important for pedestrian detection under crowded and occluded conditions, as information from a single feature level is often insufficient to provide a robust object description. Therefore, improving the accuracy of road object detection requires detectors to preserve informative local structures, utilize complementary representations at different scales, and suppress irrelevant responses introduced by complex backgrounds.
[0004] Despite significant advancements in modern detection frameworks, several limitations remain when directly applied to complex traffic scenarios. First, traditional hierarchical feature extraction often relies on relatively fixed spatial transformations, which may be insufficient to accommodate the substantial differences in geometry, scale, and spatial distribution of road objects. Second, feature representations generated at different network levels differ in spatial resolution, semantic abstraction, and response characteristics. Therefore, direct aggregation may introduce redundant information or weaken useful spatial details, necessitating appropriate contribution calibration for heterogeneous features. Third, fused representations inevitably include not only discriminative object structures but also high-frequency responses generated by road textures, vegetation, shadows, guardrails, and other complex background patterns. Recent frequency-aware detection research indicates that preserving information-rich high-frequency details is beneficial for representing small objects, while inappropriate enhancement of frequency components can introduce background interference. These observations suggest that feature extraction, cross-scale fusion, and post-fusion feature refinement should be performed in a coordinated manner. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose a road target detection method based on geometric density compensation and spectral refinement, which can significantly improve the detection accuracy of road targets in complex traffic scenarios with lower additional computational overhead.
[0006] The technical problem solved by this invention is achieved through the following technical solution: The road target detection method based on geometric density compensation and spectral refinement includes the following steps: Step 1: Acquire the road scene image to be detected and input it into the backbone network; Step 2: Extract hierarchical features from the input image through the backbone network, and introduce a geometric density compensation module after the P4 feature layer of the backbone network to perform geometric density compensation processing on the intermediate feature representation. Step 3: Introduce a residual weighted calibration fusion module at the cross-scale fusion node of the backbone network and the neck network to perform adaptive contribution weighted fusion of high-level semantic features and low-level spatial details; Step 4: After the residual weighted calibration fusion module, a spectrum balance refinement module is introduced to perform frequency-aware feature adjustment on the fused features; Step 5: Input the processed features into the detection head to perform target classification and bounding box regression, and output the detection results of road targets.
[0007] Furthermore, the processing procedure of the geometric density compensation module in step 2 is as follows: Step 2.1, Input Features X Channel alignment is performed using 1×1 convolution, resulting in... X' : Here, Align represents the channel alignment operation.
[0008] Step 2.2, for X' The neighborhood reference mean is obtained by applying 3×3 average pooling. m : Step 2.3: Extract statistics based on μ, including statistics describing local mutations. D Reflecting neighborhood structural fluctuations V Measure the overall activation intensity E : Step 2.4: Normalize the three statistics along the spatial dimension and then concatenate them: in,( A ) represents any statistical clue, Mean h,w and Std h,w Let these represent the mean and standard deviation along the spatial dimension, respectively. e ) is a constant used to avoid numerical instability. N(A) This represents the standardized result of statistical clue A along the spatial dimension. E d express N(D) , N(V) and N(E) Geometric density description features obtained by splicing along the channel dimension; Step 2.5: Map the receptive field requirement graph Q to a bounded receptive field using a 1×1 convolution. in, ψ (·) is implemented using 1×1 convolution. s q This indicates the maximum adjustment range of the demand map; Step 2.6: Generate compact receptive field basis F using two parallel depthwise separable convolutions. s and extended sensory experience F l ; And calculate the output of GDCM: in, c g Control the overall compensation range, and Q Provides location-related adjustments.
[0009] Moreover, the compact sensing field base F sand extended sensory experience F l The calculation method is as follows: Where DWConv represents depthwise separable convolution, k The kernel size is [size]. d Let be the cavitation rate, and d l >1.
[0010] Furthermore, the processing procedure of the residual weighted calibration fusion module in step 3 is as follows: Step 3.1: Obtain the upsampled high-level features X h and the corresponding lateral P4 features X l Preserve the original splicing and fusion path: Step 3.2, X h and X l Projecting the images onto the compact public hidden space via channel projection: in, F h (·)and F l (·) indicates a channel projection operation. C r It is the number of channels in the hidden space; Step 3.3: Introduce learnable coefficients and The fusion weights of high-level and low-level branches are calculated using Softmax normalization. oh h and oh l ; Step 3.4: Calculate the weighted fusion features: For F w Apply a lightweight local refinement operation ψ(·).
