A highly scalable feature extraction network detection method for efficient detection of road objects

Abstract

The present invention provides a highly scalable feature extraction network detection method for efficiently detecting road objects, and relates to the technical field of autonomous driving road object detection. The method comprises: inputting a captured road image into a spatial pyramid parallel pooling module to determine an output feature map; inputting the output feature map into a TDGL parallel branch convolution to determine a feature map with higher resolution; using the input features extracted by GL features, performing batch normalization processing and affine transformation to determine the normalized results and the output features of the affine transformation; performing Leak Relu back-propagation training on the output features to determine the partial derivative of the loss function L; obtaining the bounding box and confidence of each target; and determining the adjusted target confidence by using the Soft-NMS non-maximum suppression algorithm. According to the present invention, spatial hierarchical information can be effectively processed and the increase in computational burden can be avoided, thereby improving the detection accuracy of road objects.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving road object detection, and in particular to a highly scalable feature extraction network detection method for efficiently detecting road objects. Background Art

[0002] As a fundamental and critical research area, object detection technology is driving technological innovation in multiple fields, including autonomous driving and intelligent monitoring. With the continuous development of convolutional neural network (CNN) technology, the performance of object detection algorithms has been significantly improved. However, the substantial increase in computational burden makes it difficult to meet the real-time requirements of autonomous driving scenarios. Existing methods use a pyramid dilated convolution module in parallel with a RESA module to enhance the model's receptive field to enrich and extract global spatial feature information, but the detection speed is poor. Another approach uses the global attention MSADark module and the location attention weighted feature fusion network LAFFN to enhance the network's feature representation capabilities, but the limitation lies in the huge computational burden.

[0003] Therefore, in related technologies, edge devices have a heavy computational burden, lack effective utilization of the spatial relationships and contextual information of the target, and are unable to adaptively adjust the convolution size according to demand, resulting in a waste of computing resources and difficulty in improving the detection accuracy of road objects.

[0004] The information disclosed in the background technology section of this application is only intended to deepen the understanding of the general background technology of this application, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0005] The present invention provides a highly scalable feature extraction network detection method for efficiently detecting road objects, which can solve the technical problems of heavy computational burden on edge devices and lack of effective utilization of the spatial relationship and contextual information of targets in road detection tasks.

[0006] According to the present invention, a highly scalable feature extraction network detection method for efficiently detecting road objects is provided, comprising: inputting a captured road image into a spatial pyramid parallel pooling module to determine an output feature map; inputting the output feature map into a TDGL parallel branch convolution to determine a feature map with higher resolution, wherein the TDGL parallel branch convolution has three parallel branches, a GLConv convolution layer with a 1×1 convolution kernel, and two convolution layers with stacked 3×3 convolution kernels, with a shortcut spanning method used at the end to collect features; performing batch normalization and affine transformation on the input features of the feature map with higher resolution using GL feature extraction to determine the normalized results and the output features of the affine transformation; performing Leak Relu back-propagation training on the output features to determine the partial derivative of the loss function L; obtaining the bounding box and confidence of each target through the CIoU loss function; and determining the adjusted target confidence through the Soft-NMS non-maximum suppression algorithm based on the bounding box and the confidence.

[0007] Furthermore, the captured road image is input into a spatial pyramid parallel pooling module to determine an output feature map, including: converting the captured road image into a feature vector of fixed length to obtain an input feature map; performing maximum pooling processing on the input feature map, and then performing adaptive maximum pooling and adaptive average pooling respectively, and then splicing to determine a spliced ​​feature map; performing convolution on the spliced ​​feature map and applying an activation function to determine an output feature map.

[0008] Furthermore, after performing maximum pooling processing on the input feature map, adaptive maximum pooling and adaptive average pooling are performed respectively, and then splicing is performed to determine the spliced ​​feature map, including:

[0009] According to the formula

[0010]

[0011]

[0012]

[0013] Determine the concatenated feature map f, where 、 、 、 、 are the results after the 1st, 2nd, 3rd, e-1, and eth maximum pooling processes respectively, and , y is the initial convolution dimension reduction value of the input feature map, k is the size of the maximum pooling window, s is the stride, and p is the padding value. is the result after adaptive maximum pooling processing, is the result after adaptive average pooling processing, For adaptive pooling, the input feature map of any size is converted into a 1×1 feature map, and e is a positive integer.

