Automatic driving target detection method in night low-light environment

Abstract

The invention discloses an automatic driving target detection method in a night low-light environment. The method comprises the following steps: constructing an automatic driving target detection model AQ9; according to the AQ9, an LKNet9 network is adopted as a backbone network to extract basic features from an input image, and for local features, night target obstacles are identified; a self-adaptive pooling algorithm is adopted, and the pooling scale and type are automatically adjusted according to the content of the input feature map; a nearest neighbor interpolation up-sampling module in the Head layer is used for DySample sampling, and self-adaptive adjustment is carried out according to the content of the input features; the method comprises the following steps: extracting features in a larger range through convolution operation of enhancing a receptive field, applying an attention mechanism on a feature map after convolution operation, and respectively adjusting channel attention and space attention; a C2f module is adopted, the spatial dimension of the feature map is focused, and the distribution weight of each spatial position is enhanced; and training the AQ9 by using the enhanced KITTI data set, and performing target detection by using the trained AQ9. According to the invention, night target detection can be carried out more efficiently, stably and accurately.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving target detection, and in particular relates to an autonomous driving target detection method in a nighttime low-light environment. Background Art

[0002] In autonomous driving systems, object detection technology plays a vital role, especially in nighttime environments. Autonomous vehicles use object detection to identify and accurately locate specific objects on the road, such as vehicles, pedestrians, traffic signs, and obstacles, to ensure driving safety and achieve fully autonomous driving functions.

[0003] In the nighttime autonomous driving environment, the accuracy of target detection may be significantly challenged. In order to improve the recognition accuracy in low-light conditions, some deep learning models have introduced attention mechanisms, such as CBAM (Convolutional Block Attention Module). Although CBAM is designed as a high-precision attention module, it may not actually bring ideal results. This is mainly reflected in the multi-layer perceptron (MLP) operation in its channel attention module and the convolution operation in the spatial attention module. Moreover, the calculation steps may prolong the inference time of the model, especially in the resource-constrained nighttime autonomous driving environment.

[0004] In addition, effective use of CBAM requires careful tuning of multiple hyperparameters, such as the size of the MLP and the kernel size of the spatial attention convolution. Improper parameter settings can lead to poor performance, and the process of finding the optimal parameter configuration is complex and time-consuming. In practical applications, although CBAM can be relatively simple to integrate into existing CNN architectures, adjustments may need to be made to ensure that the attention mechanism does not affect the training dynamics or cause overfitting, especially when the dataset is small or less diverse.

[0005] While CBAM has shown effectiveness in a variety of tasks, its scalability has not been widely verified when dealing with very large networks and datasets. In these cases, the additional complexity introduced by CBAM may lead to diminishing marginal benefits of performance improvement. At the same time, attention mechanisms such as CBAM are sometimes overly sensitive to noise and outliers in the data, which may cause the model to mistakenly focus on irrelevant features, thus affecting overall performance.

[0006] Therefore, although CBAM can improve the detection accuracy in nighttime autonomous driving environments to a certain extent, it is not an ideal solution. It needs to be optimized according to specific application scenarios and requirements to cope with the low-light challenges in nighttime environments. Summary of the invention

[0007] Purpose of the invention: The present invention proposes a target detection method for autonomous driving in a low-light environment at night, which can perform target detection in a low-light environment more efficiently, stably and accurately.

[0008] Technical solution: The method for detecting an autonomous driving target in a low-light environment at nighttime according to the present invention comprises the following steps:

[0009] (1) Perform data enhancement on the KITTI dataset;

