Ultra-wide-angle traffic sign detection method and system based on adaptive dynamic pruning

Abstract

The present invention discloses an ultra-wide-angle traffic sign detection method and system based on adaptive dynamic pruning, which belongs to the field of computer vision and target detection, including: obtaining ultra-wide-angle traffic sign data to be tested and performing splicing and fusion; inputting the spliced ​​and fused data into a trained traffic sign detection model to obtain traffic sign detection results; wherein, the training of the traffic sign detection model includes: constructing an ultra-wide-angle traffic sign data set; building a traffic sign detection model, including: a human eye attention focus position detection module, a detection module based on dynamic pruning, and a detection result output module. The constructed traffic sign detection model is trained using the constructed ultra-wide-angle traffic sign data set to obtain a trained traffic sign detection model. The present invention evaluates image complexity in real time and flexibly adjusts the model's calculation path, optimizes the allocation and use of computing resources, and improves the perception and processing capabilities of large-viewing-angle targets in complex dynamic scenes.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and target detection, and in particular relates to an ultra-wide-angle traffic sign detection method and system based on adaptive dynamic pruning. Background Art

[0002] With the rapid development of drone surveillance, autonomous driving, intelligent security, and other fields, the demand for large-viewing-angle, high-precision object detection is becoming increasingly urgent. Existing object detection technologies, particularly those based on deep learning frameworks like YOLOv8, perform well under standard viewing angles but still face technical challenges in complex dynamic scenes.

[0003] First, traditional target detection methods often rely on single-view cameras. This makes it difficult for single-view cameras to fully cover all targets in complex traffic environments, resulting in detection blind spots. At the same time, detection accuracy is limited in ultra-wide-angle scenarios, especially in the edge areas of the image, where significant distortion exists. Although existing distortion correction technologies have been applied, they still find it difficult to completely eliminate image distortion, resulting in a significant decrease in target detection accuracy in edge areas. For example, although some algorithms have introduced wide-angle distortion correction technologies, the image quality and target detection effects in ultra-wide-angle fields of view are still significantly insufficient, and the viewing angle limitations cannot be effectively overcome. Second, current traffic sign detection systems typically rely on deep neural networks for target recognition. Although they can achieve high detection accuracy, they consume huge computing resources, especially when running on edge devices or embedded devices, making it difficult to meet the needs of real-time detection.

[0004] Therefore, it is necessary to design an ultra-wide-angle traffic sign detection method and system based on adaptive dynamic pruning to address the above problems. Summary of the Invention

[0005] The purpose of the present invention is to address the problem that the detection accuracy in ultra-wide-angle scenes is limited, especially in the edge areas of the image, there are significant distortion problems, and it is difficult to meet real-time detection requirements when running on edge devices or embedded devices. The present invention provides an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, which utilizes Fast Library for Approximate Nearest Neighbors (FLANN) matching and dynamic programming method to generate large-viewing angle stitching images, eliminate the influence of ultra-wide-angle distortion, dynamically prune the deep learning network by real-time evaluation of feature activation intensity, and output structured detection results in combination with non-maximum suppression and confidence filtering. It achieves both image correction and real-time requirements under ultra-wide-angle field of view on edge devices, thereby meeting increasingly complex target detection application scenarios.

[0006] According to one aspect of this specification, a method for detecting ultra-wide-angle traffic signs based on adaptive dynamic pruning is provided, comprising:

[0007] Obtain the ultra-wide-angle traffic sign data to be tested, perform pre-processing and video stream splicing and fusion;

[0008] The spliced ​​and fused data is input into a trained traffic sign detection model to obtain a traffic sign detection result; wherein the training of the traffic sign detection model includes:

[0009] Construct an ultra-wide-angle traffic sign dataset;

[0010] Build a traffic sign detection model, including: a human eye attention location detection module for dynamically adjusting detection weights; a dynamic pruning-based detection module for dynamically adjusting the network structure according to image complexity; and a detection result output module for outputting the final detection results.

[0011] The constructed ultra-wide-angle traffic sign dataset is used to train the constructed traffic sign detection model to obtain a trained traffic sign detection model.

