A target detection method, system, and device based on multi-domain perception enhancement
By introducing the Local Prior Enhancement Module (LPEM) and the Dynamic Spectral Attention Enhancement Module (DSAEM), combined with the Wise-ShapeIoU loss function, the accuracy and stability issues of target detection in complex traffic scenarios are addressed, improving the accuracy and robustness of small target and bounding box localization.
Patent Information
- Application Number
- CN202610782725.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-02
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-06-02
AI Technical Summary
Existing target detection methods struggle to reliably extract edge, texture, and local structural features of targets in complex traffic scenarios. Furthermore, the bounding box regression loss function makes insufficient contributions to the regression of samples of varying quality, leading to missed detections, inaccurate localization, and false detections of small targets.
The feature extraction capability is enhanced by employing the Local Prior Enhancement Module (LPEM) and the Dynamic Spectral Attention Enhancement Module (DSAEM), and the bounding box regression is optimized by combining the Wise-ShapeIoU loss function, thereby improving the network's representation and localization accuracy for complex backgrounds and multi-scale targets.
It improves the accuracy and robustness of target detection in complex traffic scenarios, enhances the feature extraction capability of small targets and targets with blurred boundaries, and improves the accuracy of bounding box localization and training stability.
Smart Images

Figure CN122597772B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and image processing, and in particular to a target detection method, system and device based on multi-domain perception enhancement, which can improve the accuracy of target detection in complex traffic scenarios. Background Technology
[0002] With the development of intelligent driving, assisted driving, and intelligent road traffic perception technologies, the real-time perception capability of vehicles of their surrounding traffic environment has become a crucial foundation for ensuring driving safety. Object detection, as a key task in environmental perception systems, is primarily used to identify traffic participants such as vehicles, pedestrians, and cyclists in traffic scenes, and output object category and location information to provide data support for subsequent trajectory prediction, path planning, and driving decisions. However, in real-world road traffic scenarios, object detection tasks are often affected by factors such as complex backgrounds, varying object scales, small objects at long distances, occlusion and truncation, edge blurring, and changes in illumination, making it difficult for detection models to stably extract the edge, texture, and local structural features of objects.
[0003] Existing deep learning-based object detection methods have been widely applied in traffic object detection tasks. Among them, the YOLO series of networks are often used in real-time object detection scenarios due to their high detection speed and end-to-end inference capabilities. These methods typically extract image features through a backbone network, then enhance the representation of objects at different scales through a multi-scale feature fusion structure, and finally, the detection head completes object classification and bounding box regression. Although existing methods can meet the object detection needs in general scenarios, in complex traffic environments, conventional convolutional feature extraction methods do not make sufficient use of local prior information, and high-level semantic features still have limited ability to model complex backgrounds and multi-scale objects, which can easily lead to missed detections of small objects, inaccurate object boundary localization, and false detections under similar background interference.
[0004] Furthermore, the bounding box regression loss function directly affects the target localization accuracy and training stability. While existing IoU-type loss functions can constrain the difference between predicted and ground truth bounding boxes by factors such as overlap area, center distance, or aspect ratio, their comprehensive constraints on the regression contribution of samples of different quality, differences in target shape and scale, and center offset direction are still insufficient. This is particularly problematic in traffic scenarios where target scale differences are significant, such as those involving vehicles, pedestrians, and cyclists, leading to insufficient localization accuracy and unstable training convergence. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a target detection method, system, and device based on multi-domain perception enhancement, which can improve the accuracy of target detection in complex traffic scenarios.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] In a first aspect, embodiments of the present invention provide a target detection method based on multi-domain perception enhancement, the method comprising the following steps:
[0008] Step 1: Dataset preprocessing:
[0009] The KITTI dataset was used as the original dataset. Image data and corresponding annotation information were read from the KITTI dataset. Invalid samples and abnormal annotations were removed, and the target categories were uniformly organized. Vehicle-related categories were merged into the "car" category, cyclist-related categories into the "cyclist" category, and pedestrian-related categories into the "person" category. Subsequently, the image size and bounding box coordinates were uniformly processed to generate image samples and label files suitable for training the object detection network. Finally, the processed dataset was divided into training, validation, and test sets in a 6:2:2 ratio for subsequent network training, parameter validation, and detection performance evaluation.