[0010] (2) Construct an autonomous driving target detection model AQ9; the AQ9 uses the LKNet9 network as the backbone network; the LKNet9 network includes a unified representation module Rep, a local convolution module LCB and a large kernel convolution, which extracts basic features from the input image and performs local features to identify target obstacles at night; the pooling module of the AQ9 model is a hybrid pyramid pooling structure SPPF9, which uses an adaptive pooling algorithm to automatically adjust the scale and type of pooling according to the content of the input feature map; the nearest neighbor interpolation upsampling module in the Head layer of the AQ9 model is DySample sampling, which adaptively adjusts the scale and type of pooling according to the content of the input feature map. In the Head layer of the AQ9 model, a receptive field attention convolution RFACBAMConv9 is proposed, which extracts features in a larger range by enhancing the convolution operation of the receptive field, and applies the attention mechanism to the feature map after the convolution operation, and adjusts the channel attention and spatial attention respectively; the Head layer of the AQ9 model adopts the C2f module, which is derived from the C2 module fused with dynamic sparse attention, focuses on the spatial dimension of the feature map, and strengthens the allocation weight of each spatial position; the detection head in the Head layer of the AQ9 model, for low-light environments at night, enhances local attention and enhances the ability to capture important features in a small range;

[0011] (3) Training the autonomous driving target detection model AQ9 constructed in step (2) and adjusting the training parameters of the model until the optimal model weight is obtained;

[0012] (4) Use the trained autonomous driving target detection model AQ9 for detection.

[0013] Furthermore, in step (1), the unified representation module Rep reduces network parameters by reusing features and sharing representations while maintaining high performance. The local convolution module LCB utilizes local convolution operations to improve the network's receptive field and efficiency by point-by-point convolution and local area processing; the large kernel convolution expands the receptive field to capture sparse features; and by using the idea of ​​structural reparameterization, the entire block can be equivalently converted into a large kernel convolution.

[0014] Furthermore, the large-core convolution adopts a convolution kernel size of 13x13, and a residual connection method is used to maintain the flow of information, prevent the gradient vanishing problem, capture a wider range of contextual information, and enhance the understanding of global information.

[0015] Furthermore, the hybrid pyramid pooling structure SPPF9 construction method described in step (2) is: firstly, a Conv layer is set, and three maximum pooling layers and three average pooling layers are set after the Conv layer; then, the features of the Conv layer and each pooling layer are fused through the Concat operation; and finally, a Conv layer is set.

[0016] Furthermore, the DySample sampling module described in step (2) includes an offset generating convolution and a weight generating convolution; the offset generating convolution generates an offset for each pixel; a layer generates an interpolation weight through a convolution layer, which is used to dynamically adjust how to weight neighboring pixels during upsampling; the weight generating convolution is similar to the offset prediction branch, and the generation of interpolation weights is also completed through a lightweight convolution operation; then the interpolation weight is set, which determines the weighted summation of pixels near the sampling position when performing bilinear interpolation, ensuring that the feature map after DySample upsampling can accurately retain the details and boundary information of the input feature map.

[0017] Furthermore, the DySample upsampling operation is expressed as:

[0018]

[0019] Among them, g(W(i',j'),i,j,i',j') is an interpolation function controlled by the dynamic weight W(i',j'), which is used to determine the contribution of X(i,j) to the output position (i',j').

[0020] Furthermore, the receptive field attention convolution RFACBAMConv9 described in step (2) introduces intervals in the convolution operation by dilating the convolution, so that the convolution kernel can cover a larger area; uses a fully connected layer and an activation function to calculate the weight of each channel, and achieves channel selection and enhancement by weighting the channel feature map; and at the same time integrates the CBAM attention mechanism to enhance feature extraction and capture of detailed image features.

[0021] Furthermore, the C2f module described in step (2) adopts multi-head attention, each head is responsible for capturing different feature representations; overlapping block embedding is used in the first stage, and block merging modules are used in the second to fourth stages to reduce the input spatial resolution while increasing the number of channels, and then continuous BiFormer (visual conversion) blocks are used for feature transformation; at the beginning of each block, deep convolution is used to implicitly encode relative position information; then the BRA module and the perceptron module with an expansion rate of 2 layers are applied in sequence, which are used for cross-position relationship modeling and each position respectively.