[0014] Further, convolution is performed on the concatenated feature map and an activation function is applied to determine an output feature map, including:

[0015] According to the formula

[0016]

[0017] Determine the output feature map ,in, is the result of the convolution operation, h is the height of the output feature map, w is the width of the output feature map, and c is the number of output channels. is the number of channels in the feature map, For input channels The c-th output channel weight matrix element corresponding to the position, is the value of the feature map at the corresponding position, is the bias term of output channel c, ≤ , and c, and are all positive integers, is the activation function.

[0018] Furthermore, the TDGL parallel branch convolution includes:

[0019] According to the formula

[0020]

[0021] Get the output of TDGL parallel branch convolution , where x and y are the x-axis and y-axis of the output feature map, z is the offset of the y-axis, and q is the offset of the x-axis. and is the kernel size, , , is the input of the branch convolution, and W is the convolution filter.

[0022] Furthermore, the input features extracted by GL feature extraction of the feature map with higher resolution are subjected to batch normalization processing and affine transformation, and the normalization result and the output features of the affine transformation are determined, including:

[0023] According to the formula

[0024]

[0025]

[0026]

[0027]

[0028] Determine the normalized result of the i-th sample And the output features of the affine transformation of the i-th sample ,in, is the input feature of the i-th sample, is the mean of the training batch data features, is the variance of the data feature, is a small positive value, m is the number of samples in a batch, i≤m, and both i and m are positive integers.

[0029] Furthermore, a Leak Relu back-propagation training neural network is performed on the output features to determine the partial derivative of the loss function L, including:

[0030] Apply the ReLU activation function to the output feature map of each convolutional layer, according to the formula

[0031]

[0032] Determine the The feature map of layer j ,in, is a nonlinear activation function, For the The u-th input feature map of the layer, is the input feature atlas, * is the convolution operation, is the weight matrix of the convolution kernel, is the bias value;

[0033] According to the formula

[0034]

[0035] Determine the loss function L for the Partial derivative of the layer output ,in, For the The feature map of the layer, For the propagation formula, leak is a small positive constant.

[0036] Furthermore, the CIoU loss function includes:

[0037] According to the formula

[0038]

[0039]

[0040] Determine the CIoU loss function, where is the overlap intersection-union ratio, is the center point of the prediction box, r is the diagonal length of the minimum enclosing box covering the prediction box and the real box, is the center point of the ground truth box, is the weighting coefficient, v is the consistency of the aspect ratio, is the Euclidean distance between the center point of the real box and the center point of the predicted box, and are the length and width of the real frame, and is the length and width of the prediction box.

[0041] Furthermore, determining an adjusted target confidence level based on the bounding box and the confidence level through a Soft-NMS non-maximum suppression algorithm includes:

[0042] According to the formula

[0043]

[0044] Determine the confidence level of target n ,in, is the overlap intersection-union ratio, is the bounding box of target n, is the bounding box of the target with the highest confidence, is the overlap threshold, is the weight threshold.

[0045] Technical effect: According to the present invention, a scheme for constructing a fast and accurate feature extraction module by using a normalized convolution unit is designed, and a highly scalable feature extraction network module is proposed. This module can effectively process spatial hierarchical information and avoid the increase of computational burden by reducing the internal covariate offset ICS, introducing scaling parameters and offset parameters, changing the expansion strategy and adopting grouped convolution, thereby enhancing the ability to distinguish and detect features of different scales in the road. At the same time, it solves the problem that neuron parameters cannot be updated, accelerates the convergence speed and improves the robustness.

[0046] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and not limiting of the present invention. Other features and aspects of the present invention will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] 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. Those skilled in the art can derive other embodiments based on these drawings without inventive efforts.

[0048] Figure 1 A schematic diagram exemplarily illustrates a flow chart of a highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention;

[0049] Figure 2 The TDGL parallel branch convolution structure diagram of the highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention is exemplarily shown;

[0050] Figure 3 The following is an exemplary diagram showing a GL feature extraction batch normalization structure of a highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention;

[0051] Figure 4 The structure diagram of the highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention is exemplarily shown. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 shall fall within the scope of protection of the present invention.