[0010] Step 2: Construct a multi-domain perception-enhanced target detection network:
[0011] A target detection network is built based on YOLOv11. A Local Prior Enhancement Module (LPEM) is introduced in the backbone feature extraction stage, and a Dynamic Spectral Attention Enhancement Module (DSAEM) is introduced in the high-level semantic feature extraction stage. The Wise-Shape Intersection over Union (WoU) loss function is used to optimize the bounding box regression process. Through this structure, the network can simultaneously enhance local edge texture representation, high-level semantic modeling capabilities, and bounding box localization constraints.
[0012] Step 2.1, Local Prior Enhancement Module (LPEM):
[0013] The Local Prior Enhancement (LPEM) module is used to enhance edge, texture, and local structural information during the backbone feature extraction process. The LPEM module includes functional units such as local prior extraction, local feature enhancement attention, MLP (Multi-Layer Perceptron) recalibration, DropPath and batch normalization, and residual fusion.
[0014] In the local prior extraction process, input features Local prior maps are generated after processing with edge enhancement and smoothing enhancement branches respectively. As shown in equation (1):
[0015]
[0016] In the formula, For input features, This is a local prior map. For the feature extraction function of the edge enhancement branch, To smooth the feature extraction function of the enhanced branch, This represents the local prior branch adjustment coefficient.
[0017] Local Feature Enhancement Attention for Local Prior Maps Input features Perform joint modeling to make the local prior map For input features Guided enhancement is performed on the edges, textures, and local structural responses in the image to obtain local enhancement features. .
[0018] Local enhancement features After MLP recalibration, channel recalibration is performed through two 1×1 convolutional layers to obtain recalibrated features. Subsequently, the features were recalibrated. After DropPath and batch normalization processing, and compared with the input features... Perform residual fusion to obtain enhanced feature output. As shown in equation (2):
[0019]
[0020] In the formula, This indicates the output of enhanced features; This indicates the DropPath operation; This indicates batch normalization operation.
[0021] Subsequently, The input is fed into the SPPF (Spatial Pyramid Pooling-Fast) module. The SPPF module expands the feature receptive field through successive pooling operations, aggregating contextual information at different scales. This allows the edge, texture, and local structure enhancement information obtained from LPEM to be further integrated into high-level semantic representations, resulting in high-level semantic features. .
[0022] Step 2.2, Dynamic Spectral Attention Enhancement Module (DSAEM):
[0023] The Dynamic Spectral Attention Enhancement Module (DSAEM) is used to enhance the expressive power of high-level semantic features for complex backgrounds and multi-scale targets. The DSAEM includes functional units such as channel mapping, preservation branching, enhancement branching, feature concatenation, and convolutional fusion.
[0024] Channel mapping receives high-level semantic features And high-level semantic features are processed through 1×1 convolution. Perform channel mapping to obtain the channel-mapped features Including preserving branch features and enhanced branch features Preserve branch features The retained branches are directly passed to feature concatenation to enhance the branch features. The input enhancement branch performs feature enhancement processing;
[0025] The augmentation branch first applies dynamic modulation and TSSA (Token Statistics Self-Attention) attention to the augmentation branch features. Global modeling is performed, and intermediate features are obtained through residual fusion and local context enhancement. As shown in equation (3):
[0026]
[0027] In the formula, Represents the TSSA attention function; Indicates dynamic modulation 1; This indicates local context enhancement 1.
[0028] intermediate features The spectral features are then enhanced using Dynamic Modulation 2 and the Spectral Enhanced Feed-Forward Network (SEFFN), and the enhanced branch output is obtained through residual fusion and local context enhancement 2. As shown in equation (4):
[0029]
[0030] In the formula, Indicates dynamic modulation 2; This indicates local context enhancement 2; This represents the SEFFN function of the spectrum enhancement feedforward network.