[0022] Furthermore, the C2f module implementation process is as follows:

[0023] Initial convolution:

[0024] F0=Conv(X)

[0025] BiLevel Routing Attention is applied after each convolution operation:

[0026] F i =Conv i (F i-1 )

[0027] From the previous convolution result F i-1 Generate the feature map F of the current layer i ; Extract higher-level features layer by layer through multiple convolutional layers:

[0028] F i =A BLA (F i )=α·A global (F i )+β·A local (F i )

[0029] A BLA (F i ) represents the feature map F i Apply BiLevel Routing Attention; A global (F i ) is a feature calculated by the global attention mechanism, capturing the overall context information of the input feature map; A local (F i ) is a feature calculated by the local attention mechanism, focusing on the detailed information of the local area; α and β are learnable weight parameters used to control the contribution of global and local attention in the final feature;

[0030] Finally, all feature maps to which the attention mechanism is applied are fused:

[0031] Fout =Fusion(F0,F1,…,F n ).

[0032] Furthermore, the detection head in the Head layer of the AQ9 model described in step (2) uses a distributed focus loss function and an attention mechanism to improve the recognition and positioning capabilities of targets in the image through multi-level feature extraction and integration; the calculation formula of the distributed focus loss function is as follows:

[0033]

[0034] Among them: the first term is the classification loss, which uses Focal Loss to measure the difference between the predicted probability distribution and the actual distribution; the second term is the regression loss, which uses L1 loss to measure the difference between the predicted value x^ and the true value x; λ is a balancing factor used to control the weight between classification loss and regression loss.

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

[0036] 1. Reduce computational burden: Optimize the algorithm structure of the attention mechanism and reduce additional computational steps to improve the inference speed and efficiency of the model in resource-constrained environments;

[0037] 2. Simplify parameter adjustment: Design a more adaptive parameter configuration method to reduce the need for manual hyperparameter adjustment, so that the module can automatically optimize performance on different network architectures and data sets;

[0038] 3. Enhance the universality and robustness of the model: Improve the integration method of the attention mechanism to ensure that it can run more stably in various existing CNN architectures, while improving tolerance to noise and data anomalies;

[0039] 4. Expand the application scope of the model: By optimizing the design of the attention module, expand its scalability and effectiveness on large-scale networks and datasets, and ensure that it can continue to provide performance advantages in complex tasks;

[0040] In summary, the present invention provides a more efficient, stable and accurate method for automatic driving target detection in low-light environments at night. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of the autonomous driving target detection model AQ9 proposed in the present invention;

[0042] Figure 2 It is a schematic diagram of the overall framework structure diagram of LKNet9 proposed by the present invention;

[0043] Figure 3It is a schematic diagram of the structure of SPPF9 proposed in the present invention;

[0044] Figure 4 This is a schematic diagram of the RFACBAMConv9 module structure proposed in the present invention;

[0045] Figure 5 This is a schematic diagram of the C2f module structure proposed by the present invention;

[0046] Figure 6 Schematic diagram of the large-kernel convolution structure of the LKNet9 module proposed in this invention. DETAILED DESCRIPTION

[0047] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0048] As attached Figure 1 As shown, the present invention proposes a method for automatic driving target detection in a nighttime low-light environment, comprising the following steps:

[0049] Step 1: Perform data augmentation on the KITTI dataset and use it as input to train the autonomous driving target detection model.

[0050] Step 2: Build the autonomous driving target detection model AQ9, mainly considering the need for high-precision target recognition in low-light environments at night.

[0051] (1) Figure 2 As shown, the present invention uses LKNet9 as the backbone network. This network structure includes efficient modules such as the Rep module and the neural network architecture combined with the LCB module to increase the network depth, and successfully decouples the relationship between the expansion of the receptive field, the improvement of the abstract level and the increase of parameters in the large convolution kernel architecture.

[0052] The Rep (unified representation) module aims to reduce network parameters while maintaining high performance through feature reuse and representation sharing. This module achieves efficient feature expression through a multi-branch structure. The LCB (local convolution) module uses local convolution operations to improve the network's receptive field and efficiency through point-by-point convolution and local area processing.

[0053] Use large kernel convolution to expand the receptive field to capture sparse features. Using the idea of ​​structural reparameterization, the entire block can be equivalently converted to a large kernel convolution. This is because a small kernel + dilated convolution is equivalent to a large kernel + non-dilated convolution, such as Figure 6As shown. Regarding kernel size: the kernel size is selected according to the downstream task and the specific framework used; for the task involved in this invention, 13x13 is used. At the same time, the residual connection method is used to maintain the flow of information and prevent the gradient disappearance problem. Especially in deep networks, the network can capture a wider range of contextual information and enhance the understanding of global information.