[0053] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0054] Figure 1The flowchart of the highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention is exemplarily shown. The method includes: step S101, inputting the captured road image into the spatial pyramid parallel pooling module to determine the output feature map; step S102, inputting the output feature map into the TDGL parallel branch convolution to determine a feature map with higher resolution, wherein the TDGL parallel branch convolution has three parallel branches, a GLConv convolution layer with a 1×1 convolution kernel, and two convolution layers with stacked 3×3 convolution kernels, and a shortcut spanning method is used to collect features at the end; step S103, performing batch normalization and affine transformation on the input features of the feature map with higher resolution using GL feature extraction, and determining the output features of the normalization result and the affine transformation; step S104, performing Leak on the output features. Relu back-propagation trains the neural network to determine the partial derivative of the loss function L; step S105, obtains the bounding box and confidence of each target through the CIoU loss function; step S106, determines the adjusted target confidence based on the bounding box and the confidence through the Soft-NMS non-maximum suppression algorithm.

[0055] According to one embodiment of the present invention, in step S101 , the captured road image is input into a spatial pyramid parallel pooling module to determine an output feature map.

[0056] According to one embodiment of the present invention, step S101 includes: converting the captured road image into a feature vector of a fixed length to obtain an input feature map; performing maximum pooling processing on the input feature map, and then performing adaptive maximum pooling and adaptive average pooling respectively, and then splicing to determine a spliced ​​feature map; convolving the spliced ​​feature map and applying an activation function to determine an output feature map.

[0057] According to one embodiment of the present invention, a captured road image is input into a spatial pyramid parallel pooling module to convert it into a fixed-length feature vector to ensure data consistency. A pooling operation is then performed to gradually reduce the pooling window size.

[0058] According to one embodiment of the present invention, convolution is performed on the spliced ​​feature map and an activation function is applied to determine an output feature map, including: performing maximum pooling on the input feature map, performing adaptive maximum pooling and adaptive average pooling respectively, and then splicing to determine the spliced ​​feature map, including: determining the spliced ​​feature map f according to formulas (1), (2) and (3),

[0059] (1),

[0060] (2),

[0061] (3),

[0062] in, 、 、 、 、 are the results of the 1st, 2nd, 3rd, e-1, and eth maximum pooling processes respectively, and , y is the initial convolution dimension reduction value of the input feature map, k is the size of the maximum pooling window, s is the stride, and p is the padding value. is the result after adaptive maximum pooling processing, is the result after adaptive average pooling processing, For adaptive pooling, the input feature map of any size is converted into a 1×1 feature map, and e is a positive integer.

[0063] According to one embodiment of the present invention, in formula (1), the padding value can add an additional zero layer at the edge of the input feature map to maintain the size of the output feature map. In formula (2), the significant features in the feature map, such as edges, textures, etc., are retained by adaptive maximum pooling, and adaptive average pooling is used to retain the average features of the image, reduce information loss, and smooth the noise in the image. In formula (3), the results of all pooling operations are spliced ​​along the channel dimension to obtain the spliced ​​feature map. . Minimize redundant operations while effectively capturing features at various scales.

[0064] According to one embodiment of the present invention, convolution is performed on the concatenated feature map and an activation function is applied to determine an output feature map, including: determining the output feature map according to formula (4) ,

[0065] (4),

[0066] in, is the result of the convolution operation, h is the height of the output feature map, w is the width of the output feature map, and c is the number of output channels. is the number of channels in the feature map, For input channels The c-th output channel weight matrix element corresponding to the position, is the value of the feature map at the corresponding position, is the bias term of output channel c, ≤ , and c, and are all positive integers, is the activation function.

[0067] According to one embodiment of the present invention, in formula (4), the SiLU activation function can introduce nonlinearity in the output. The input feature map is convolved with the convolution kernel (weight matrix) to generate a new feature map. The convolution result is then nonlinearly transformed using the SiLU activation function to obtain the output feature map.

[0068] According to one embodiment of the present invention, in step S102, Figure 2 The TDGL parallel branch convolution structure diagram of the highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention is exemplified. The output feature map is input into the TDGL parallel branch convolution, which has three parallel branches. The image detail features are extracted through a GLConv convolution layer with a 1×1 convolution kernel and two convolution layers with 3×3 convolution kernels stacked. The shortcut spanning method is used to collect features at the end. The specific parameters of the TDGL parallel branch convolution are adjusted, and the step value of the control convolution operation is set to 1 by default. The reduction factor of the number of channels in the intermediate layer is calculated. The default scaling factor of the jump connection (shortcut) is 0.1, which improves the efficiency of network information transmission and training convergence. In order to better focus on multi-scale target features, the expansion rate of the convolution in each branch is set differently. The expansion rate of the 3×3 convolution at the last layer of branch 1 is set to 1 to focus on smaller image areas and retain detail features. The expansion rate of the convolution at the last layer of branch 2 is set to 3 to enable it to capture medium-scale features, such as part of the target or background. The expansion rate of the convolution at the last layer of branch 3 is set to 5 to enable it to capture large-scale global information and understand the structure and background of the entire image.