[0012] Furthermore, pre-processing and video stream splicing and fusion are performed, including:

[0013] Preprocessing of ultra-wide-angle traffic sign data, including denoising, normalization, and resizing;

[0014] The pre-processed data is input into the video stream splicing and fusion module for splicing and fusion to obtain a large-viewing angle video stream.

[0015] Furthermore, the human eye attention focus position detection module includes an attention recognition module and a dynamic weight adjustment module:

[0016] The attention recognition module locates the areas that need priority attention by analyzing image texture, moving targets and visual blind spot characteristics;

[0017] The dynamic weight adjustment module automatically increases the attention to blind spots or high-speed moving targets according to the complexity of the real-time scene.

[0018] Furthermore, the dynamic pruning-based detection module includes a pruning evaluation module and a network structure optimization module:

[0019] The pruning evaluation module filters out redundant computation paths by analyzing feature activation strength in real time;

[0020] The network structure optimization module adaptively retains key feature extraction paths according to image complexity.

[0021] Furthermore, the detection result output module includes a result integration module and a visual transmission module:

[0022] The result integration module removes redundant and low-quality detection frames through non-maximum suppression and confidence filtering;

[0023] The visual transmission module is displayed or stored in real time through the vehicle terminal, cloud platform or security system.

[0024] Furthermore, the video stream splicing and fusion module further includes:

[0025] Frame stitching parameter calculation module, used to extract feature points and calculate the optimal stitching line;

[0026] The video frame fusion module is used to quickly stitch together a large-viewing angle video stream based on the optimal stitching line.

[0027] According to one aspect of this specification, an ultra-wide-angle traffic sign detection system based on adaptive dynamic pruning is provided, comprising:

[0028] The data acquisition module is used to obtain the ultra-wide-angle traffic sign data to be tested, perform pre-processing and video stream splicing and fusion;

[0029] The traffic sign detection module is used to input the spliced ​​and fused data into the trained traffic sign detection model to obtain traffic sign detection results. The training of the traffic sign detection model includes:

[0030] Construct an ultra-wide-angle traffic sign dataset;

[0031] Build a traffic sign detection model, including: a human eye attention location detection module for dynamically adjusting detection weights; a dynamic pruning-based detection module for dynamically adjusting the network structure according to image complexity; and a detection result output module for outputting the final detection results.

[0032] The constructed ultra-wide-angle traffic sign dataset is used to train the constructed traffic sign detection model to obtain a trained traffic sign detection model.

[0033] According to one aspect of this specification, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning are implemented.

[0034] According to one aspect of the present specification, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning are implemented.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention extracts feature points of multiple video streams by using the Speeded Up Robust Features (SURF) algorithm, and uses FLANN matching and dynamic programming to generate large-viewing angle stitching images to eliminate the influence of ultra-wide-angle distortion.

[0037] 2. The present invention simulates the visual characteristics of the human eye, dynamically corrects the line of sight error through radial basis function interpolation and cross-ratio invariance, identifies key areas in the image and assigns attention weights, prioritizes blind spots and moving targets, and then dynamically prunes the deep learning network through real-time evaluation of feature activation intensity, adaptively retains key computing paths, significantly reduces redundant computing, effectively reduces resource waste in high-complexity image processing, and improves the system's computing efficiency and application adaptability.

[0038] 3. The present invention combines non-maximum suppression and confidence filtering to output structured detection results, and superimposes them on the original image for visual transmission, realizing high-precision traffic sign detection in ultra-wide-angle scenes, taking into account both image correction effects and real-time requirements, and effectively balancing detection accuracy and resource consumption on edge devices, meeting the needs of real-time monitoring and autonomous driving applications in complex traffic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of multi-view video stream splicing and fusion according to an embodiment of the present invention;

[0042] Figure 3 This is a structural diagram of a human eye attention focus position detection module according to an embodiment of the present invention;

[0043] Figure 4 This is a structural diagram of a dynamic pruning traffic sign detection module according to an embodiment of the present invention;

