Traffic scene target detection method based on feature fusion and attention mechanism

Abstract

The invention relates to the technical field of traffic scene target detection, in particular to a traffic scene target detection method based on feature fusion and an attention mechanism, and the method comprises the steps: obtaining a public traffic scene target image training data set; training a traffic target detection model by using the traffic scene target image training data set; and inputting a to-be-detected traffic scene image into the trained traffic scene target detection model to obtain an output detection result graph. A hierarchical receptive field is constructed through a feature fusion module, a fine structure of a target edge is reserved in a deep convolution stage, cross-channel semantic information is dynamically aggregated through point convolution, an attention module is put forward to dynamically adjust a convolutional receptive field and a space attention weight, interference of background noise is suppressed, and a target is obtained. The local detail feature expression of the target is better enhanced, a PIoU loss function is adopted, and the positioning precision and robustness are further improved by combining a target size adaptive penalty factor and an anchor frame quality-based gradient adjustment strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic scene target detection, and in particular to a traffic scene target detection method based on feature fusion and attention mechanism. Background Art

[0002] Object detection is increasingly used in many fields, especially in video surveillance and autonomous driving. Deep learning-based object detection technology provides strong support for these fields. In autonomous driving, achieving a higher level of performance requires the system to accurately perceive the driving environment to ensure that the vehicle can safely identify and respond to surrounding obstacles.

[0003] However, autonomous driving faces many challenges, especially when the vehicle is driving at high speed, object detection faces many difficulties, including object occlusion, lighting changes, weather conditions (such as rain, fog, snow, etc.), and other complex scenes (such as occlusion that may occur when multiple lanes are driving in parallel). The influence of these factors may cause the image quality acquired by the sensor to deteriorate, which in turn affects the accuracy and robustness of the target detection model. In severe cases, it may affect the decision-making ability of the system, and thus affect the safety and reliability of the autonomous driving vehicle. Therefore, improving the accuracy and robustness of target detection technology in complex environments is still a major challenge in current research.

[0004] At present, target detection in traffic scenes often encounters recognition difficulties when facing problems such as object occlusion, shape change, and perspective change. These changes may cause the appearance of the target to change significantly, thereby increasing the difficulty of detection. To address these problems, a common solution is to enhance the model's understanding and attention to the target by improving feature fusion technology and introducing attention mechanisms. However, simple feature fusion methods are not enough to capture the detailed information and overall context of the target, especially when the background is complex or the target is partially occluded. How to further improve the performance of the model in these situations remains a difficult problem to be solved in the field of traffic scene target detection. Summary of the invention

[0005] The purpose of the present invention is to provide a traffic scene target detection method based on feature fusion and attention mechanism, which solves the problem that target detection in existing traffic scenes often encounters recognition difficulties when facing problems such as object occlusion, shape change, and perspective change, and common solutions are not sufficient to capture the detailed information and overall context of the target, especially in complex backgrounds or when the target is partially occluded.

[0006] To achieve the above object, the present invention provides a traffic scene target detection method based on feature fusion and attention mechanism, comprising the following steps:

[0007] Prepare the datasets for training, including the PASCAL VOC 2007 and 2012 datasets, and the COCO 2017 dataset;

[0008] Train the original network model on the two datasets respectively, and record the metrics for evaluating performance as the criteria for subsequent performance evaluation;

[0009] Propose an improved method;

[0010] Train the network model of the improved method on the two datasets respectively, then evaluate the network model and compare it with the metric criteria of the original method;

[0011] If the improved method meets the performance requirements, conduct detection and analysis on the improved method. If the improved method does not meet the performance requirements, continue to tune the parameters or improve the network model until the performance requirements are met;

[0012] Conduct detection and analysis on the improved method that meets the performance requirements, and verify and analyze the improved method on the test set.