[0031] Feature splicing will preserve branch features With enhanced branch output The data is concatenated and then fused using convolution to obtain enhanced feature output. As shown in equation (5):
[0032]
[0033] In the formula, This indicates enhanced feature output. Represents the convolution fusion function. This indicates feature splicing.
[0034] Step 2.3, Wise-ShapeIoU loss function:
[0035] The Wise-ShapeIoU loss function is used in the bounding box regression process. The Wise-ShapeIoU loss function introduces Wise dynamic focusing weights on the basis of ShapeIoU loss, which are used to simultaneously constrain the overlap error, center offset error, shape scale error and regression contribution of samples of different quality between the predicted bounding box and the true bounding box, as shown in Equation (6):
[0036]
[0037] In the formula, Represents the Wise-ShapeIoU loss function; Indicates dynamic focus weight; This represents the regression error between the predicted bounding box and the true bounding box. The shape distance term is used to apply a weighted constraint on the center offset between the predicted bounding box and the true bounding box based on the width, height, and shape of the true bounding box. This represents the shape cost term, used to constrain the width and height differences between the predicted bounding box and the true bounding box.
[0038] Among them, the introduced dynamic focus weight This allows the bounding box regression training process to adaptively adjust the loss contribution based on the sample regression quality. This weight is determined by a relative factor of sample quality. The relative relationship between the current sample regression error and the average regression error during training is described, and the regression contribution in the Wise-ShapeIoU loss is dynamically adjusted accordingly. The calculation method is shown in Equations (7) and (8):
[0039]
[0040] In the formula, and This represents the adjustment parameter in the dynamic focus weight; This represents a constant to prevent the denominator from being zero; This represents the moving average of the regression error during training.
[0041] The shape distance term in the formula This can improve the stability of bounding box center position regression. This term combines the width and height shape of the real bounding box to adaptively weight the center offset of the predicted bounding box from the real bounding box in the horizontal and vertical directions, as shown in Equation (9):
[0042]
[0043] In the formula, and These represent the x-coordinate and y-coordinate of the predicted bounding box center point, respectively. and These represent the x and y coordinates of the center point of the actual bounding box, respectively. and These represent the width and height of the smallest bounding rectangle that simultaneously encloses the predicted bounding box and the ground truth bounding box, respectively. This represents a constant to prevent the denominator from being zero; and These represent the shape weights determined by the width and height of the actual bounding box, respectively, and are used to adaptively weight the lateral and vertical offsets of the center point.
[0044] The shape cost term in the formula This can further constrain the width and height shape differences between the predicted bounding box and the true bounding box. This term measures the shape deviation between the predicted bounding box and the true bounding box in the width and height directions, respectively, and enhances the penalty for large shape deviations through exponential functions and power operations, so that the bounding box regression process not only focuses on the overlapping area and center position, but also on the consistency of the target width-to-height ratio and shape scale, as shown in Equation (10):
[0045]
[0046] In the formula, This indicates the shape difference between the predicted bounding box and the true bounding box in the width direction. This indicates the shape difference between the predicted bounding box and the true bounding box in the height direction; Represents a constant.
[0047] Step 3: Train and evaluate the multi-domain perception-enhanced object detection network:
[0048] The pre-built multi-domain perception-enhanced object detection network was trained using the training set, and the network model was validated using the validation set. Subsequently, it was evaluated on the test set, with evaluation metrics including accuracy. Recall rate and Among them, accuracy Recall rate represents the proportion of correctly detected targets in the test results. This indicates the proportion of real targets that were correctly detected. This represents the average AP for each category when the IoU threshold is 0.5.
[0049] Secondly, the present invention also provides a target detection system based on multi-domain perception enhancement, which uses the above-mentioned method to detect targets in complex traffic scenarios. The system includes:
[0050] The data preprocessing module is used to preprocess traffic scene image data;
[0051] A multi-domain perception-enhanced target detection network performs target detection based on preprocessed traffic scene image data.
