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

Abstract

The invention discloses an ultra-wide-angle traffic sign detection method and system based on adaptive dynamic pruning, and belongs to the field of computer vision and target detection, and the method comprises the steps: obtaining to-be-detected ultra-wide-angle traffic sign data, and carrying out the splicing and fusion; inputting the spliced and fused data into the trained traffic sign detection model to obtain a traffic sign detection result; wherein the training of the traffic sign detection model comprises the following steps: constructing an ultra-wide-angle traffic sign data set; a traffic sign detection model is established and comprises a human eye attention concentration position detection module, a detection module based on dynamic pruning and a detection result output module. And training the constructed traffic sign detection model by using the constructed ultra-wide-angle traffic sign data set to obtain a trained traffic sign detection model. According to the method, the image complexity is evaluated in real time, the calculation path of the model is flexibly adjusted, the allocation and use of calculation resources are optimized, and the sensing and processing capability on a large-view-angle target in a complex dynamic scene is improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and object detection, and particularly 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 fields such as drone monitoring, autonomous driving, and intelligent security, the demand for large-angle and high-precision object detection has become increasingly urgent. Existing object detection technologies, especially algorithms based on deep learning frameworks such as YOLOv8, although perform excellently in the standard view, still have some technical challenges in complex dynamic scenarios.

[0003] First, traditional object detection methods mostly rely on single-view cameras, resulting in that in a complex traffic environment, a single-view camera is difficult to comprehensively cover all objects, thus generating detection blind spots. At the same time, the detection accuracy in the ultra-wide-angle scenario is limited. Especially in the image edge area, there are significant distortion problems. Although existing distortion correction technologies have been applied, it is still difficult to completely eliminate the image distortion, resulting in a significant decrease in object detection accuracy in the edge area. For example, although some algorithms introduce wide-angle distortion correction technologies, in the ultra-wide-angle field of view, the image quality and object detection effect still have obvious deficiencies and cannot effectively overcome the perspective limitation. Second, the current traffic sign detection system usually relies on deep neural networks for object recognition. Although high detection accuracy can be obtained, the consumption of computing resources is huge. Especially when running on edge devices or embedded devices, it is difficult to meet the requirements 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 for the above problems. Summary of the Invention

[0005] The object of the present invention is to provide an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning for the problems that the detection accuracy in the ultra-wide-angle scenario is limited, especially in the image edge area, there are significant distortion problems, and it is difficult to meet the real-time detection requirements when running on edge devices or embedded devices. By using the Fast Library for Approximate Nearest Neighbors (FLANN) matching and dynamic programming method to generate a large-angle stitched image, eliminating the influence of ultra-wide-angle distortion, dynamically pruning the deep learning network by real-time evaluating the feature activation intensity, and combining non-maximum suppression and confidence filtering to output a structured detection result, it realizes both image correction in the ultra-wide-angle field of view and real-time requirements on edge devices, thus meeting the increasingly complex object detection application scenarios.

[0006] According to one aspect of the present specification, an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning is provided, including:

[0007] Obtain ultra-wide-angle traffic sign data to be measured, and perform preprocessing and video stream stitching and fusion;

[0008] Input the stitched and fused data into the trained traffic sign detection model to obtain the traffic sign detection result; wherein, the training of the traffic sign detection model includes:

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

[0010] Build a traffic sign detection model, including: a human eye attention concentration position detection module for dynamically adjusting the detection weight; a detection module based on dynamic pruning for dynamically adjusting the network structure according to the image complexity; a detection result output module for outputting the final detection result;

[0011] Use the constructed ultra-wide-angle traffic sign data set to train the built traffic sign detection model to obtain a trained traffic sign detection model.

[0012] Further, the preprocessing and video stream stitching and fusion include:

[0013] Perform preprocessing on the ultra-wide-angle traffic sign data, including denoising, normalization, and size adjustment;

[0014] Input the preprocessed data into the video stream stitching and fusion module for stitching and fusion to obtain a large-view video stream.

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

[0016] The attention recognition module locates the area that needs to be prioritized by analyzing image texture, moving objects, and visual blind spot features;

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

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

[0019] The pruning evaluation module screens redundant calculation paths by analyzing the feature activation intensity in real time;

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

[0021] Further, the detection result output module includes a result integration module and a visualization transmission module:

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

[0023] The visualization transmission module is displayed or stored in real time through a vehicle-mounted terminal, a cloud platform or a security system.