[0013] Among them, the proposed improved method includes the following steps:

[0014] The proposed improved method includes the following steps:

[0015] Reconstruct the C2PSA structure to enhance the network's perception ability of the target. By introducing large separable kernel convolution (LSKC) to construct a hierarchical large receptive field, it strengthens the ability of multi-scale feature fusion to capture target edge and texture information, reduces detail loss, and improves detection accuracy;

[0016] Use the scalable residual attention mechanism to improve the C3k2 module. Expand the receptive field through dilated convolution, capture multi-scale global context information, dynamically adjust the convolutional receptive field and spatial attention weights, suppress the interference of background noise, and enhance the expression of local detail features of the target;

[0017] Adopt the PIoU loss function. By combining the target size adaptive penalty factor and the gradient adjustment strategy based on the anchor box quality, it further improves the localization accuracy and robustness;

[0018] Add each module to the baseline network model for training respectively, and record the results;

[0019] Combine all modules together to form a new single-stage object detector;

[0020] Train the network model.

[0021] Among them, the reconstruction of the C2PSA structure includes:

[0022] The C2PSA module is reconstructed using large separable kernel convolution (LSKC), fundamentally solving the feature expression bottleneck of the original module in object detection tasks.

[0023] Meanwhile, the reconstructed module constructs a hierarchical receptive field through separable large convolution kernels, preserves the fine structure of the object edges in the deep convolution stage, and then dynamically aggregates cross-channel semantic information through point convolution. Its gated attention mechanism can suppress the activation intensity of the background interference area while strengthening the channel weight allocation of the object area.

[0024] In addition, it not only reduces the number of module parameters but also establishes a semantic association between the object and the long-range context in complex traffic scenes through spatial-channel two-dimensional attention guidance, improving the accuracy of the model.

[0025] Among them, the improved C3k2 module includes:

[0026] The C3k2 module is improved using an extensible residual attention mechanism to capture multi-scale global context information, enabling the model to better understand the spatial relationships in complex scenes.

[0027] Meanwhile, the attention mechanism is used to dynamically adjust the feature weights, highlighting the key features of the object and suppressing background noise to enhance the separation ability between the object and the background.

[0028] In addition, the feature propagation is further optimized through residual connections to alleviate the vanishing gradient problem in deep neural networks.

[0029] Among them, the PIoU loss function includes:

[0030]

[0031] Among them, A and B are the areas of the predicted box and the ground truth box respectively; ∩ represents the intersection; ∪ represents the union; the default value of λ is 1.3, which adjusts the decay rate of the dynamic modulation factor and controls the intensity of the boundary alignment optimization; P represents the boundary alignment penalty term, which calculates the offsets of the four boundaries of the predicted box and the target box, dw1 represents the left boundary offset; dw2 represents the right boundary offset; d h1 represents the lower boundary offset; d h2 represents the upper boundary offset; w2 and h2 represent the width and height of the target box, which are used to normalize the offsets.

[0032] The PIoU loss function explicitly calculates the offsets of the four boundaries, uses the width and height of the target box as the normalization factors, so that a slight offset of the object significantly increases the loss value after normalization, thereby forcing the model to accurately align the boundaries.

[0033] In addition, through the combined effect of exponential decay and sub-exponential decay, the dynamic modulation factor of the PIoU loss function generates the maximum gradient at medium errors, accelerating the convergence of the model; while at extreme errors, it suppresses gradient explosion and improves training stability. This design enables the PIoU loss function to significantly improve the localization accuracy of bounding boxes in object detection scenarios.

[0034] Among them, the improved new single-stage object detector includes three parts: the feature fusion of the improved new single-stage object detector, enhanced attention, and the loss function.

[0035] In the feature fusion part, the C2PSA structure is reconstructed to build a hierarchical receptive field, and the fine structure of the object edge is retained in the depth convolution stage. Then, point convolution is used to dynamically aggregate cross-channel semantic information to suppress background interference information. In the enhanced attention part, the C3k2 module is improved using an extensible residual attention mechanism to capture multi-scale global context information, enabling the model to better understand the spatial relationships in complex scenes. In the loss function part, the PIoU loss function is adopted, and by combining an object size adaptive penalty factor and a gradient adjustment strategy based on anchor box quality, the localization accuracy and robustness are further improved.