[0052] Thirdly, the present invention also provides an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the above-described target detection method based on multi-domain perception enhancement.
[0053] Compared with the prior art, the present invention has at least the following beneficial effects:
[0054] 1. This invention can improve the accuracy of target detection in complex traffic scenarios.
[0055] 2. This invention introduces a Local Prior Enhancement (LPEM) module, a Dynamic Spectral Attention Enhancement (DSAEM) module, and an improved ShapeIoU loss function (Wise-ShapeIoU loss function) into the target detection network. LPEM enhances the expression of target edges, textures, and local structures, improving the backbone network's feature extraction capabilities for small targets, targets with blurred boundaries, and targets with weak local features. DSAEM enhances the dynamic semantic relationships, global dependency information, and spectral feature expression in high-level semantic features, improving the network's ability to represent complex backgrounds and multi-scale targets. The Wise-ShapeIoU loss function jointly constrains overlap errors, center offset errors, shape and scale errors, and the regression contributions of samples of different quality, improving the bounding box localization accuracy and training stability.
[0056] Compared with existing technologies, this invention improves segmentation accuracy and robustness while ensuring efficient computation, making it suitable for various remote sensing image segmentation tasks. The method of this invention can better address complex features, targets of different scales, and class imbalances in remote sensing images, and has broad application prospects.
[0057] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.
[0058] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.
[0061] Figure 1 This is a flowchart illustrating the target detection method based on multi-domain perception enhancement provided by the present invention.
[0062] Figure 2 This is a schematic diagram of the structure of the multi-domain perception-enhanced target detection network provided by the present invention.
[0063] Figure 3 This is a schematic diagram of the Local Prior Enhancement Module (LPEM) provided by the present invention.
[0064] Figure 4 A schematic diagram of the structure of the Dynamic Spectral Attention Enhancement Module (DSAEM) provided by this invention.
[0065] Figure 5 This is a schematic diagram of vehicle detection in complex traffic scenarios provided by the present invention.
[0066] Figure 6 This is a schematic diagram illustrating the detection of cyclists and pedestrians in complex traffic scenarios provided by the present invention.
[0067] Figure 7 A schematic diagram of the electronic device provided by the present invention. Detailed Implementation
[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0069] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0070] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0071] See Figures 1 to 4 As shown, this invention provides a target detection scheme based on multi-domain perception enhancement, which mainly includes the following steps:
[0072] Step 1: Dataset preprocessing:
[0073] The KITTI dataset was used as the original dataset. Image data and corresponding annotation information were read from the KITTI dataset. Invalid samples and abnormal annotations were removed, and the target categories were uniformly organized. Vehicle-related categories were merged into the "car" category, cyclist-related categories into the "cyclist" category, and pedestrian-related categories into the "person" category. Subsequently, the image size and bounding box coordinates were uniformly processed to generate image samples and label files suitable for training the object detection network. Finally, the processed dataset was divided into training, validation, and test sets in a 6:2:2 ratio for subsequent network training, parameter validation, and detection performance evaluation.
[0074] Step 2: Construct a multi-domain perception-enhanced target detection network:
[0075] A multi-domain perception-enhanced object detection network was built based on YOLOv11. For example... Figure 2As shown, the image is first input into the backbone network, where convolutional layers extract basic image features. During the backbone feature extraction stage, a Local Prior Enhancement (LPEM) module is introduced to enhance edge, texture, and local structural information in the features, improving the backbone network's ability to express local details. Subsequently, the high-level features output from the backbone network are input into the SPPF module, which expands the receptive field and aggregates multi-scale contextual information through fast spatial pyramid pooling to obtain high-level semantic features. These high-level semantic features are further input into the Dynamic Spectral Attention Enhancement (DSAEM) module, which improves the network's ability to model complex backgrounds, multi-scale targets, and high-level semantic relationships through dynamic modulation, TSSA attention, a spectrum-enhanced feedforward network (SEFFN), and local context enhancement. Then, the network performs multi-scale feature fusion through upsampling, feature concatenation, and the C3k2 module, forming detection features for small, medium, and large targets, respectively. The detection head outputs classification results and bounding box localization results for traffic targets such as vehicles, cyclists, and pedestrians. During training, the Wise-ShapeIoU loss function is used to optimize the bounding box regression process to improve target localization accuracy and training stability.