[0011] Furthermore, the fusion weight of the high-level branches and low-level branches oh h and oh l The calculation method is as follows: in, e To avoid dividing by zero small constants.
[0012] Furthermore, the processing procedure of the spectrum balance refinement module in step 4 is as follows: Step 4.1: Obtain the input feature X, where X is the fusion feature output by the residual weighted calibration fusion module. Use a local averaging filter to perform three-band decomposition to obtain the low-frequency component F. L , intermediate frequency component F M and high frequency component F H ; Step 4.2: Estimate the structural consistency between intermediate frequency (IF) and high frequency (HF) components through inter-channel relationships, and calculate the reliability map R. s ; Step 4.3: The modulation amplitude g of the intermediate frequency component and the high frequency component is controlled by a learnable factor. m and g h ; Step 4.4: Calculate candidate refined residuals: in, ΔF Indicates candidate refined residuals; R s A structural reliability diagram showing the relationship between intermediate frequency (IF) and high frequency (HF) components is used to characterize the degree to which the high-frequency response is supported by the IF structure; 2 R s -1 maps the reliability map to a modulation interval centered at zero, which is used to determine the modulation direction of high-frequency components; Step 4.5: Calculate the modulation coefficient using local residual energy constraints. c m and γ h ; Step 4.6, Output Y SBRM : .
[0013] in, R loc The local root mean square response of the input features is used to characterize the local feature energy at different spatial locations. R glob Represents the global root mean square response of the input features, used to provide a reference for the overall feature energy; R Δ This represents the root mean square response of the candidate refined residual, used to measure the energy intensity of the candidate residual; R ref This represents the reference energy determined by both the local and global characteristic energies.
[0014] The advantages and positive effects of this invention are: 1. This invention proposes a Geometry-Density Compensation Module (GDCM) to enhance intermediate feature representations. GDCM considers geometric features and feature density variations to improve the adaptability of the backbone network to road targets with different scales, structures, and spatial distributions.
[0015] 2. This invention designs a Residual Weighted Calibration Fusion Module (RWCFM) to improve cross-scale feature interactions. RWCFM adaptively calibrates the relative contributions of heterogeneous feature representations using residual information, thereby enhancing complementary semantic and spatial information while reducing redundant or inconsistent responses.
[0016] 3. This invention proposes a Spectral-Balance Refinement Module (SBRM) for feature refinement after fusion. The SBRM performs frequency-aware feature adjustment to preserve structurally rich details and suppress unsupported high-frequency interference, thereby improving the quality of the fused representation. Attached Figure Description
[0017] Figure 1 This is a diagram showing the overall network architecture of the road target detection method based on geometric density compensation and spectral refinement of the present invention. Figure 2 This is a schematic diagram of the geometric density compensation module (GDCM) in this invention; Figure 3 This is a schematic diagram of the residual weighted calibration fusion module (RWCFM) in this invention; Figure 4 This is a schematic diagram of the Spectrum Balance Refinement Module (SBRM) in this invention. Detailed Implementation
[0018] The present invention will be further described in detail below with reference to the accompanying drawings.
[0019] Example 1 like Figure 1 The diagram shows the overall architecture of this invention. YOLOv12 was chosen as the baseline detector because of its efficient hierarchical representation and multi-scale detection capabilities. However, complex traffic scenarios still face three consecutive challenges: insufficient adaptability of backbone features to scale and structural changes, imbalance of semantic-detail interactions during cross-scale fusion, and interference between useful structural details and high-frequency responses caused by the background after feature aggregation.
[0020] Therefore, this invention sequentially introduces GDCM, RWCFM, and SBRM. GDCM, located after the P4 backbone feature, is used to dynamically compensate for the effective context based on local geometry; RWCFM, located at the first P5→P4 fusion node, performs weighted calibration on the two cross-scale feature flows while preserving the original Concat path; SBRM further evaluates the reliability of the fused frequency components and performs bidirectional adjustments. These three modules correspond to three different stages: feature generation, cross-scale fusion, and post-fusion optimization.