[0036] Among them, the training of the network model includes the following steps:

[0037] Select the PyTorch framework to implement the network model;

[0038] Configure the network according to the parameters of the original model, set the training image resolution to 640×640, the number of training epochs to 100, and keep the other parameters consistent with the original model;

[0039] Conduct experiments on a server equipped with Linux = 3.10.0-1127.el7.x8664 and GPU = A100;

[0040] Adjust the parameters of the network model according to the experimental results to maximize the detection accuracy and speed;

[0041] Repeat the above steps and continuously adjust the parameters of the network model to achieve the best detection effect.

[0042] A traffic scene object detection method based on feature fusion and attention mechanism of the present invention uses the convolutional neural network CSPDarkNet53 as the backbone network of the model. It reconstructs the C2PSA structure through large separable kernel convolution (LSKC) to construct a hierarchical receptive field, preserves the fine structure of the object edge in the deep convolution stage, then dynamically aggregates cross-channel semantic information through point convolution to suppress background interference information, uses an extensible residual attention mechanism to improve the C3k2 module, captures multi-scale global context information, extracts multi-scale feature information with strong robustness, including deep and shallow feature map information, and adopts the PIoU loss function. By combining the object size adaptive penalty factor and the gradient adjustment strategy based on the anchor box quality, the positioning accuracy and robustness are further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art.

[0044] Figure 1 It is a schematic structural diagram of a traffic scene object detection model of the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0045] Figure 2 It is a schematic diagram of the reconstructed C2PSA structure of the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0046] Figure 3 It is a schematic diagram of the extensible residual attention mechanism of the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0047] Figure 4 It is a schematic diagram of the Bottleneck_2 module of the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0048] Figure 5 It is a schematic diagram of the improved C3k2 module of the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0049] Figure 6 It is a flow step diagram of the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0050] Figure 7 It is a specific step diagram of the improved method proposed for the traffic scene object detection method based on feature fusion and attention mechanism in the first embodiment of the present invention.

[0051] Figure 8 This is the specific step diagram for training the network model of the traffic scene target detection method based on feature fusion and attention mechanism in the first embodiment of the present invention. Detailed implementation manners

[0052] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present invention, but should not be construed as limiting the present invention.

[0053] First embodiment:

[0054] Please refer to Figures 1 to 8 , the present invention provides a traffic scene target detection method based on feature fusion and attention mechanism, including the following steps:

[0055] S1. Prepare the data sets for training, including PASCAL VOC 2007 and 2012 data sets, and COCO 2017 data set.

[0056] Specifically, the PASCAL VOC 2007 and 2012 data sets contain 16,551 training images and 4,952 test images, and the COCO 2017 data set contains 118,287 training images, 5,000 validation images and 40,670 test images. Using the PASCAL VOC 2007 and 2012 data sets and the COCO 2017 data set, the data sets used cover common traffic scenes, and each image has detailed annotations, annotating the specific positions and sizes of traffic targets in the image. When implementing the present invention, the image resolution input into the model in the network training experiment is uniformly adjusted to 640×640.

[0057] S2. Train the original network model on the two data sets respectively, and record the metrics for evaluating performance as the criteria for subsequent performance evaluation.

[0058] Specifically, the metrics for evaluating performance include mAP@.5, mAP@.5:.95, Params(M) and FLOPs(G).

[0059] S3. Propose an improvement method.

[0060] Specifically, the proposed improvement method includes the following steps:

[0061] S31. Reconstruct the C2PSA structure to enhance the network's perception ability of targets, strengthen the capture ability of multi-scale feature fusion for target edge and texture information, reduce detail loss, and improve detection accuracy.