[0076] Step 2.1, Local Prior Enhancement Module (LPEM):
[0077] The Local Prior Enhancement (LPEM) module is used to enhance edge, texture, and local structural information during the backbone feature extraction process. For example... Figure 3 As shown, the Local Prior Enhancement Module (LPEM) includes local prior extraction, local feature enhancement attention, MLP recalibration, DropPath and batch normalization, and residual fusion.
[0078] In the local prior extraction process, input features Local prior maps are generated after processing with edge enhancement and smoothing enhancement branches respectively. As shown in equation (1):
[0079]
[0080] In the formula, For input features, This is a local prior map. For the feature extraction function of the edge enhancement branch, To smooth the feature extraction function of the enhanced branch, This represents the local prior branch adjustment coefficient.
[0081] Local Feature Enhancement Attention for Local Prior Maps Input features Perform joint modeling to make the local prior map For input features Guided enhancement is performed on the edges, textures, and local structural responses in the image to obtain local enhancement features. .
[0082] Local enhancement features After MLP recalibration, channel recalibration is performed through two 1×1 convolutional layers to obtain recalibrated features. Subsequently, the features were recalibrated. After DropPath and batch normalization processing, and compared with the input features... Perform residual fusion to obtain enhanced feature output. As shown in equation (2):
[0083]
[0084] In the formula, This indicates the output of enhanced features; This indicates the DropPath operation; This indicates batch normalization operation.
[0085] Subsequently, The input is fed into the SPPF module. The SPPF module expands the feature receptive field through successive pooling operations, aggregating contextual information at different scales. This allows the edge, texture, and local structure enhancement information obtained from LPEM to be further integrated into high-level semantic representations, resulting in high-level semantic features. .
[0086] Step 2.2, Dynamic Spectral Attention Enhancement Module (DSAEM):
[0087] like Figure 4 As shown, the Dynamic Spectral Attention Enhancement Module (DSAEM) is used to enhance the expressive power of high-level semantic features for complex backgrounds and multi-scale targets. The DSAEM module includes channel mapping, preservation branch, enhancement branch, feature concatenation, and convolutional fusion.
[0088] Channel mapping receives high-level semantic features And high-level semantic features are processed through 1×1 convolution. Perform channel mapping to obtain the channel-mapped features Including preserving branch features and enhanced branch features Preserve branch features The retained branches are directly passed to feature concatenation to enhance the branch features. The input enhancement branch performs feature enhancement processing;
[0089] The enhancement branch is first enhanced by dynamic modulation 1 and TSSA attention to the enhancement branch features. Global modeling is performed, and intermediate features are obtained through residual fusion and local context enhancement. As shown in equation (3):
[0090]
[0091] In the formula, Represents the TSSA attention function; Indicates dynamic modulation 1; This indicates local context enhancement 1.
[0092] intermediate features The spectral features are then enhanced using Dynamic Modulation 2 and the Spectrum Enhancement Feedforward Network (SEFFN), and the enhanced branch output is obtained through residual fusion and Local Context Enhancement 2. As shown in equation (4):
[0093]
[0094] In the formula, Indicates dynamic modulation 2; This indicates local context enhancement 2; This represents the SEFFN function of the spectrum enhancement feedforward network.
[0095] Feature splicing will preserve branch features With enhanced branch output The data is concatenated and then fused using convolution to obtain enhanced feature output. As shown in equation (5):
[0096]
[0097] In the formula, This indicates enhanced feature output. Represents the convolution fusion function. This indicates feature splicing.