[0021] The road target detection method based on geometric density compensation and spectral refinement includes the following steps: Step 1: Obtain the road scene image to be detected and input it into the backbone network.
[0022] Step 2: Extract hierarchical features from the input image through the backbone network, and introduce a geometric density compensation module after the P4 feature layer of the backbone network to perform geometric density compensation processing on the intermediate feature representation.
[0023] like Figure 2 As shown, state-of-the-art vehicle detection algorithms based on YOLO have demonstrated that adaptive spatial sampling and receptive field modeling can effectively handle geometric and scale variations in complex traffic scenarios. However, existing strategies primarily focus on improving the capture of contextual information, paying less attention to whether different spatial locations require the same degree of receptive field expansion. To address this issue, this invention proposes a geometric density compensation module (GDCM). Instead of directly aggregating multiple receptive field features, GDCM estimates the contextual requirements of different spatial locations and selectively injects incremental information provided by the expanded receptive field.
[0024] Step 2.1, Input Features X ∈R B×Cin×H×W Channel alignment is performed using 1×1 convolution, resulting in... X' :
[0025] Here, Align represents the channel alignment operation.
[0026] Step 2.2, for X' The neighborhood reference mean is obtained by applying 3×3 average pooling. m :
[0027] Step 2.3: Extract statistics based on μ, including statistics describing local mutations. D Reflecting neighborhood structural fluctuations V Measure the overall activation intensity E :
[0028] Wherein, the geometric density is determined by local mutations. D Neighborhood structural fluctuations V and overall activation intensity E A common representation is used to describe the structural response state at different spatial locations, rather than the number of physical object instances.
[0029] Step 2.4: Normalize the three statistics along the spatial dimension and then concatenate them:
[0030] in,( A ) represents any statistical clue, Mean h,w and Std h,w Let these represent the mean and standard deviation along the spatial dimension, respectively. e ) is a constant used to avoid numerical instability. N(A) This represents the standardized result of statistical clue A along the spatial dimension. E d express N(D) , N(V) and N(E) Geometric density description features obtained by splicing along the channel dimension; Step 2.5: Map the receptive field requirement graph Q to a bounded receptive field using a 1×1 convolution.
[0031] in, ψ (·) is implemented using 1×1 convolution. s q This indicates the maximum adjustment range of the demand graph. (This invention...) s q Set to 0.25; therefore, Q ∈[0.25,0.75]. Larger values indicate a stronger need for contextual information, while smaller values emphasize the preservation of local details.
[0032] Step 2.6: Generate compact receptive field basis F using two parallel depthwise separable convolutions. s and extended sensory experience F l ; And calculate the output of GDCM:
[0033] Where DWConv represents depthwise separable convolution, k The kernel size is [size]. d Let be the cavitation rate, and d l 1. Compact and comfortable feel of the wilderness Fs Generated by standard depthwise separable convolution, primarily preserving vehicle boundaries, local textures, and precise positional information. (Extended sensory field base) F l Generated by depth-separable dilated convolution, it expands the spatial coverage without significantly increasing the number of parameters and is used to extract contextual information around occluded and scale-changing objects.
[0034] F s and F l Instead of direct addition or concatenation, their differences are treated as incremental contextual information provided by the expanded receptive field. Therefore, the final GDCM output can be expressed as:
[0035] in, c g Control the overall compensation range, and Q Location-dependent adjustments are provided. In the implementation, the initial response of the projection branch is close to zero to prevent the compensation path from excessively perturbing the original feature distribution in the early stages of training.
[0036] Step 3: Introduce a residual weighted calibration fusion module at the cross-scale fusion node of the backbone network and the neck network to perform adaptive contribution weighted fusion of high-level semantic features and low-level spatial details.
[0037] like Figure 3 As shown, the YOLO-based autonomous driving detector demonstrates that dynamic multi-scale feature interaction can improve the utilization of complementary information across different network layers. However, heterogeneous features exhibit significant differences in semantic abstraction and spatial details, and directly modifying the fusion topology cannot explicitly guarantee the balanced contribution of each fusion node. Therefore, this invention proposes a residual weighted calibration fusion module. This module preserves the original cascaded path and learns an additional lightweight residual branch to calibrate high-level semantic information and low-level spatial details.