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

[0025] A frame stitching parameter calculation module for feature point extraction and calculation of the best stitching line;

[0026] A video frame fusion module for quickly stitching a large-view video stream according to the best stitching line.

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

[0028] A data acquisition module for acquiring ultra-wide-angle traffic sign data to be measured, performing preprocessing and video stream stitching and fusion;

[0029] A traffic sign detection module for inputting the stitched and fused data into a trained traffic sign detection model to obtain a traffic sign detection result; wherein, the training of the traffic sign detection model includes:

[0030] Constructing an ultra-wide-angle traffic sign data set;

[0031] Building a traffic sign detection model, including: a human eye attention concentration position detection module for dynamically adjusting the detection weight; a detection module based on dynamic pruning for dynamically adjusting the network structure according to the image complexity; a detection result output module for outputting the final detection result;

[0032] Training the built traffic sign detection model with the constructed ultra-wide-angle traffic sign data set to obtain a trained traffic sign detection model.

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

[0034] According to one aspect of the present specification, there is provided a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps of the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning when executed by a processor.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. The present invention extracts feature points of multiple video streams through the Speeded Up Robust Features (SURF) algorithm, generates a large - view stitching image using FLANN matching and dynamic programming method, and eliminates the influence of ultra - wide - angle distortion.

[0037] 2. The present invention utilizes the characteristics of human eye vision, dynamically corrects the line - of - sight landing error through radial basis function interpolation and cross - ratio invariance, identifies key regions in the image and assigns attention weights, preferentially focuses on blind areas and moving targets, and then dynamically prunes the deep - learning network by real - time evaluating the feature activation intensity, adaptively retains key computational paths, significantly reduces the redundant computational amount, effectively reduces resource waste in high - complexity image processing, and improves the computational efficiency and application adaptability of the system.

[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 an ultra - wide - angle scene, taking into account both the image correction effect and real - time requirements, effectively balancing detection accuracy and resource consumption on edge devices, and meeting the real - time monitoring and autonomous driving application requirements in complex traffic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0040] Figure 1 is the flowchart of the method of the embodiment of the present invention;

[0041] Figure 2 is the flowchart of multi - view video stream stitching and fusion of the embodiment of the present invention;

[0042] Figure 3 is the structural diagram of the human eye attention - focused position detection module of the embodiment of the present invention;

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

[0044] Figure 5 is the structural diagram of the detection result output module of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0046] As Figure 1 shown, the embodiment of the present invention provides an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, including: Step S1, collecting the video stream I that needs to be detected for traffic signs and performing preprocessing; Step S2, multi-view video stream stitching and fusion, inputting the preprocessed multiple video streams for image stitching and alignment to obtain a large-view image; Step S3, detecting the position where the human eye's attention is concentrated, dynamically adjusting the detection weight by simulating the visual characteristics of the human eye, and giving priority to focusing on 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 the 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 result, and combining the optimized deep learning network model in the previous steps to output the final detection result.

[0047] Specifically, the preprocessing of the video stream I in the embodiment of the present invention includes denoising, normalization processing, and image size adjustment 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 multi-view video stream stitching and fusion. By inputting the preprocessed multiple video streams for image stitching and alignment, a large-view image is obtained. As Figure 2 shown, the frame stitching parameter calculation module includes a feature extraction unit and a suture calculation unit. The input of the feature extraction unit is the first-frame image information of multiple video streams, and the output is the extracted feature points. The input of the suture calculation unit is the feature points of the first-frame images of multiple video streams, and the output is the best suture of the large-view image formed after the matching of the feature points of multiple images; the input of the video frame fusion module is the best suture calculated by the frame stitching parameter calculation module, and the output is the subsequent large-view video stream quickly stitched according to the suture.

[0049] Specifically, in Step S2, the frame stitching parameter calculation module extracts image feature points based on the feature extraction unit from the first-frame image information of multiple video streams, and then calculates the best suture of multiple images through the suture calculation unit, and finally outputs a large-view 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 it belongs to is shown in Equation (1).

[0050] (1)

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

[0052] (2)

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

[0054] (3)

[0055] Wherein, is the Gaussian pyramid image at the scale of . and represent the second-order partial derivatives of the image at the scale of , represents the mixed derivative of the image at the scale of . In actual calculation, it is approximately accelerated by the BoxFilter.