[0062] Specifically, please refer to Figure 2, in the embodiments of the present invention, a large separable kernel convolution (LSKC) is used to reconstruct the C2PSA module, fundamentally solving the feature expression bottleneck of the original module in the object detection task. The module reconstructed by the large separable kernel convolution constructs a hierarchical receptive field through separable large kernel convolutions, retains the fine structure of the object edges in the deep convolution stage, and then dynamically aggregates cross-channel semantic information through point convolutions. Its gated attention mechanism can suppress the activation intensity of the background interference area while strengthening the channel weight allocation of the object area. This design not only reduces the number of module parameters, but also establishes a semantic association between the object and the long-range context in complex traffic scenarios through spatial-channel two-dimensional attention guidance, improving the accuracy of the model.

[0063] S32. Improve the C3k2 module using an extensible residual attention mechanism, expand the receptive field through dilated convolutions, capture multi-scale global context information, and obtain an enhanced feature map.

[0064] Specifically, please refer to Figure 3 , in the embodiments of the present invention, the extensible residual attention mechanism expands the receptive field through dilated convolutions, captures multi-scale global context information, enabling the model to better understand the spatial relationships in complex scenarios. At the same time, the attention mechanism dynamically adjusts the feature weights, highlighting the key features of the object and suppressing background noise, thereby enhancing the separation ability between the object and the background. In addition, the residual connection further optimizes the feature propagation, alleviates the vanishing gradient problem in deep networks, and ensures that the features of the object can be effectively transmitted to subsequent network layers.

[0065] Please refer to Figure 4 , in the embodiments of the present invention, the extensible residual attention mechanism is embedded into the Bottleneck sub-module of C3k2, and the embedded module is renamed Bottleneck_2. The input feature map first passes through the extensible residual attention to capture multi-scale global context information, understand the spatial relationships in complex scenarios, dynamically adjust the feature weights, highlight the key features of the object, and suppress background noise. At the same time, after two 3×3 convolutions, normalization, and SiLU activation function layers, it is added element-wise to the original feature map to obtain an enhanced feature map.

[0066] Please refer to Figure 5, in the embodiments of the present invention, when the input parameter c3k is True, the C3k2 module uses C3k as an internal component, and when the input parameter c3k is False, it uses Bottleneck as an internal component. The present invention adds an expandable residual attention mechanism to Bottleneck, and when the input parameter c3k is False, Bottleneck_2 is embedded into C3k2 to replace the original Bottleneck to complete the improvement. Its feature map first passes through a 1×1 convolutional layer, then passes through the Bottleneck_2 module embedded with the expandable residual attention mechanism, and is added to the original feature map in the channel dimension through a residual connection. Finally, the enhanced feature map is output after passing through a 1×1 convolutional layer.

[0067] S33. The PIoU loss function is adopted, and by combining the target size adaptive penalty factor and the gradient adjustment strategy based on the anchor box quality, the positioning accuracy and robustness are further improved.

[0068] Specifically, the PIoU loss function includes:

[0069]

[0070] where A and B are the areas of the predicted box and the ground truth box respectively; ∩ represents the intersection; ∪ represents the union; the default value of λ is 1.3, which adjusts the decay rate of the dynamic modulation factor and controls the intensity of the boundary alignment optimization; P represents the boundary alignment penalty term, which calculates the offsets of the four boundaries between the predicted box and the target box, dw1 represents the left boundary offset; dw2 represents the right boundary offset; d h1 represents the lower boundary offset; d h2 represents the upper boundary offset; w2 and h2 represent the width and height of the target box, which are used to normalize the offsets.

[0071] The PIoU loss function explicitly calculates the offsets of the four boundaries, and uses the width and height of the target box as the normalization factors, so that a slight offset of the target significantly increases the loss value after normalization, thereby forcing the model to accurately align the boundaries. In addition, the dynamic modulation factor of the PIoU loss function, through the combined effect of exponential decay and sub-exponential decay, generates the maximum gradient at medium errors, accelerating the convergence of the model; while suppressing the gradient explosion at extreme errors, improving the training stability. This design enables the PIoU loss function to significantly improve the positioning accuracy of the box in the object detection scenario.

[0072] S34. Each module is added to the baseline network model for training respectively, and the results are recorded;

[0073] S35. All the modules are combined together to form a new single-stage object detector.