[0044] Figure 5 This is a structural diagram of the detection result output module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] like Figure 1 As shown, an embodiment of the present invention provides an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, including: step S1, collecting a video stream I that needs to be detected for traffic signs and performing preprocessing; step S2, multi-view video stream splicing and fusion, splicing and aligning the preprocessed multiple video stream input line images to obtain a wide-view image; step S3, detecting the position where the human eye focuses attention, by simulating the visual characteristics of the human eye, dynamically adjusting the detection weight, and giving priority to important areas in the image; step S4, traffic sign detection based on dynamic pruning, optimizing the calculation path of the deep learning network model through a dynamic pruning strategy, and dynamically adjusting the network structure according to the image complexity during the real-time detection process; step S5, outputting the detection results, combining the deep learning network model optimized in the previous steps to output the final detection results.

[0047] Specifically, in the embodiment of the present invention, the preprocessing of the video stream I includes denoising, normalization and image resizing to ensure the consistency and accuracy of the video stream input data. The preprocessing operation ensures the accuracy of subsequent image stitching and target detection.

[0048] Specifically, the embodiment of the present invention also provides a multi-view video stream splicing and fusion method, which splices and aligns the pre-processed multiple video stream input line images to obtain a large-view image. Figure 2 As shown, the frame splicing parameter calculation module includes a feature extraction unit and a stitching line calculation unit. The feature extraction unit takes as input the image information of the first frame of multiple video streams and outputs the extracted feature points. The stitching line calculation unit takes as input the feature points of the first frame of multiple video streams and outputs the optimal stitching line for the wide-angle image formed by matching the feature points of the multiple images. The video frame fusion module takes as input the optimal stitching line calculated by the frame splicing parameter calculation module and outputs the subsequent wide-angle video streams that are quickly spliced ​​together according to the stitching line.

[0049] Specifically, in step S2, the frame splicing parameter calculation module extracts the feature points of the first frame image information of the multi-channel video stream based on the feature extraction unit to obtain the image feature points, and then calculates the optimal stitching line of the multi-channel image through the stitching line calculation unit, and finally outputs the wide-angle image. Among them, the feature extraction unit uses the SURF algorithm to extract feature points. The SURF algorithm detects the extreme points of the image through the Hessian matrix, and the binary function The Hessian where is is shown in formula (1).

[0050] (1)

[0051] in, The image is at the pixel point The pixel value at , the Hessian matrix discriminant is shown in formula (2).

[0052] (2)

[0053] Where det(H) is the determinant of the Hessian matrix, and the point where det(H) is greater than 0 is the extreme point. The characteristic of the Hessian matrix expression of this pixel in the Gaussian pyramid is shown in Equation (3).

[0054] (3)

[0055] in, The scale is Gaussian pyramid image. 、 Indicates that the image is at scale The second-order partial derivative under , Indicates that the image is at scale The mixed derivatives under , the actual calculation is accelerated by the box filter (BoxFilter) approximation.

[0056] Specifically, the seam calculation unit first uses the FLANN algorithm to match feature points. FLANN is a matching operation method based on Kd-tree. Kd-tree is a balanced binary tree, mainly used to divide the K-dimensional data information distribution space, which can quickly find high-dimensional data information. It is mainly aimed at the repeated use of super-distribution planes with mutually perpendicular distribution coordinate axes, so that the data information distribution space is divided into two major components. Among them, the steps of constructing Kd-tree are:

[0057] 1. Determine the partition domain. Calculate the variance of each dimension in the data space, and the partition domain represents the dimension with the largest variance.

[0058] 2. Divide the space. Sort the calculated data from small to large according to the partition domain. After finding the middle node, divide the data space into two subspaces according to the partition domain.

[0059] 3. Continue recursively until the subspace contains no elements. If the subspace contains only a single element, partition it based on the first dimension of the element.