[0098] Step 2.3, Wise-ShapeIoU loss function:
[0099] The Wise-ShapeIoU loss function is used in the bounding box regression process. The Wise-ShapeIoU loss function introduces Wise dynamic focusing weights on the basis of ShapeIoU loss, which are used to simultaneously constrain the overlap error, center offset error, shape scale error and regression contribution of samples of different quality between the predicted bounding box and the true bounding box, as shown in Equation (6):
[0100]
[0101] In the formula, Represents the Wise-ShapeIoU loss function; Indicates dynamic focus weight; This represents the regression error between the predicted bounding box and the true bounding box. The shape distance term is used to apply a weighted constraint on the center offset between the predicted bounding box and the true bounding box based on the width, height, and shape of the true bounding box. This represents the shape cost term, used to constrain the width and height differences between the predicted bounding box and the true bounding box.
[0102] To enable the bounding box regression training process to adaptively adjust the loss contribution based on the sample regression quality, dynamic focusing weights are introduced. This weight is determined by the relative factor of sample quality. The relative relationship between the current sample regression error and the average regression error during training is described, and the regression contribution in the Wise-ShapeIoU loss is dynamically adjusted accordingly. The calculation method is shown in Equations (7) and (8):
[0103]
[0104] In the formula, and This represents the adjustment parameter in the dynamic focus weight; This represents a constant to prevent the denominator from being zero; This represents the moving average of the regression error during training.
[0105] The shape distance term in the formula This can improve the stability of bounding box center position regression. This term combines the width and height shape of the real bounding box to adaptively weight the center offset of the predicted bounding box from the real bounding box in the horizontal and vertical directions, as shown in Equation (9):
[0106]
[0107] In the formula, and These represent the x-coordinate and y-coordinate of the predicted bounding box center point, respectively. and These represent the x and y coordinates of the center point of the actual bounding box, respectively. and These represent the width and height of the smallest bounding rectangle that simultaneously encloses the predicted bounding box and the ground truth bounding box, respectively. This represents a constant to prevent the denominator from being zero; and These represent the shape weights determined by the width and height of the actual bounding box, respectively, and are used to adaptively weight the lateral and vertical offsets of the center point.
[0108] The shape cost term in the formula This can further constrain the width and height shape differences between the predicted bounding box and the true bounding box. This term measures the shape deviation between the predicted bounding box and the true bounding box in the width and height directions, respectively, and enhances the penalty for large shape deviations through exponential functions and power operations, so that the bounding box regression process not only focuses on the overlapping area and center position, but also on the consistency of the target width-to-height ratio and shape scale, as shown in Equation (10):
[0109]
[0110] In the formula, This indicates the shape difference between the predicted bounding box and the true bounding box in the width direction. This indicates the shape difference between the predicted bounding box and the true bounding box in the height direction; Represents a constant.
[0111] Step 3: Train and evaluate the multi-domain perception-enhanced object detection network:
[0112] The multi-domain perception-enhanced object detection network was trained using the training set, and the network model was validated using the validation set. In this embodiment, the input image size was set to 640×640, the number of training epochs was set to 300, the batch size was set to 16, the optimizer was SGD (Stochastic Gradient Descent), the initial learning rate was set to 0.01, and the weight decay coefficient was set to 0.0005. After training, the detection accuracy was evaluated on the test set, and the evaluation metrics included precision. Recall rate and Among them, accuracy Recall rate represents the proportion of correctly detected targets in the test results. This indicates the proportion of real targets that were correctly detected. The average AP (Average Precision) for each category is represented when the IoU threshold is 0.5, as shown in equations (11) to (13):
[0113]
[0114] In the formula, Indicates the number of targets that were correctly detected; Indicates the number of false positives; Indicates the number of targets missed; Indicates the detection category, in this embodiment These are car, cyclist, and person, respectively. Indicates category With an IoU threshold of 0.5 and a recall rate of The corresponding accuracy.