[0054] The LKNet9 network can efficiently handle a variety of data recognition tasks including audio and video through structures such as dilated convolution and ReparamBlock (super-large core). Through the design of different modules and model sizes, it demonstrates its excellent performance in various tasks.

[0055] When adding a 3×3 convolutional layer (conv) to a ConvNet (small convolution kernel), three significant effects occur: first, the receptive field becomes larger; second, the level of abstraction of spatial patterns is improved (such as recognition from edges and textures to object shapes); and finally, the generalization ability of the model is enhanced by introducing more learnable parameters and nonlinearities. In contrast, in a large convolution kernel architecture, these three effects should be decoupled. This is because the main advantage of large convolution kernels is that a wide receptive field can be achieved even without increasing the depth of the network. Since increasing the convolution kernel size is more efficient in expanding the receptive field than stacking more layers, a sufficient effective receptive field (ERF) can be constructed with a small number of large convolution kernel layers, thereby saving computational resources for other efficient structures that can increase the level of abstraction of spatial patterns or the depth of the network.

[0056] ReparamBlock (extra-large kernel): Ignoring input pixels in a dilated convolution is equivalent to inserting extra zeros into the conv kernel, so a dilated conv layer with a small kernel can be equivalently converted to a non-dilated (i.e., r=1) layer with a sparse larger kernel.

[0057] Original convolution kernel: E∈R k×k .

[0058] After zero insertion: W'∈R ((k-1)r+1)×((k-1)r+1) .

[0059] This can be achieved by transposed convolution with a step size of r:

[0060] W'=conv_transpose2d(W,I,stride=r)

[0061] Among them, k represents the convolution kernel, r represents the step size, and W represents the model after convolution.

[0062] (2) The pooling layer of the AQ9 detection model is the SPPF9 hybrid pyramid pooling module, which reduces the spatial size of the input feature map by retaining the maximum value in each region;

[0063] like Figure 3 As shown in Figure 1, the construction method of the SPPF9 hybrid pyramid pooling module is as follows: First, set a convolution layer (Conv layer), then add three maximum pooling layers and three average pooling layers. Then, the convolution layer is fused with the features of each pooling layer through the Concat operation. Finally, set another convolution layer to complete the overall architecture of the H-SPPF module. Among them, MaxPool2d k=5 represents the maximum pooling operation, and the size of the pooling window is 5×5.

[0064] The height and width of the output feature map after the MaxPool2d operation are Hout and Wout, respectively, and their calculation formulas are as follows:

[0065]

[0066] in, represents the rounding down operation, k is the size of the pooling window (assumed to be a square, i.e. k×k); s is the stride; and p is the padding.

[0067] The SPPF9 hybrid pyramid pooling module reduces the spatial size of the feature map by retaining the maximum value in each region, thereby reducing the amount of computation and the number of parameters of the model. The SPPF9 hybrid pyramid pooling module has a certain degree of translation invariance, that is, when the target in the input feature map moves to a certain extent, the output of the maximum pooling remains stable, which helps to improve the robustness of the model. It helps the network extract important features in the input feature map, because the maximum value usually represents the most significant feature in the region. It reduces the size of the feature map in the network, thereby reducing memory consumption.

[0068] (3) The sampling method of the AQ9 detection model is DySample, which can adaptively adjust the sampling strategy according to the input feature map, which can usually improve the quality of the upsampled image, thereby enhancing the overall performance of the model. The core idea of ​​DySample is to generate dynamic sampling weights through the network. These weights guide the interpolation method of pixels during the upsampling process. Unlike the fixed interpolation weights in traditional methods, DySample dynamically calculates these weights based on the local features of the input image or feature map, so as to better adapt to different areas of the image. The following is the calculation formula of DySample: After generating the weight matrix W, DySample uses these weights to calculate the upsampled feature map Y. Assuming the target size is H′×W′, the upsampling operation of DySample can be expressed as:

[0069]

[0070] Among them, g(W(i',j'),i,j,i',j') is an interpolation function controlled by the dynamic weight W(i',j'), which is used to determine the contribution of X(i,j) to the output position (i',j').