[0060] Specifically, the FLANN algorithm first analyzes the generated Kd-tree, then generates an index structure for the corresponding data information. Based on this data structure, feature point search is primarily based on comparing the Euclidean distance of the point, i.e., finding the closest Euclidean distance to the query point. Matching point pairs are determined based on the ratio of the nearest neighbor to the next nearest neighbor. After feature point matching, the overlapping area is obtained, and dynamic programming is used to find the optimal seam line in the overlapping area. The steps for constructing the seam line are as follows:

[0061] 1. Initialize the energy matrix. Each column of pixels has a corresponding seam line. Starting from the first row of pixels and continuing to the last row, we can get the standard value of the pixels corresponding to the seam line.

[0062] 2. Locate the direction of the stitching line. Starting from the second row, compare the stitching line energy values ​​of the current pixel with the three adjacent pixels in the next row, as shown in Equation (4).

[0063] (4)

[0064] in, For the current pixel The energy through the seam is at its lowest value. Indicates the current pixel At the same time, the corresponding upstream pixels are recorded, and the minimum criterion value point is regarded as the direction of the suture line.

[0065] 3. Repeat the calculation in step 2 until the last row of pixels is calculated, and take the one with the smallest criterion value as the best stitching line.

[0066] Specifically, in step S2, the video frame fusion module, after obtaining the optimal stitching line calculated by the frame stitching parameter calculation module, quickly stitches the subsequent video frames of the multiple video streams together into a subsequent wide-angle video stream based on the stitching line. Using the information obtained from the initialization template frame omits many steps in the subsequent video frame stitching, significantly reducing video stitching time.

[0067] Specifically, the embodiment of the present invention also provides a method for detecting the position where the human eye focuses its attention. By simulating the visual characteristics of the human eye, the detection weight is dynamically adjusted to give priority to the important areas in the image. Figure 3 As shown, the human eye attention focus position detection module includes an attention recognition module and a dynamic weight adjustment module.

[0068] Specifically, in step S3, the attention recognition module inputs the preprocessed image or video stream information, and outputs the key area recognition results in the image that simulate the visual characteristics of the human eye. By analyzing the image texture, moving targets and visual blind spot characteristics, the areas that need priority attention are located.

[0069] Specifically, the sight point correction and attention recognition are first performed. The visual area is divided into 48 areas with 6 rows and 8 columns, and a calibration point is set in the center of each area. The tester looks at these calibration points in turn, and the system records the sight point error of each point. The error is decomposed into X-axis and Y-axis components, and the error value of each calibration point is calculated separately. The radial basis function (RBF) interpolation method is used to construct a two-dimensional error surface, and the sight point of any visual area is dynamically corrected. The reflection position of four infrared LED spots on the cornea ( ) and its projection ( ), the pupil center is determined by the cross ratio invariance. The cross ratio is defined as shown in formula (5).

[0070] (5)

[0071] Among them, | |Expressed as the absolute value of the distance between two points.

[0072] Specifically, the error caused by corneal non-planarity is then corrected. The actual spot position is converted to Mapping to virtual plane position , as shown in formula (6).

[0073] (6)

[0074] in, represents the reference position on the retina (i.e. the intersection of the optical axis of the human eye and the retina), Through the pupil center The condition for coincidence with the virtual light spot is calculated as shown in formula (7).

[0075] (7)

[0076] Specifically, the error between the actual and predicted sight points is measured at each calibration point, decomposed into X / Y components, and vector correction is performed. The global error surface is constructed through Gaussian radial basis interpolation, as shown in Equation (8).

[0077] (8)

[0078] Specifically, the interpolation function is a linear combination of the basis functions, as shown in formula (9).

[0079] (9)

[0080] Determine weights by solving a system of linear equations Realize error compensation at any position.

[0081] Specifically, in step S3, the dynamic weight adjustment module takes as input the key area recognition results and outputs a dynamically adjusted detection weight allocation scheme. This module automatically increases attention to blind spots or high-speed moving targets based on the complexity of the real-time scene, ensuring that the system can still accurately capture important targets in complex environments. Based on a convolutional neural network, the module calculates the key areas in the image and allocates attention to those areas that may contain important targets. The attention weights output by this module can be expressed as shown in Equation (10).

[0082] (10)

[0083] in, is the image point The attention weight, is the visual attention response computed by the network.