[0038] Step 3.1: Obtain the upsampled high-level features X h and the corresponding lateral P4 features X l Preserve the original splicing and fusion path:
[0039] Step 3.2, X h and X l Projecting the images onto the compact public hidden space via channel projection:
[0040] in, F h (·)and F l (·) indicates a channel projection operation. C r It is the number of channels in the hidden space; in the final implementation, the hidden channel reduction rate is set to 4.
[0041] Step 3.3: Introduce learnable coefficients and The fusion weights of high-level and low-level branches are calculated using Softmax normalization. oh h and oh l ;
[0042] in, oh h and oh l These represent the fusion weights of the high-level semantic branch and the low-level detail branch, respectively. e To avoid dividing by zero small constants.
[0043] Step 3.4: To further enhance the local consistency of weighted fusion features, RWCFM... F w A lightweight local thinning operation is applied. This operation includes depthwise separable convolutions and 1×1 convolutions, supplementing local spatial interaction information while controlling the number of parameters. A lightweight depthwise separable operator is then used. R (·) is used to enhance local spatial interaction. After channel reprojection, the calibrated information is injected into the original stitching features through residual connections:
[0044] in, ψ (·) indicates a lightweight local refinement operation. This indicates an output projection operation. c It is a learnable residual scaling factor. Residual coefficients c Initializing to zero makes this module equivalent to the original stitching operation at the start of optimization. Therefore, RWCFM can progressively learn useful cross-scale calibration information without discarding the original feature fusion path. Compared to direct weighted fusion, this design separates information preservation from information calibration: the stitching path preserves complete multi-scale information, while the residual branch is only responsible for correcting semantic-detail imbalance.
[0045] Step 4: After the residual weighted calibration fusion module, a spectrum balance refinement module is introduced to perform frequency-aware feature adjustment on the fused features.
[0046] like Figure 4 As shown in the YOLO-based vehicle perception research, Fourier domain modeling can enhance weak structural representations and improve feature discrimination when vehicle appearance is degraded or blurred. However, frequency enhancement itself cannot distinguish whether the high-frequency response originates from meaningful object structures or irrelevant background patterns, which may lead to the simultaneous amplification of useful details and interference. To address this issue, this invention introduces a spectral balancing refinement module after RWCFM. Instead of simply enhancing high-frequency information, SBRM evaluates whether the high-frequency response is structurally supported by adjacent intermediate frequency components before determining the modulation direction.
[0047] Step 4.1: Obtain input features X ,in X To obtain the fusion characteristics output by the residual weighted calibration fusion module, a local averaging filter is used for three-band decomposition to obtain the low-frequency component F. L , intermediate frequency component F M and high frequency component F H ;
[0048] F L , F M and F H These correspond to the low-frequency context, intermediate structural information, and fine high-frequency details, respectively, while simultaneously satisfying... X = F L + F M + F H .
[0049] Step 4.2: Estimate the structural consistency between intermediate frequency (IF) and high frequency (HF) components through inter-channel relationships, and calculate the reliability map R. s ;
[0050] in, R M and R H These represent the root-mean-square amplitudes between channels for the intermediate frequency and high frequency components, respectively. In implementation, k =8.0, t =0.15, r=0.5. Therefore, a high reliability score can only be obtained when a strong high-frequency response is accompanied by consistent structural support.
[0051] Step 4.3: The modulation amplitude g of the intermediate frequency component and the high frequency component is controlled by a learnable factor. m and g h ;
[0052] Step 4.4: Calculate candidate refined residuals:
[0053] Reliable high-frequency structures can be preserved or enhanced, while unsupported high-frequency responses can be actively suppressed.
[0054] Step 4.5: Calculate the modulation coefficient using local residual energy constraints. c m and γ h And output Y SBRM :
[0055] in, R loc , R glob and R Δ These represent the local input root mean square (RMS), the global input root mean square (RMS), and the candidate residual root mean square (RMS), respectively. c m and γ h All values are initialized to zero. Therefore, SBRM starts with the identity mapping and gradually learns the appropriate spectral correction. Furthermore, the frequency analysis branch is constructed from a separated reference of the input features, preventing the refinement residuals from introducing additional direct gradient paths into the preceding RWCFM and GDCM modules.