[0071] (4) The AQ9 detection model uses large kernel convolution, which combines the receptive field enhancement of RFA and the attention mechanism of CBAM in the convolution operation. Figure 4 As shown in the figure, RFACBAMConv9 is improved by integrating RFA convolution and CMAM attention mechanism. First, RFA convolution is used to enhance the convolution operation of the receptive field (such as dilated convolution) to extract features in a larger range. The CBAM module is applied to the feature map after the convolution operation, and the channel attention and spatial attention are adjusted respectively. Finally, the final feature map is extracted through the convolution operation.

[0072] The calculation process of RFACBAMConv9 is:

[0073] RFA Convolution:

[0074]

[0075] Channel Attention:

[0076] M c (F rfa )=σ(MLP(AvgPool(F rfa ))+MLP(MaxPool(F rfa )))

[0077] Spatial Attention:

[0078] M s (F rfa )=σ(f 7×7 ([AvgPool(F rfa );MaxPool(F rfa )]))

[0079] Attention Fusion:

[0080] F cbam =M c (F rfa )×F rfa

[0081] F″ cbam =M s (F cbam )×F cbam

[0082] Standard convolution:

[0083]

[0084] Among them, y[i,j] is the value of the output feature map at position (i,j), x[i+m,j+n] is the value of the input feature map at position (i+m,j+n), w[m,n] is the weight of the convolution kernel, and F rfa Represents RFA convolution.

[0085] (5) The Head layer of the AQ9 detection model uses the C2f module, which is the C2 module fused with dynamic sparse attention (BiLevelRoutingAttention).

[0086] Dynamic sparse attention uses multi-head attention, each head is responsible for capturing different feature representations. But the difference is that each head can dynamically select a sparse attention pattern, which is more efficient in calculation. Secondly, it contains a sparse pattern selector: this is the core part of dynamic sparse attention. This module dynamically selects a part of the attention connections as sparse patterns through a certain strategy (such as based on data content or predefined rules). Specific methods include threshold selection based on query-key scores, soft threshold methods, or learned sparsity strategies. Then normalization is performed to ensure that their sum is 1 and maintain numerical stability. Through selective calculation and dynamic sparse patterns, the computational cost of the attention mechanism is significantly reduced while retaining model performance.

[0087] like Figure 5 As shown, the C2 module is fused with BiLevel Routing Attention to generate a new C2f module, which can usually be achieved by adding an attention mechanism during the feature extraction process. The following is the fusion method:

[0088] Initial convolution:

[0089] F0=Conv(X)

[0090] Multi-layer feature extraction, combined with attention mechanism: BiLevel RoutingAttention is applied after each convolution operation:

[0091] F i =Conv i (F i-1 )

[0092] Here it represents the convolution result F from the previous layer i-1 Generate the feature map F of the current layer i This operation extracts higher-level features layer by layer through multiple convolutional layers:

[0093] F i =ABLA (F i )=α·A global (F i )+β·A local (F i )

[0094] A BLA (F i ) represents the feature map F i Apply BiLevel Routing Attention; A global (F i ) is a feature computed through the global attention mechanism, which captures the overall contextual information of the input feature map; A local (F i ) is a feature calculated by the local attention mechanism, which focuses on the detailed information of the local area; α and β are learnable weight parameters used to control the contribution of global and local attention in the final feature.

[0095] Finally, all feature maps to which the attention mechanism is applied are fused:

[0096] F out =Fusion(F0,F1,…,F n )

[0097] The fusion operation here can be concatenation, weighted summation or other fusion methods of feature maps. It fuses the feature maps F0, F1, ..., Fn generated by all convolutional layers into a comprehensive feature map Fout. This comprehensive feature map contains high-level features processed by multiple layers of convolution and attention, which can provide richer representation for subsequent detection.