[0056] Specifically, the suture calculation unit first uses the FLANN algorithm for feature point matching. FLANN is an operation method for matching correspondence based on the Kd-tree. The Kd-tree is a balanced binary tree, mainly used in dividing the K-dimensional data information distribution space, and can complete the rapid search and discovery of high-dimensional data information. It is mainly for the repeated use of the super-distribution plane with mutually perpendicular distribution coordinate axes, so that the data information distribution space is divided into two major components. Among them, the steps to construct the Kd-tree are as follows:

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

[0058] 2. Divide the space. Sort the calculated data from small to large according to the division domain, find the middle node, and then divide the data space into two sub-spaces according to the division domain.

[0059] 3. Continuously recurse until the sub-space contains no elements. If the sub-space only has a single element, it is divided according to the first dimension of the element.

[0060] Specifically, the FLANN algorithm first conducts research and analysis on the generated Kd-tree, then generates an index composition structure for the corresponding data information, and then, with reference to the data information composition structure, the feature point search is mainly obtained by comparing the points with the Euclidean distance, that is, finding the point with the closest Euclidean distance to the query point. The confirmation of the matching point pairs is determined based on the ratio of the nearest neighbor to the second nearest neighbor. After the feature point matching, the overlapping area can be obtained, and the dynamic programming method is used in the overlapping area to obtain the optimal suture line. The steps to construct the suture line are as follows:

[0061] 1. Initialization of the energy matrix. There is a corresponding suture line for each column of pixels. Taking the pixel points in the first row as the starting points until the last row, the criterion values of the pixel points corresponding to the suture line can be obtained.

[0062] 2. Locate the advancing direction of the suture line. Starting from the second row, compare the current pixel with the suture line energy values of the 3 adjacent points in the next row. The calculation is shown in Equation (4).

[0063] (4)

[0064] Among them, is the minimum energy value of the current pixel passing through the seam, represents the energy value of the current pixel . At the same time, record the corresponding upper-row pixels and take the point with the minimum criterion value as the direction of the suture line.

[0065] 3. Continuously repeat the calculation in step 2 until the pixels in the last row are calculated, and take the one with the minimum criterion value as the optimal suture line.

[0066] Specifically, in step S2, after the video frame fusion module obtains the optimal suture line calculated by the frame stitching parameter calculation module, it quickly stitches the subsequent video frames of multiple video streams into a subsequent large-view video stream according to the suture line. Using the information obtained from the initialized template frame omits many steps for the subsequent video frame stitching, greatly reducing the video stitching time.

[0067] Specifically, the embodiment of the present invention also provides the detection of the position where the human eye's attention is concentrated. 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. As Figure 3 shown, the module for detecting the position where the human eye's attention is concentrated includes an attention recognition module and a dynamic weight adjustment module.

[0068] Specifically, in step S3, the input of the attention recognition module is the preprocessed image or video stream information, and the output is the recognition result of the key areas in the image that simulate the visual characteristics of the human eye. By analyzing the image texture, moving targets, and visual blind spot features, the areas that need to be prioritized for attention are located.

[0069] Specifically, first, the correction of the gaze landing point and attention recognition are performed. The visible area is divided into 48 regions of 6 rows × 8 columns, and calibration points are set at the center of each region. The tester fixates on these calibration points in sequence, and the system records the gaze landing point error of each point. The error is decomposed into X-axis and Y-axis components, and the error values of each calibration point are calculated respectively. The radial basis function (RBF) interpolation method is used to construct a two-dimensional error surface to dynamically correct the gaze landing point of any visible area. Using the reflection positions of four infrared LED spots on the cornea ( ), and their projections ( ), the pupil center is determined by the cross-ratio invariance. The cross-ratio is defined as shown in Equation (5).

[0070] (5)

[0071] where, | | represents the absolute value of the distance between two points.

[0072] Specifically, then the error caused by the non-planarity of the cornea is corrected. The actual spot position is mapped to the virtual plane position through a formula, as shown in Equation (6).

[0073] (6)

[0074] where, represents the reference position on the retina (i.e., the intersection of the optical axis of the human eye and the retina), is calculated by the condition that the pupil center coincides with the virtual spot, as shown in Equation (7).

[0075] (7)

[0076] Specifically, finally, the error between the actual and predicted gaze landing points is measured at each calibration point, decomposed into X / Y components, and vector correction is performed. A 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 each basis function, as shown in Equation (9).

[0079] (9)

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

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

[0082] (10)

[0083] where is the attention weight of image point , is the visual attention response calculated through the network.