[0056] Step 5: Input the processed features into the detection head to perform target classification and bounding box regression, and output the detection results of road targets.
[0057] Example 2 Based on the aforementioned road target detection method based on geometric density compensation and spectral refinement, experiments were conducted to verify the effectiveness of the invention. All experiments were performed in an Ubuntu 20.04 environment, using an Intel Core i9-14900k processor, an NVIDIA GeForce RTX 4090 (24 GB) GPU, Python 3.13, PyTorch 2.7.1, and CUDA 12.8. During training, images underwent mosaic data augmentation, and the image size was consistently set to 640 × 640 pixels. To ensure comparability, all comparison methods were trained for 300 epochs. The initial learning rate was 0.01, and the AdamW optimizer was used. The best model from each iteration was saved.
[0058] This invention uses the KITTI dataset as its base dataset, whose images are all collected from real urban, rural, and highway scenes. This invention selects images from this dataset for training and testing, dividing them in an 8:2 ratio, resulting in a training set containing 6028 images and a testing set containing 1453 images.
[0059] Using YOLOv12n as the base network, and with the same training conditions, ablation verification was performed by adding the proposed GDCM, RWCFM, and SBRM. The results are shown in Table 1.
[0060] Table 1 Ablation Experiment Results
[0061] The baseline model achieved Precision, Recall, mAP@50, and mAP@50-95 of 88.6%, 80.0%, 88.0%, and 64.0%, respectively. Adding GDCM alone improved Recall to 81.4%, and mAP@50 and mAP@50-95 to 89.2% and 65.1%, respectively. Adding RWCFM alone improved Recall to 81.8% and mAP@50-95 to 65.0%. Adding SBRM alone improved Precision to 90.0%, indicating that different modules improved detection performance in terms of intermediate feature representation, cross-scale fusion, and fusion feature refinement. In some dual-module combinations, slight fluctuations were observed in certain metrics. For example, combining SBRM with only GDCM or RWCFM improved Precision, while some Recall or mAP metrics slightly decreased. The reason for this is that the spectral refinement process suppresses uncertain high-frequency responses lacking structural support. Before the preceding features have undergone simultaneous geometric compensation and cross-scale calibration, the responses of some weak targets may be conservatively adjusted. When the GDCM, RWCFM, and SBRM modules are applied simultaneously, Precision, Recall, mAP@50, and mAP@50-95 reach 91.4%, 81.4%, 89.5%, and 65.7%, respectively. Compared to the baseline network, these represent improvements of 2.8, 1.4, 1.5, and 1.7 percentage points, respectively. Simultaneously, the number of parameters in the complete model increases from 2.51 M to 2.61 M, and GFLOPs increase from 5.8 to 6.2. These experimental results demonstrate that sequentially performing geometric density compensation, cross-scale contribution calibration, and spectral balancing refinement can achieve better overall target detection performance with relatively low additional computational complexity.
[0062] To verify the effectiveness of the proposed algorithm in vehicle detection, a comparative experiment was conducted on the KITTI dataset with seven YOLO series algorithms, namely YOLOv5, YOLOv6, YOLOv8, YOLOv9, YOLOv10, YOLOv11 and YOLOv12. The experimental results are shown in Table 2.
[0063] Table 2 Comparison of experimental results on the KITTI dataset
[0064] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.
Claims
1. A road target detection method based on geometric density compensation and spectral refinement, characterized in that: Includes the following steps: Step 1: Acquire the road scene image to be detected and input it into the backbone network; Step 2: Extract hierarchical features from the input image through the backbone network, and introduce a geometric density compensation module after the P4 feature layer of the backbone network to perform geometric density compensation processing on the intermediate feature representation. Step 3: Introduce a residual weighted calibration fusion module at the cross-scale fusion node of the backbone network and the neck network to perform adaptive contribution weighted fusion of high-level semantic features and low-level spatial details; Step 4: After the residual weighted calibration fusion module, a spectrum balance refinement module is introduced to perform frequency-aware feature adjustment on the fused features; Step 5: Input the processed features into the detection head to perform target classification and bounding box regression, and output the detection results of road targets.