[0115] To further verify the actual technical effect of the method of the present invention, under the same dataset partitioning, the same input image size, and the same training parameters, the method of the present invention was compared with the current mainstream YOLO model. The results are shown in Table 1 below:
[0116] Table 1 Comparison results of different models
[0117]
[0118] As shown in Table 1, under the same training parameters, the accuracy of the method of the present invention is... Recall rate and The accuracy rates were 0.919, 0.920, and 0.850 respectively, all outperforming existing mainstream YOLO models. Compared to the YOLOv11 base model, the accuracy... The recall rate increased from 0.877 to 0.919, a 4.2% improvement; It increased from 0.739 to 0.920, a rise of 1.81%; It increased from 0.830 to 0.850, a 2% improvement.
[0119] In summary, this invention provides a target detection method based on multi-domain perception enhancement. Through dataset preprocessing, construction of a multi-domain perception enhancement target detection network, and network training and evaluation, it can achieve accurate detection of targets such as vehicles, cyclists, and pedestrians in complex traffic scenarios, with detection results as shown in the figure. Figure 5 and Figure 6 As shown, this method introduces a Local Prior Enhancement (LPEM) module in the backbone feature extraction stage and a Dynamic Spectral Attention Enhancement (DSAEM) module in the high-level semantic feature extraction stage. It also uses the Wise-ShapeIoU loss function to optimize the bounding box regression process, thereby improving the target detection performance from three aspects: local structure representation, semantic feature modeling, and localization loss constraints.
[0120] Furthermore, the present invention also provides a target detection system based on multi-domain perception enhancement, which uses the above-described method to detect targets in complex traffic scenarios. The system includes:
[0121] The data preprocessing module is used to preprocess traffic scene image data;
[0122] A multi-domain perception-enhanced target detection network performs target detection based on preprocessed traffic scene image data.
[0123] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment, and will not be repeated here.
[0124] Additionally, refer to Figure 7 As shown, this embodiment of the invention also provides an electronic device, which may include a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may also include a computer program stored in the memory 11 and executable on the processor 10. The processor executes the computer program to implement a target detection method based on multi-domain perception enhancement in the above method embodiment.
[0125] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units, microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control core of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 and calls data stored in the memory 11 to perform various functions and process data within the electronic device.
[0126] The memory 11 may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, and any suitable combination thereof.
[0127] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, software systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0128] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.
[0129] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0130] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A target detection method based on multi-domain perception enhancement, characterized in that, The method includes the following steps: Preprocessing traffic scene image data; A multi-domain perception-enhanced object detection network is built based on the YOLO network. A local prior enhancement module is introduced during the backbone feature extraction process, and a dynamic spectral attention enhancement module is introduced during the high-level semantic feature extraction process. An improved ShapeIoU loss function is used to optimize the bounding box regression process. The local prior enhancement module includes a local prior extraction unit, a local feature enhancement attention unit, an MLP recalibration unit, a DropPath and batch normalization unit, and a residual fusion unit. Wherein: The local prior extraction unit includes an edge enhancement branch, a smoothing enhancement branch, and a local prior map unit; the local prior extraction unit generates a local prior map based on the features extracted by the edge enhancement branch and the smoothing enhancement branch, and its expression is: In the formula, For input features, This is a local prior map. For the feature extraction function of the edge enhancement branch, To smooth the feature extraction function of the enhanced branch, This represents the local prior branch adjustment coefficient; The local feature enhancement attention unit on the local prior map Input features Perform joint modeling to make the local prior map For input features Local enhancement features are obtained by guiding and enhancing the edge, texture, and local structure responses in the data. ; The MLP recalibration unit enhances local features. Recalibrated features are obtained by performing channel recalibration through convolution. ; The DropPath and batch normalization unit recalibrate features. Perform regularization and normalization processing; The residual fusion unit will recalibrate the features after regularization and normalization. Input features Residual fusion is performed to obtain enhanced feature output. Its expression is: In the formula, This indicates the output of enhanced features; This indicates the DropPath operation; This indicates a batch normalization operation; Enhanced feature output through fast spatial pyramid pooling Expanding the receptive field