[0084] Specifically, the embodiment of the present invention also provides traffic sign detection based on dynamic pruning, which optimizes the calculation path of the deep learning network model through the dynamic pruning strategy and dynamically adjusts the network structure according to the image complexity during the real-time detection process. Figure 4 As shown in FIG, the dynamic pruning traffic sign detection module includes a pruning evaluation module and a network structure optimization module.

[0085] Specifically, in step 4, the pruning evaluation module inputs the feature map information of the current frame image, and outputs the importance score of each computing unit (such as channel, neuron) in the deep learning model. Redundant computing paths are screened by real-time analysis of feature activation intensity.

[0086] Specifically, feature importance evaluation is first performed. The pruning evaluation module performs real-time importance scoring on feature channels or regions in the neural network, measuring their contribution by activation strength or gradient sensitivity, as shown in Equation (11).

[0087] (11)

[0088] in, For the The importance score of each channel, Indicates the spatial position of the channel The activation value of is the feature map size.

[0089] Specifically, in step 4, the input of the network structure optimization module is the importance scoring result, and the output is a dynamically adjusted lightweight network structure. The key feature extraction path is adaptively retained according to the image complexity, which significantly reduces the computational load while ensuring detection accuracy, ensuring a balance between real-time performance and resource efficiency.

[0090] Specifically, dynamic pruning decisions are first made. The network structure optimization module dynamically selects the channels and regions to be retained based on the importance scores and adjusts the computational load of the deep learning network model based on the image features. , only retain the parts with scores higher than the threshold, such as retaining channels , , where the threshold It can be adaptively adjusted according to the complexity of the input image. When the scene is more complex, the threshold is increased to reduce the computational burden.

[0091] Secondly, sparse computation and acceleration are performed. After pruning, convolution is performed only on the retained channels, skipping the redundant parts. The computational reduction ratio is shown in Equation (12).

[0092] (12)

[0093] in is the total number of channels, is the indicator function.

[0094] Specifically, the embodiment of the present invention also provides a detection result output. Combined with the deep learning network model optimized in step S1-step S4, the final detection result is output. Figure 5 As shown, the detection result output module includes a result integration module and a visualization transmission module.

[0095] Specifically, in step S5, the result integration module inputs the optimized network detection information and outputs structured traffic sign data, including target category, location coordinates, size, and confidence score. Post-processing optimization is then performed to remove redundant and low-quality detection frames through non-maximum suppression and confidence filtering. The result integration module combines the deep learning network model optimized in the previous steps to output the final detection results. The detection results include the category, location, and confidence score of the identified traffic sign. This information is shown in Equation (13).

[0096] (13)

[0097] in, Indicates the detected traffic sign category, is the target's position coordinate, is the width and height of the target, is the detection confidence of the target.

[0098] Specifically, post-processing optimization involves non-maximum suppression (NMS) and confidence filtering. NMS filters out highly overlapping detection boxes, retaining the most confident ones. Confidence filtering eliminates detections with low confidence levels to reduce false positives. In step S5, the visualization transmission module overlays the detection results onto the original image, annotating them with categories and confidence levels, and then transmits them to the onboard system or cloud platform via ROS or HTTP protocols.

[0099] Specifically, an embodiment of the present invention provides an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, which aims to significantly improve the perception and processing capabilities of wide-angle targets in complex dynamic scenes. By drawing on the bionic structure of nature and adopting an optimization strategy of multi-perspective fusion and distortion correction, the viewing angle limitations of traditional visual detection systems are broken through, and excellent detection accuracy is maintained in an ultra-wide-angle field of view, exceeding the conventional viewing angle range. By introducing an intelligent focusing mechanism based on the optimization of human eye visual characteristics, this mechanism dynamically adjusts the target detection weight, imitates the focusing characteristics of the human eye, and gives priority to key elements in the target area, especially effectively improving the sensitivity of target detection in visual blind spots. This mechanism is combined with a deep neural network to achieve adaptive learning and optimization of visual attention distribution, significantly improving the detection accuracy and real-time response capability of the system, and is particularly suitable for highly complex environments such as drone monitoring, intelligent security and autonomous driving. To address the high consumption of computing resources in complex scenarios, an embodiment of the present invention also introduces a dynamic pruning strategy driven by edge devices. This strategy assesses image complexity in real time and flexibly adjusts the computational path of the deep neural network model, optimizing the allocation and use of computing resources. This effectively reduces resource waste in high-complexity image processing and improves the system's computational efficiency and application adaptability. This invention overcomes the limitations of traditional detection systems in wide-angle fields of view and highly dynamic environments, significantly improving the robustness of target perception and detection, and demonstrating enormous innovation potential and application prospects.