2. The road target detection method based on geometric density compensation and spectral refinement according to claim 1, characterized in that: The geometric density compensation module in step 2 processes the intermediate features according to the following steps: Step 2.1, Input Features X Channel alignment is performed using 1×1 convolution, resulting in... X' : Here, Align represents the channel alignment operation. Step 2.2, for X' The neighborhood reference mean is obtained by applying 3×3 average pooling. μ : Step 2.3: Extract statistics based on μ, including statistics describing local mutations. D Reflecting neighborhood structural fluctuations V Measure the overall activation intensity E : Step 2.4: Normalize the three statistics along the spatial dimension and then concatenate them: in,( A ) represents any statistical clue, Mean h,w and Std h,w Let these represent the mean and standard deviation along the spatial dimension, respectively. ε ) is a constant used to avoid numerical instability. N(A) This represents the standardized result of statistical clue A along the spatial dimension. E d express N(D) , N(V) and N(E) Geometric density description features obtained by splicing along the channel dimension; Step 2.5: Map the receptive field requirement graph Q to a bounded receptive field using a 1×1 convolution. in, ψ (·) is implemented using 1×1 convolution. s q This indicates the maximum adjustment range of the demand map; Step 2.6: Generate compact receptive field basis F using two parallel depthwise separable convolutions. s and extended sensory experience F l ; And calculate the output of GDCM: in, γ g Control the overall compensation range, and Q Provides location-related adjustments.
3. The road target detection method based on geometric density compensation and spectral refinement according to claim 2, characterized in that: The compact sensing field base F s and extended sensory experience F l The calculation method is as follows: Where DWConv represents depthwise separable convolution, k The kernel size is [size]. d Let be the cavitation rate, and d l >1.
4. The road target detection method based on geometric density compensation and spectral refinement according to claim 1, characterized in that: The processing procedure of the residual weighted calibration fusion module in step 3 is as follows: Step 3.1: Obtain the upsampled high-level features X h and the corresponding lateral P4 features X l Preserve the original splicing and fusion path: Step 3.2, X h and X l Projecting the images onto the compact public hidden space via channel projection: in, Φ h (·)and Φ l (·) indicates a channel projection operation. C r It is the number of channels in the hidden space; Step 3.3: Introduce learnable coefficients and The fusion weights of high-level and low-level branches are calculated using Softmax normalization. ω h and ω l ; Step 3.4: Calculate the weighted fusion features: For F w Apply a lightweight local refinement operation ψ(·).
5. The road target detection method based on geometric density compensation and spectral refinement according to claim 4, characterized in that: The fusion weight of the high-level branch and the low-level branch ω h and ω l The calculation method is as follows: in, ε To avoid dividing by zero small constants.
6. The road target detection method based on geometric density compensation and spectral refinement according to claim 1, characterized in that: The processing procedure of the spectrum balancing and refinement module in step 4 is as follows: Step 4.1: Obtain input features X ,in X To obtain the fusion characteristics output by the residual weighted calibration fusion module, a local averaging filter is used for three-band decomposition to obtain the low-frequency component F. L , intermediate frequency component F M and high frequency component F H ; Step 4.2: Estimate the structural consistency between intermediate frequency (IF) and high frequency (HF) components through inter-channel relationships, and calculate the reliability map R. s ; Step 4.3: The modulation amplitude g of the intermediate frequency component and the high frequency component is controlled by a learnable factor. m and g h ; Step 4.4: Calculate candidate refined residuals: in, ΔF Indicates candidate refined residuals; R s A structural reliability diagram showing the relationship between intermediate frequency (IF) and high frequency (HF) components is used to characterize the degree to which the high-frequency response is supported by the IF structure; 2 R s -1 maps the reliability map to a modulation interval centered at zero, which is used to determine the modulation direction of high-frequency components; Step 4.5: Calculate the modulation coefficient using local residual energy constraints. γ m and γ h ; Step 4.6, Output Y SBRM : in, R loc The local root mean square response of the input features is used to characterize the local feature energy at different spatial locations. R glob Represents the global root mean square response of the input features, used to provide a reference for the overall feature energy; R Δ This represents the root mean square response of the candidate refined residual, used to measure the energy intensity of the candidate residual; R ref This represents the reference energy determined by both the local and global characteristic energies.