[0074] Specifically, the improved new single-stage object detector includes three parts: the feature fusion of the improved new single-stage object detector, enhanced attention, and loss function.

[0075] In the feature fusion part, the C2PSA structure is reconstructed to build a hierarchical receptive field, and the fine structure of the object edge is retained in the depth convolution stage. Then, point convolution is used to dynamically aggregate cross-channel semantic information to suppress background interference information. In the enhanced attention part, the C3k2 module is improved using an extensible residual attention mechanism to capture multi-scale global context information, enabling the model to better understand the spatial relationships in complex scenes. In the loss function part, the PIoU loss function is adopted, and by combining the object size adaptive penalty factor and the gradient adjustment strategy based on the anchor box quality, the localization accuracy and robustness are further improved.

[0076] S36. Train the network model.

[0077] Specifically, the training of the network model includes the following steps:

[0078] S361. Implement the network model using the PyTorch framework.

[0079] S362. Configure the network according to the parameters of the original model, set the training image resolution to 640×640, and the number of training epochs to 100. Keep the other parameters the same as those of the original model.

[0080] S363. Conduct experiments on a server equipped with Linux = 3.10.0 - 1127.el7.x8664 and GPU = A100.

[0081] S364. Adjust the parameters of the network model according to the experimental results to maximize the detection accuracy and speed.

[0082] S365. Repeat steps S362, S363, and S364 above, continuously adjust the parameters of the network model to achieve the best detection effect.

[0083] Specifically, this model is implemented using the PyTorch framework, with PyTorch version 1.10.0 and CUDA version 11.1. The experimental equipment and instruments are Linux = 3.10.0 - 1127.el7.x8664 and GPU = A100. The number of training epochs is 100, and the input image resolution is uniformly set to 640×640.

[0084] S4. Train the network model of the improved method on the two datasets respectively, then evaluate the network model and compare it with the original method's metric standards.

[0085] S5. If the improved method meets the performance requirements, conduct detection and analysis on the improved method. If the improved method fails to meet the performance requirements, continue to adjust the parameters or improve the network model until the performance requirements are met;

[0086] S6. Conduct detection and analysis on the improved method that meets the performance requirements, and verify and analyze the improved method on the test set.

[0087] Specifically, the improved new single-stage object detector of the present invention consists of three parts: feature fusion, enhanced attention, and loss function. In the feature fusion part, reconstruct the C2PSA structure, construct a hierarchical receptive field, retain the fine structure of the object edge in the depth convolution stage, and then dynamically aggregate cross-channel semantic information through point convolution to suppress background interference information; in the enhanced attention part, use an extensible residual attention mechanism to improve the C3k2 module, capture multi-scale global context information, enabling the model to better understand the spatial relationship in complex scenes; in the loss function part, adopt the PIoU loss function, and further improve the localization accuracy and robustness by combining the target size adaptive penalty factor and the gradient adjustment strategy based on the anchor box quality.

[0088] The above-disclosed are only one or more preferred embodiments of the present application, and the scope of rights of the present application cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.

Claims

1. A traffic scene object detection method based on feature fusion and attention mechanism, characterized in that, It includes the following steps: Prepare the datasets for training, including the PASCAL VOC 2007 and 2012 datasets, and the COCO 2017 dataset; Train the original network model on the two datasets respectively, and record the metrics for evaluating performance as the criteria for subsequent performance evaluation; Propose an improvement method; Train the network model of the improvement method on the two datasets respectively, then evaluate the network model, and compare it with the metric criteria of the original method; If the improvement method meets the performance requirements, conduct detection and analysis on the improvement method. If the improvement method does not meet the performance requirements, continue to adjust the parameters or improve the network model until the performance requirements are met; Conduct detection and analysis on the improvement method that meets the performance requirements, and verify and analyze the improvement method on the test set.