[0098] (6) The AQ9 detection model uses the AQ9 detection head to improve the recognition and positioning capabilities of targets in images in low-light environments at night. The detection head uses multi-level feature extraction and integration, distributed focal loss function and an attention mechanism to improve the recognition and positioning capabilities of targets in images.

[0099] The calculation formula of the distribution focal loss function is as follows:

[0100]

[0101] Among them, the first item is the classification loss, and Focal Loss is used to measure the difference between the predicted probability distribution and the actual distribution; the second item is the regression loss, and L1 loss is used to measure the difference between the predicted value x^ and the true value x; λ is a balancing factor used to control the weight between classification loss and regression loss.

[0102] Step 3: Use the enhanced KITTI data to train the detection model AQ9 and adjust the model training parameters until the optimal model weights are obtained; then detect the image to be detected.

[0103] In the case of 300 epoch training, the learning rate lr is set to 0.0001, a 4070 graphics card is used, and the batch-size is set to 16. The results show that under the KITTI dataset, the ablation experiment results are compared as shown in Table 1:

[0104] Table 1 Comparison of ablation experiment results

[0105]

[0106]

[0107] The meanings of the data in the table are as follows: Class represents the model type. Precision (P): Precision refers to the ratio of the number of entities correctly predicted by the model to the number of all predicted entities. Specifically, precision is equal to the intersection of the number of predicted entities and the number of true entities divided by the number of predicted entities. Recall (R): Recall refers to the ratio of the number of entities correctly predicted by the model to the number of all true entities. Specifically, recall is equal to the intersection of the number of predicted entities and the number of true entities divided by the number of true entities. Mean Average Precision (mAP50): It is a measure of the performance of the target detection model. Mean Average Precision is one of the commonly used performance evaluation indicators in target detection tasks. It takes into account the detection accuracy of different categories and calculates their average. mAP 50 is the mAP score when the IoU threshold is 0.5, which reflects the comprehensive performance of the model on different categories.

[0108] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for automatic driving target detection in a nighttime low-light environment, characterized in that: The following steps are involved: (1) Perform data enhancement on the KITTI dataset; (2) Build the autonomous driving target detection model AQ9; The AQ9 uses the LKNet9 network as the backbone network; the LKNet9 network includes a unified representation module Rep, a local convolution module LCB and a large kernel convolution, which extracts basic features from the input image and performs local features to identify target obstacles at night; the pooling module of the AQ9 model is a hybrid pyramid pooling structure SPPF9, which uses an adaptive pooling algorithm to automatically adjust the scale and type of pooling according to the content of the input feature map; the nearest neighbor interpolation upsampling module in the Head layer of the AQ9 model is DySample sampling, which is adaptively adjusted according to the content of the input feature; the AQ9 model A receptive field attention convolution RFACBAMConv9 is proposed in the Head layer of the AQ9 model, which extracts features in a larger range by enhancing the convolution operation of the receptive field, and applies the attention mechanism to the feature map after the convolution operation, and adjusts the channel attention and spatial attention respectively; the Head layer of the AQ9 model adopts the C2f module, which is derived from the C2 module fused with dynamic sparse attention, focuses on the spatial dimension of the feature map, and strengthens the allocation weight of each spatial position; the detection head in the Head layer of the AQ9 model, for low-light environments at night, enhances local attention and enhances the ability to capture important features in a small range; (3) Training the autonomous driving target detection model AQ9 constructed in step (2) and adjusting the training parameters of the model until the optimal model weight is obtained; (4) Use the trained autonomous driving target detection model AQ9 for detection.

2. The method for detecting an automatic driving target in a nighttime low-light environment according to claim 1, characterized in that: In step (1), the unified representation module Rep reduces network parameters by reusing features and sharing representations while maintaining high performance. The local convolution module LCB uses local convolution operations to improve the network's receptive field and efficiency by processing point-by-point convolution and local areas; the large kernel convolution expands the receptive field to capture sparse features. Using the idea of ​​structural reparameterization, the entire block can be equivalently converted into a large kernel convolution.

3. The method for detecting an automatic driving target in a nighttime low-light environment according to claim 1, characterized in that: The large-core convolution adopts a convolution kernel size of 13x13, and uses a residual connection method to maintain the flow of information, prevent the gradient disappearance problem, capture a wider range of context information, and enhance the understanding of global information.