[0069] According to one embodiment of the present invention, the TDGL parallel branch convolution includes: obtaining the output of the TDGL parallel branch convolution according to formula (5): ,

[0070] (5),

[0071] Among them, x and y are the x-axis and y-axis of the output feature map, z is the offset of the y-axis, and q is the offset of the x-axis. and is the kernel size, , , is the input of the branch convolution, and W is the convolution filter.

[0072] According to one embodiment of the present invention, in formula (5), the convolution bias term is ignored for simplicity, and the value of the output feature map is calculated by weighted summation of the convolution kernel and the input feature map. The network structure has multiple parallel branches that process different parts of the input feature map in parallel, thereby capturing a wider range of information and more context while maintaining the same number of parameters, thereby improving detection speed.

[0073] According to one embodiment of the present invention, in step S103, Figure 3 The GL feature extraction batch normalization structure diagram of the highly scalable feature extraction network detection method for efficiently detecting road objects according to an embodiment of the present invention is exemplarily shown. GL feature extraction batch normalization is adopted, and an integrated batch normalization layer is used for the feature number after the two-dimensional convolution layer. A non-zero value is added to the batch processing variance to increase numerical stability. Each instance feature is standardized through centering and scaling operations so that its mean is 0 and the variance is 1, so that the input of each layer remains relatively stable, the internal covariate shift is reduced, and the network is easier to converge, thereby accelerating the convergence speed and improving the generalization ability of the model.

[0074] According to one embodiment of the present invention, the input features extracted by GL feature extraction on the feature map with higher resolution are subjected to batch normalization processing and affine transformation, and the normalization result and the output features of the affine transformation are determined, including: determining the normalization result of the i-th sample according to formulas (6), (7), (8) and (9): And the output features of the affine transformation of the i-th sample ,

[0075] (6),

[0076] (7),

[0077] (8),

[0078] (9),

[0079] in, is the input feature of the i-th sample, is the mean of the training batch data features, is the variance of the data feature, is a small positive value, m is the number of samples in a batch, i≤m, and both i and m are positive integers.

[0080] According to one embodiment of the present invention, in formula (6), A small positive value is used to prevent the denominator from being zero when calculating normalization, usually The default value is , considering that road targets contain diverse target categories and have complex characteristics, Adjust the value to , increase The value will improve the numerical stability, help deal with the differences in image data captured under different environmental conditions, and reduce the impact of noise on the normalization process. In formula (7), the input features of multiple samples are averaged to obtain the mean of the training batch data features. In formula (8), the variance of the input features of multiple samples is calculated to obtain the variance of the data features. In formula (9), although the normalized data helps the stability of training, its distribution characteristics, such as zero mean and unit variance, may not be the most suitable for the feature expression requirements of the current layer. Therefore, the affine Boolean parameter value is adjusted to true, and the scaling parameter and offset parameter are introduced to perform affine transformation on the normalized data, thereby restoring or retaining some characteristics of the original data.

[0081] According to one embodiment of the present invention, in step S104, the Leak Relu back-propagation activation function is used to process the feature map data that has undergone convolution operation and batch normalization, so as to solve the problem that the corresponding neuron parameters cannot be updated when the input is less than 0.

[0082] According to one embodiment of the present invention, performing Leak Relu back-propagation training on the output features to determine the partial derivative of the loss function L includes:

[0083] Apply the ReLU activation function to the output feature map of each convolutional layer and determine the The feature map of layer j ,

[0084] (10),

[0085] in, is a nonlinear activation function, For the The u-th input feature map of the layer, is the input feature atlas, * is the convolution operation, is the weight matrix of the convolution kernel, is the bias value;

[0086] According to formula (11), the loss function L is determined for the Partial derivative of the layer output ,

[0087] (11),

[0088] in, For the The feature map of the layer, For the propagation formula, leak is a small positive constant.

[0089] According to one embodiment of the present invention, in formula (10), after the convolution operation is completed, the ReLU activation function is applied to the output feature map of each convolution layer. The definition of the ReLU function is f ( x )=max(0, x ), setting negative values ​​to 0 and keeping positive values ​​unchanged, thereby introducing nonlinear characteristics and enabling the network to better learn complex patterns.