[0100] The implementation of each embodiment of the present invention is based on programmed processing performed by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functionalities of each embodiment of the present invention are packaged into various modules. Based on this reality, and in addition to the aforementioned embodiments, embodiments of the present invention provide an ultra-wide-angle traffic sign detection system based on adaptive dynamic pruning. This system is used to implement the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning described in the aforementioned method embodiments.

[0101] The system includes: a data acquisition module for acquiring ultra-wide-angle traffic sign data to be tested, performing preprocessing and video stream splicing and fusion; a traffic sign detection module for inputting the spliced ​​and fused data into a trained traffic sign detection model to obtain traffic sign detection results. Training the traffic sign detection model includes constructing an ultra-wide-angle traffic sign dataset; building a traffic sign detection model, including a human eye focus position detection module for dynamically adjusting detection weights; a dynamic pruning-based detection module for dynamically adjusting the network structure based on image complexity; and a detection result output module for outputting the final detection results. The constructed traffic sign detection model is trained using the constructed ultra-wide-angle traffic sign dataset to obtain a trained traffic sign detection model.

[0102] The ultra-wide-angle traffic sign detection system based on adaptive dynamic pruning provided by the embodiment of the present invention addresses the problem that the detection accuracy in ultra-wide-angle scenarios is limited, especially in the edge areas of the image, there are significant distortion problems, and it is difficult to meet real-time detection requirements when running on edge devices or embedded devices. The system adopts several modules and uses FLANN matching and dynamic programming methods to generate large-angle stitched images to eliminate the influence of ultra-wide-angle distortion. The deep learning network is dynamically pruned by real-time evaluation of feature activation intensity, and structured detection results are output in combination with non-maximum suppression and confidence filtering. This achieves both image correction and real-time requirements under ultra-wide-angle field of view on edge devices, thereby meeting increasingly complex target detection application scenarios.

[0103] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning as proposed in the aforementioned embodiment.

[0104] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the program overcomes the problems of limited detection accuracy in ultra-wide-angle scenes, particularly significant distortion in image edge regions, and difficulty meeting real-time detection requirements when running on edge devices or embedded devices. The program effectively eliminates the effects of ultra-wide-angle distortion, significantly reduces redundant computations, effectively reduces resource waste in high-complexity image processing, improves the system's computational efficiency and application adaptability, and achieves high-precision traffic sign detection in ultra-wide-angle scenes, balancing image correction effects with real-time requirements. It effectively balances detection accuracy and resource consumption on edge devices, meeting the needs of real-time monitoring and autonomous driving applications in complex traffic environments.

[0105] The storage medium can be any non-volatile storage device such as a hard disk, solid-state drive, flash drive, optical disk, etc., which is used to store computer program code and necessary data files. The stored computer program includes: a data acquisition module and a traffic sign detection module.

[0106] Finally, it should be noted that the above specific embodiments are merely representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and is susceptible to numerous variations. Any simple modifications, equivalent variations, and modifications to the above specific embodiments based on the technical essence of the present invention shall be deemed to fall within the scope of protection of the present invention.