4. The method for detecting an automatic driving target in a low-light environment at night according to claim 1, characterized in that: The hybrid pyramid pooling structure SPPF9 construction method described in step (2) is: first set a Conv layer, and set three maximum pooling layers and three average pooling layers after the Conv layer; then fuse the features of the Conv layer and each pooling layer through the Concat operation; finally set a Conv layer.

5. The method for detecting an automatic driving target in a low-light environment at night according to claim 1, characterized in that: The DySample sampling module described in step (2) includes an offset generating convolution and a weight generating convolution; the offset generating convolution generates an offset for each pixel; a layer generates an interpolation weight through a convolution layer, which is used to dynamically adjust how to weight neighboring pixels during upsampling; the weight generating convolution is similar to the offset prediction branch, and the generation of interpolation weights is also completed through a lightweight convolution operation; then the interpolation weight is set, which determines the weighted summation of pixels near the sampling position when performing bilinear interpolation, ensuring that the feature map after DySample upsampling can accurately retain the details and boundary information of the input feature map.

6. The method for detecting an automatic driving target in a low-light environment at night according to claim 5, characterized in that: The DySample upsampling operation is expressed as: Among them, g(W(i',j'),i,j,i',j') is an interpolation function controlled by the dynamic weight W(i',j'), which is used to determine the contribution of X(i,j) to the output position (i',j').

7. The method for detecting an automatic driving target in a nighttime low-light environment according to claim 1, characterized in that: The receptive field attention convolution RFACBAMConv9 described in step (2) introduces intervals in the convolution operation by dilating the convolution, so that the convolution kernel can cover a larger area; uses a fully connected layer and an activation function to calculate the weight of each channel, and achieves channel selection and enhancement by weighting the channel feature map; at the same time, it integrates the attention mechanism CBAM to enhance feature extraction and capture of detailed image features.

8. The method for detecting an automatic driving target in a nighttime low-light environment according to claim 1, characterized in that: The C2f module described in step (2) adopts multi-head attention, each head is responsible for capturing different feature representations; overlapping block embedding is used in the first stage, and block merging modules are used in the second to fourth stages to reduce the input spatial resolution while increasing the number of channels, and then continuous BiFormer (visual conversion) blocks are used for feature transformation; at the beginning of each block, deep convolution is used to implicitly encode relative position information; then the BRA module and the perceptron module with an expansion rate of 2 layers are applied in sequence, which are used for cross-position relationship modeling and each position respectively.

9. The method for detecting an automatic driving target in a low-light environment at night according to claim 1, characterized in that: The C2f module implementation process is as follows: Initial convolution: F0=Conv(X) Apply BiLevel Routing Attention after each convolution operation: F i =Conv i (F i-1 ) From the previous convolution result F i-1 Generate the feature map F of the current layer i ; Extract higher-level features layer by layer through multiple convolutional layers: F i =A BLA (F i )=α·A global (F i )+β·A local (F i ) A BLA (F i ) represents the feature map F i Apply BiLevel Routing Attention; A global (F i ) is a feature calculated by the global attention mechanism, capturing the overall context information of the input feature map; A local (F i ) is a feature calculated by the local attention mechanism, focusing on the detailed information of the local area; α and β are learnable weight parameters used to control the contribution of global and local attention in the final feature; Finally, all feature maps to which the attention mechanism is applied are fused: F out =Fusion(F0,F1,…,F n )。 10. The method for detecting an automatic driving target in a nighttime low-light environment according to claim 1, characterized in that: The detection head in the Head layer of the AQ9 model described in step (2) uses a distributed focus loss function and an attention mechanism to improve the recognition and positioning capabilities of targets in the image through multi-level feature extraction and integration; the calculation formula of the distributed focus loss function is as follows: Among them: the first term is the classification loss, which uses Focal Loss to measure the difference between the predicted probability distribution and the actual distribution; the second term is the regression loss, which uses L1 loss to measure the difference between the predicted value x^ and the true value x; λ is a balancing factor used to control the weight between classification loss and regression loss.

Images