[0084] Specifically, the embodiment of the present invention also provides traffic sign detection based on dynamic pruning. By means of a dynamic pruning strategy, the calculation path of the deep learning network model is optimized, and the network structure is dynamically adjusted according to the image complexity during real-time detection. As Figure 4 shown, the dynamic pruning traffic sign detection module includes a pruning evaluation module and a network structure optimization module.

[0085] Specifically, in step 4, the input of the pruning evaluation module is the feature map information of the current frame image, and the output is the importance score of each calculation unit (such as channels, neurons) in the deep learning model, and redundant calculation paths are screened by real-time analysis of the feature activation intensity.

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

[0087] (11)

[0088] where is the importance score of the th channel, represents the activation value of this channel at the spatial position , is the size of the feature map.

[0089] Specifically, in step 4, the input of the network structure optimization module is the importance scoring result, and the output is the lightweight network structure after dynamic adjustment. It adaptively retains the key feature extraction paths according to the image complexity, significantly reducing the computational load while ensuring the detection accuracy, and ensuring the balance between real-time performance and resource efficiency.

[0090] Specifically, dynamic pruning decision is first performed. The network structure optimization module dynamically selects the channels and regions to be retained according to the importance scoring, and adjusts the computational amount of the deep learning network model according to the image features. A threshold is set , and only the part with a score higher than the threshold is retained. For example, channels , are retained, where the threshold can be adaptively adjusted according to the complexity of the input image. When the scene is relatively complex, the threshold is increased to reduce the computational burden.

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

[0092] (12)

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

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

[0095] Specifically, in step S5, the input of the result integration module is the optimized network detection information, and the output is the structured traffic sign data, including the target category, position coordinates, size, and confidence score. Then, post-processing optimization is performed to eliminate redundant and low-quality detection frames through non-maximum suppression and confidence filtering. The result integration module combines the optimized deep learning network model in the previous steps to output the final detection result. The detection result includes the category, position, and confidence of the recognized traffic sign. This information is shown in Equation (13).

[0096] (13)

[0097] Where, represents the detected traffic sign category, are the position coordinates of the target, are the width and height of the target, is the detection confidence of this target.

[0098] Specifically, the post-processing optimization is divided into non-maximum suppression (NMS) and confidence filtering. Non-maximum suppression screens the detection boxes with high overlap rates and retains the target box with the highest confidence; confidence filtering eliminates the detection results with too low confidence to reduce false alarms. In step S5, the visualization transmission module superimposes the detection results on the original image, annotates the category and confidence, and transmits them to the vehicle-mounted system or the cloud platform through the ROS or HTTP protocol.

[0099] Specifically, the embodiment of the present invention provides an ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, aiming to significantly improve the perception and processing ability of large-view targets in complex dynamic scenarios. By drawing on the bionic structure of nature, an optimization strategy of multi-view fusion and distortion correction is adopted to break through the perspective limitation of the traditional vision detection system and maintain excellent detection accuracy in the ultra-wide-angle field of view, exceeding the conventional perspective range. By introducing an intelligent focusing mechanism optimized based on the human eye vision characteristics, this mechanism dynamically adjusts the target detection weight, imitates the focusing characteristics of the human eye, and preferentially focuses on the key elements in the target area, especially effectively improving the sensitivity of target detection in the visual blind area. This mechanism combines with the deep neural network to achieve adaptive learning and optimization of the visual attention distribution, significantly improving the detection accuracy and real-time response ability of the system, and is especially suitable for high-complexity environments such as drone monitoring, intelligent security, and autonomous driving. To address the problem of high consumption of computing resources in complex scenarios, the embodiment of the present invention also introduces a dynamic pruning strategy driven by edge devices, which can evaluate the image complexity in real time and flexibly adjust the calculation path of the deep neural network model, optimize the allocation and use of computing resources, effectively reduce the resource waste in high-complexity image processing, and improve the computing efficiency and application adaptability of the system. The present invention breaks through the limitations of the traditional detection system in the wide-angle field of view and high-dynamic environment, significantly improves the robustness of target perception and detection, and shows great innovation potential and application prospects.

[0100] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functions. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this actual situation, on the basis of the above embodiments, the embodiment of the present invention provides an ultra-wide-angle traffic sign detection system based on adaptive dynamic pruning, which is used to execute the ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning in the above method embodiments.