[0090] According to one embodiment of the present invention, in formula (11), when the Leak Relu input is negative, a smaller negative slope is allowed, thereby alleviating the problem of neuron death and using back propagation as a layer, and the loss function L is obtained for the The partial derivative of the layer output helps alleviate the vanishing gradient problem, allowing the neural network to more effectively propagate gradient information during training, thereby accelerating convergence and improving robustness.

[0091] According to one embodiment of the present invention, in step S105 , the CIoU loss function is used to evaluate the boundary loss, taking into account the shape information of the target box.

[0092] According to one embodiment of the present invention, the CIoU loss function includes: determining the CIoU loss function according to formulas (12) and (13),

[0093] (12),

[0094] (13),

[0095] in, is the overlap intersection-union ratio, is the center point of the prediction box, r is the diagonal length of the minimum enclosing box covering the prediction box and the real box, is the center point of the ground truth box, is the weighting coefficient, v is the consistency of the aspect ratio, is the Euclidean distance between the center point of the real box and the center point of the predicted box, and are the length and width of the real frame, and is the length and width of the prediction box.

[0096] According to one embodiment of the present invention, the CIoU loss function in formula (12) is a metric used to evaluate the degree of overlap between a predicted box and a ground-truth box in object detection tasks. It not only takes into account the overlapping area (measured by Intersection over Union (IoU), where a higher IoU indicates a better degree of overlap), but also introduces the consistency of the box's center point distance and aspect ratio, making the loss more robust to target boxes of different shapes. It also improves the accuracy of the predicted box by focusing on the center of mass to evaluate the degree of overlap and positional relationship between the predicted box and the ground-truth box. In formula (13), the consistency of the aspect ratio between the predicted box and the ground-truth box is measured.

[0097] According to one embodiment of the present invention, in step S106, an adjusted target confidence is determined by a Soft-NMS non-maximum suppression algorithm based on the bounding box and the confidence.

[0098] According to one embodiment of the present invention, based on the bounding box and the confidence, the adjusted target confidence is determined by the Soft-NMS non-maximum suppression algorithm, including: determining the confidence of the target n according to formula (14) ,

[0099] (14),

[0100] in, is the overlap intersection-union ratio, is the bounding box of target n, is the bounding box of the target with the highest confidence, is the overlap threshold, is the weight threshold.

[0101] According to one embodiment of the present invention, in formula (14), the Soft-NMS non-maximum suppression algorithm is used to screen out target frames with confidence levels higher than a threshold, and the target frame with the highest confidence level is selected as the reference frame. This algorithm is executed separately for each category, and the degree of overlap between other target frames and the reference frame is compared. Target frames with an overlap level higher than a certain threshold are removed.

[0102] Figure 4 This diagram illustrates the structure of a highly scalable feature extraction network detection method for efficient road object detection according to an embodiment of the present invention. First, captured road images are fed into a spatial pyramid parallel pooling module, converting them into fixed-length feature vectors. Second, TDGL parallel branch convolution is used to obtain a rich feature representation, generating a higher-resolution feature map. Finally, the Leak ReLU backpropagation activation function is used to process the convolutional and batch-normalized feature map data.

[0103] According to an embodiment of the present invention, a highly scalable feature extraction network detection method for efficiently detecting road objects is proposed by designing a solution for constructing a fast and accurate feature extraction module by using a normalized convolution unit. This module can effectively process spatial hierarchical information and avoid an increase in computational burden by reducing the internal covariate offset (ICS), introducing scaling parameters and offset parameters, changing the expansion strategy, and adopting grouped convolution. This enhances the ability to distinguish and detect features of different scales in the road. At the same time, it solves the problem of neuron parameters being unable to be updated, accelerates convergence, and improves robustness.

[0104] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0105] Those skilled in the art will appreciate that the embodiments of the present invention described above and shown in the accompanying drawings are intended to be illustrative only and are not intended to limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and illustrated in the embodiments. Any variations or modifications may be made to the embodiments of the present invention without departing from the principles described.