Claims

1. An ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, characterized by: include: Obtain the ultra-wide-angle traffic sign data to be tested, perform pre-processing and video stream splicing and fusion; The spliced ​​and fused data is input into a trained traffic sign detection model to obtain a traffic sign detection result; wherein the training of the traffic sign detection model includes: Construct an ultra-wide-angle traffic sign dataset; Build a traffic sign detection model, including: a human eye attention location detection module for dynamically adjusting detection weights; a dynamic pruning-based detection module for dynamically adjusting the network structure according to image complexity; and a detection result output module for outputting the final detection results. The human eye attention focus position detection module includes: an attention recognition module for correcting the sight point, dividing the visual area into 48 areas (6 rows × 8 columns), setting a calibration point at the center of each area, sequentially gazing at these calibration points and recording the sight point error of each calibration point, decomposing the sight point error into X-axis and Y-axis components, calculating the error value of each calibration point respectively, and then using radial basis function interpolation to construct a two-dimensional error surface, dynamically correcting the sight point in any visual area, and obtaining a key area recognition result in the image that simulates the visual characteristics of the human eye; a dynamic weight adjustment module for calculating the key areas in the image based on a convolutional neural network, allocating attention to areas that may contain important targets, and outputting the attention weights of the image points; The constructed ultra-wide-angle traffic sign dataset is used to train the constructed traffic sign detection model to obtain a trained traffic sign detection model.

2. The ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to claim 1, characterized in that: Perform preprocessing and video stream splicing and fusion, including: Preprocessing of ultra-wide-angle traffic sign data, including denoising, normalization, and resizing; The pre-processed data is input into the video stream splicing and fusion module for splicing and fusion to obtain a large-viewing angle video stream.

3. The ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to claim 1, characterized in that: The human eye attention focus position detection module includes an attention recognition module and a dynamic weight adjustment module: The attention recognition module locates the areas that need priority attention by analyzing image texture, moving targets and visual blind spot characteristics; The dynamic weight adjustment module automatically increases the attention to blind spots or high-speed moving targets according to the complexity of the real-time scene.

4. The ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to claim 1, characterized in that: The dynamic pruning-based detection module includes a pruning evaluation module and a network structure optimization module: The pruning evaluation module filters out redundant computation paths by analyzing feature activation strength in real time; The network structure optimization module adaptively retains key feature extraction paths according to image complexity.

5. The ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to claim 1, characterized in that: The detection result output module includes a result integration module and a visual transmission module: The result integration module removes redundant and low-quality detection frames through non-maximum suppression and confidence filtering; The visual transmission module is displayed or stored in real time through the vehicle terminal, cloud platform or security system.

6. The ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to claim 2, characterized in that: The video stream splicing and fusion module also includes: Frame stitching parameter calculation module, used to extract feature points and calculate the optimal stitching line; The video frame fusion module is used to quickly stitch together a large-viewing angle video stream based on the optimal stitching line.

7. Ultra-wide-angle traffic sign detection system based on adaptive dynamic pruning, characterized by: include: The data acquisition module is used to obtain the ultra-wide-angle traffic sign data to be tested, perform pre-processing and video stream splicing and fusion; The traffic sign detection module is used to input the spliced ​​and fused data into the trained traffic sign detection model to obtain traffic sign detection results. The training of the traffic sign detection model includes: Construct an ultra-wide-angle traffic sign dataset; Build a traffic sign detection model, including: a human eye attention location detection module for dynamically adjusting detection weights; a dynamic pruning-based detection module for dynamically adjusting the network structure according to image complexity; and a detection result output module for outputting the final detection results. The human eye attention focus position detection module includes: an attention recognition module for correcting the sight point, dividing the visual area into 48 areas (6 rows × 8 columns), setting a calibration point at the center of each area, sequentially gazing at these calibration points and recording the sight point error of each calibration point, decomposing the sight point error into X-axis and Y-axis components, calculating the error value of each calibration point respectively, and then using radial basis function interpolation to construct a two-dimensional error surface, dynamically correcting the sight point in any visual area, and obtaining a key area recognition result in the image that simulates the visual characteristics of the human eye; a dynamic weight adjustment module for calculating the key areas in the image based on a convolutional neural network, allocating attention to areas that may contain important targets, and outputting the attention weights of the image points; The constructed ultra-wide-angle traffic sign dataset is used to train the constructed traffic sign detection model to obtain a trained traffic sign detection model.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning according to any one of claims 1 to 6 are implemented.

Images