[0101] The system includes: a data acquisition module for acquiring ultra-wide-angle traffic sign data to be measured, performing preprocessing and video stream stitching and fusion; a traffic sign detection module for inputting the stitched 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 concentration position detection module for dynamically adjusting the detection weight; a detection module based on dynamic pruning for dynamically adjusting the network structure according to the image complexity; a detection result output module for outputting the final detection result. The constructed ultra-wide-angle traffic sign data set is used to train the built traffic sign detection model 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 aims at the problem that the detection accuracy in the ultra-wide-angle scenario is limited, especially in the image edge area, there are significant distortion problems, and it is difficult to meet the real-time detection requirements when running on edge devices or embedded devices. It adopts several modules to generate a large-view stitching image by using FLANN matching and dynamic programming method to eliminate the influence of ultra-wide-angle distortion. The deep learning network is dynamically pruned by real-time evaluating the feature activation intensity, and the structured detection result is output by combining non-maximum suppression and confidence filtering, realizing both image correction in the ultra-wide-angle field of view and real-time requirements on edge devices, so as to meet the increasingly complex target detection application scenarios.

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

[0104] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it overcomes the problems that the detection accuracy in the ultra-wide-angle scenario is limited, especially in the image edge area, there are significant distortion problems, and it is difficult to meet the real-time detection requirements when running on edge devices or embedded devices. It effectively eliminates the influence of ultra-wide-angle distortion, significantly reduces the redundant calculation amount, effectively reduces the resource waste in high-complexity image processing, improves the computing efficiency and application adaptability of the system, realizes high-precision traffic sign detection in the ultra-wide-angle scenario, takes into account the image correction effect and real-time requirements, effectively balances the detection accuracy and resource consumption on edge devices, and meets the real-time monitoring and autonomous driving application requirements in complex traffic environments.

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

[0106] Finally, it should be noted that the above specific embodiments are only relatively representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and there can be many variations. Any simple modification, equivalent change and modification made to the above specific embodiments based on the technical essence of the present invention shall be considered as falling within the protection scope of the present invention.

Claims

1. An ultra-wide-angle traffic sign detection method based on adaptive dynamic pruning, characterized in that Including: Obtain ultra-wide-angle traffic sign data to be measured, perform preprocessing and video stream stitching and fusion; Input the stitched and fused data into the trained traffic sign detection model to obtain the traffic sign detection result; wherein, the training of the traffic sign detection model includes: Construct an ultra-wide-angle traffic sign data set; Build a traffic sign detection model, including: a human eye attention concentration position detection module for dynamically adjusting the detection weight; a detection module based on dynamic pruning for dynamically adjusting the network structure according to the image complexity; a detection result output module for outputting the final detection result; Use the constructed ultra-wide-angle traffic sign data set to train the built 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 stitching and fusion, including: Preprocess the ultra-wide-angle traffic sign data, including denoising, normalization and size adjustment; Input the preprocessed data into the video stream stitching and fusion module for stitching and fusion to obtain a large-view 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 concentration position detection module includes an attention recognition module and a dynamic weight adjustment module: The attention recognition module locates the area that needs to be prioritized by analyzing image texture, moving targets and visual blind spot features; The dynamic weight adjustment module automatically enhances the attention to blind spots or high-speed moving targets according to the real-time scene complexity.

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

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

6. The method for ultra-wide-angle traffic sign detection based on adaptive dynamic pruning according to claim 1, characterized in that, The video stream stitching and fusion module further includes: A frame stitching parameter calculation module for performing feature point extraction and calculating the best stitching line; A video frame fusion module for quickly stitching a large-view video stream according to the best stitching line.

7. The ultra-wide-angle traffic sign detection system based on adaptive dynamic pruning is characterized in that Including: A data acquisition module for obtaining ultra-wide-angle traffic sign data to be measured, performing preprocessing and video stream stitching and fusion; A traffic sign detection module for inputting the stitched and fused data into the trained traffic sign detection model to obtain the traffic sign detection result; wherein, the training of the traffic sign detection model includes: Construct an ultra-wide-angle traffic sign data set; Build a traffic sign detection model, including: a human eye attention concentration position detection module for dynamically adjusting the detection weight; a detection module based on dynamic pruning for dynamically adjusting the network structure according to the image complexity; a detection result output module for outputting the final detection result; Use the constructed ultra-wide-angle traffic sign data set to train the built traffic sign detection model to obtain a trained traffic sign detection model.

8. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, 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 the 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