Claims

1. A highly scalable feature extraction network detection method for efficiently detecting road objects, characterized in that: include: Input the captured road image into the spatial pyramid parallel pooling module, convert it into a fixed-length feature vector, and obtain the input feature map; After performing maximum pooling on the input feature map, adaptive maximum pooling and adaptive average pooling are performed respectively, and then splicing is performed to determine a spliced ​​feature map, including: According to the formula Determine the concatenated feature map f, where 、 、 、 、 They are 1st, 2nd, 3rd, 、 The result after sub-maximal pooling, and , y is the initial convolution dimension reduction value of the input feature map, k is the size of the maximum pooling window, s is the stride, and p is the padding value. is the result after adaptive maximum pooling processing, is the result after adaptive average pooling processing, Convert the input feature map of any size into a 1×1 feature map for adaptive pooling, and is a positive integer; Convolving the concatenated feature map and applying an activation function to determine an output feature map includes: According to the formula Determine the output feature map ,in, is the result of the convolution operation, h is the height of the output feature map, w is the width of the output feature map, and c is the number of output channels. is the number of channels in the feature map, For input channels The c-th output channel weight matrix element corresponding to the position, is the value of the feature map at the corresponding position, is the bias term of output channel c, , and c, and are all positive integers, is the activation function; Input the output feature map into the TDGL parallel branch convolution to determine a feature map with higher resolution, wherein the TDGL parallel branch convolution has three parallel branches, a GLConv convolution layer with a 1×1 convolution kernel, and two convolution layers with stacked 3×3 convolution kernels. A shortcut leap is used at the end to collect features. The step value controlling the convolution operation is set to 1 by default, the reduction factor for calculating the number of channels in the intermediate layer is set to 8, the scaling factor of the jump connection is set to 0.1 by default, the convolution expansion rate of the last layer of branch 1 is set to 1, the convolution expansion rate of the last layer of branch 2 is set to 3, and the convolution expansion rate of the last layer of branch 3 is set to 5; The TDGL parallel branch convolution includes: According to the formula Get the output of TDGL parallel branch convolution , where x and y are the x-axis and y-axis of the output feature map, z is the offset of the y-axis, and q is the offset of the x-axis. and is the kernel size, , , Is the input of the branch convolution, W is the convolution filter; Performing batch normalization and affine transformation on the input features of the feature map with higher resolution using GL feature extraction, and determining the normalization result and the output features of the affine transformation; Performing Leak Relu back propagation training on the output features to determine the partial derivative of the loss function L; Obtain the bounding box and confidence of each target through the CIoU loss function; According to the bounding box and the confidence level, an adjusted target confidence level is determined by a Soft-NMS non-maximum suppression algorithm.

2. The highly scalable feature extraction network detection method for efficiently detecting road objects according to claim 1, characterized in that: The input features extracted by using GL features on the feature map with higher resolution are subjected to batch normalization processing and affine transformation, and the normalization results and the output features of the affine transformation are determined, including: According to the formula Determine the normalized result of the i-th sample And the output features of the affine transformation of the i-th sample ,in, is the input feature of the i-th sample, is the mean of the training batch data features, is the variance of the data feature, for , is the scaling parameter, is the offset parameter, m is the number of samples in a batch, i≤m, and both i and m are positive integers.

3. The highly scalable feature extraction network detection method for efficiently detecting road objects according to claim 1, characterized in that: Performing Leak Relu back propagation training on the output features to determine the partial derivative of the loss function L includes: Apply the ReLU activation function to the output feature map of each convolutional layer, according to the formula Determine the Feature map of layer j ,in, is a nonlinear activation function, For the The u-th input feature map of the layer, is the input feature atlas, * is the convolution operation, is the weight matrix of the convolution kernel, is the bias value; According to the formula Determine the loss function L for the Partial derivative of the layer output ,in, For the The feature map of the layer, For the propagation formula, leak is a small positive constant.

4. The highly scalable feature extraction network detection method for efficiently detecting road objects according to claim 1, characterized in that: The CIoU loss function includes: According to the formula Determine the CIoU loss function, where is the overlap intersection-union ratio, is the center point of the prediction box, r is the diagonal length of the minimum enclosing box covering the prediction box and the real box, is the center point of the ground truth box, is the trade-off coefficient, To measure aspect ratio consistency, is the Euclidean distance between the center point of the real box and the center point of the predicted box, and are the length and width of the real frame, and is the length and width of the prediction box.

5. The highly scalable feature extraction network detection method for efficiently detecting road objects according to claim 1, characterized in that: Determining an adjusted target confidence level based on the bounding box and the confidence level through a Soft-NMS non-maximum suppression algorithm includes: According to the formula Determine the confidence level of target n ,in, is the overlap intersection-union ratio, is the bounding box of target n, is the bounding box of the target with the highest confidence, is the overlap threshold, is the weight threshold.

6. A computer-readable storage medium, characterized in that Computer program instructions are stored thereon, and when the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images