of features and aggregating multi-scale contextual information yields high-level semantic features. ; The dynamic spectral attention enhancement module includes a channel mapping unit, a preservation branch, an enhancement branch, a feature concatenation unit, and a convolutional fusion unit; wherein: The channel mapping unit receives high-level semantic features. And through convolution to process high-level semantic features Perform channel mapping to obtain the channel-mapped features Including preserving branch features and enhanced branch features Preserve branch features The retained branches are directly passed to the feature assembly unit to enhance the branch features. The input enhancement branch performs feature enhancement processing; The enhancement branch includes a first dynamic modulation unit, a TSSA attention unit, a first residual fusion unit, a first local context enhancement unit, a second dynamic modulation unit, a spectrum enhancement feedforward network, a second residual fusion unit, and a second local context enhancement unit; the enhancement branch first passes the first dynamic modulation unit and the TSSA attention unit to enhance the branch features. Global modeling is performed, and intermediate features are obtained through the first residual fusion unit and the first local context enhancement unit. Then, the intermediate features are processed through a second dynamic modulation unit and a spectrum enhancement feedforward network. Spectral feature enhancement is performed, and the enhanced branch output is obtained through the second residual fusion unit and the second local context enhancement unit. Its expression is: In the formula, Represents the TSSA attention function; and These represent the first dynamic modulation and the second dynamic modulation, respectively. and These represent the first local context enhancement and the second local context enhancement, respectively. This represents the spectrum-enhancing feedforward network function; The feature splicing unit will retain the branch features With the enhanced branch output The data are concatenated and then fused using convolution to obtain the final enhanced feature output. ; A multi-domain perception-enhanced target detection network is trained and evaluated based on preprocessed traffic scene image data. The evaluated multi-domain perception-enhanced target detection network is then used to detect targets in traffic scenes.
2. The target detection method based on multi-domain perception enhancement according to claim 1, characterized in that, The preprocessing includes removing invalid samples and anomaly labels, and uniformly organizing the target categories and merging related categories.
3. The target detection method based on multi-domain perception enhancement according to claim 1, characterized in that, The improved ShapeIoU loss function introduces dynamic focusing weights on top of the ShapeIoU loss to simultaneously constrain the overlap error, center offset error, shape scale error, and regression contribution of samples of different quality between the predicted and ground truth bounding boxes. Its expression is: In the formula, This represents the ShapeIoU loss function that incorporates dynamic focusing weights; Indicates dynamic focus weight; This represents the regression error between the predicted bounding box and the true bounding box. This represents the moving average of the regression error during training. The shape distance term is used to apply a weighted constraint on the center offset between the predicted bounding box and the true bounding box based on the width, height, and shape of the true bounding box. This represents the shape cost term, used to constrain the width and height differences between the predicted bounding box and the true bounding box; This represents the relative factor of sample quality, used to characterize the degree of deviation of the current sample regression error from the average regression error during training; and This represents the adjustment parameter in the dynamic focus weight; and These represent the x-coordinate and y-coordinate of the predicted bounding box center point, respectively. and These represent the x and y coordinates of the center point of the actual bounding box, respectively. and These represent the width and height of the smallest bounding rectangle that simultaneously encloses the predicted bounding box and the ground truth bounding box, respectively. and Both represent constants to prevent the denominator from being zero; and These represent the shape weights determined by the width and height of the actual bounding box, used for adaptive weighting of the lateral and vertical offsets of the center point; This indicates the shape difference between the predicted bounding box and the true bounding box in the width direction. This indicates the shape difference between the predicted bounding box and the true bounding box in the height direction; Represents a constant.
4. A target detection system based on multi-domain perception enhancement, characterized in that, The system, comprising the method described in any one of claims 1-3, for detecting targets in a traffic scene, includes: The data preprocessing module is used to preprocess traffic scene image data; A multi-domain perception-enhanced target detection network performs target detection based on preprocessed traffic scene image data.
5. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement a target detection method based on multi-domain perception enhancement as described in any one of claims 1-3.
Citation Information
Patent Citations
Substation peripheral potential safety hazard identification method and system based on target detection
CN118397545A
Improved YOLOv11 unmanned aerial vehicle small target detection method
CN119992524A