2. The traffic scene object detection method based on feature fusion and attention mechanism according to claim 1, wherein: The proposed improvement method includes the following steps: Reconstruct the C2PSA structure to enhance the network's perception ability of targets. By introducing large separable kernel convolution (LSKC) to construct a hierarchical large receptive field, it strengthens the ability of multi-scale feature fusion to capture target edge and texture information, reduces detail loss, and improves detection accuracy; Use the scalable residual attention mechanism to improve the C3k2 module. Expand the receptive field through dilated convolution, capture multi-scale global context information, dynamically adjust the convolutional receptive field and spatial attention weights, suppress the interference of background noise, and enhance the local detail feature expression of targets; Adopt the PIoU loss function. By combining the target size adaptive penalty factor and the gradient adjustment strategy based on anchor box quality, it further improves the localization accuracy and robustness; Add each module to the baseline network model for training respectively, and record the results; Combine all the modules together to form a new single-stage object detector; Train the network model.

3. The traffic scene object detection method based on feature fusion and attention mechanism according to claim 2, wherein: The reconstruction of the C2PSA structure includes: Use large separable kernel convolution (LSKC) to reconstruct the C2PSA module, fundamentally solving the feature expression bottleneck of the original module in object detection tasks. At the same time, the reconstructed module constructs a hierarchical receptive field through large separable kernel convolution, retains the fine structure of the target edge in the deep convolution stage, and then dynamically aggregates cross-channel semantic information through point convolution. Its gated attention mechanism can suppress the activation intensity of the background interference area while strengthening the channel weight allocation of the target area. In addition, it not only reduces the module parameter quantity, but also establishes the semantic association between the target and the long-range context in complex traffic scenes through spatial-channel two-dimensional attention guidance, improving the accuracy of the model.

4. The traffic scene object detection method based on feature fusion and attention mechanism according to claim 2, wherein: The improvement of the C3k2 module includes: Improve the C3k2 module using an expandable residual attention mechanism to capture multi-scale global context information, enabling the model to better understand spatial relationships in complex scenarios. Meanwhile, dynamically adjust the feature weights through the attention mechanism to highlight the key features of the target and suppress background noise, thereby enhancing the ability to separate the target from the background. In addition, further optimize feature propagation through residual connections to alleviate the vanishing gradient problem in deep neural networks.

5. The traffic scene object detection method based on feature fusion and attention mechanism according to claim 2, wherein: The PIoU loss function includes: Among them, A and B are the areas of the predicted box and the ground truth box respectively; ∩ represents the intersection; ∪ represents the union; the default value of λ is 1.3, which represents the decay rate of adjusting the dynamic modulation factor and controls the intensity of boundary alignment optimization; P represents the boundary alignment penalty term, which calculates the offsets of the four boundaries of the predicted box and the target box; dw1 represents the left boundary offset; dw2 represents the right boundary offset; d h1 represents the lower boundary offset; d h2 represents the upper boundary offset; w2 and h2 represent the width and height of the target box, which are used to normalize the offsets. The PIoU loss function explicitly calculates the offsets of the four boundaries, using the width and height of the target box as the normalization factor, such that a slight offset of the target significantly increases the loss value after normalization, thereby forcing the model to accurately align the boundaries. Furthermore, the dynamic modulation factor of the PIoU loss function, through the combined effect of exponential decay and sub-exponential decay, generates the maximum gradient at medium errors to accelerate model convergence; while suppressing gradient explosion at extreme errors to enhance training stability. This design enables the PIoU loss function to significantly improve the localization accuracy of the bounding box in the object detection scenario.

6. The traffic scene object detection method based on feature fusion and attention mechanism according to claim 2, wherein: The improved new single-stage object detector includes: The improved new single-stage object detector consists of three parts: feature fusion, enhanced attention, and loss function. In the feature fusion part, reconstruct the C2PSA structure to build a hierarchical receptive field, retain the fine structure of the target edge in the deep convolution stage, and then dynamically aggregate cross-channel semantic information through point convolution to suppress background interference information; in the enhanced attention part, use the expandable residual attention mechanism to improve the C3k2 module to capture multi-scale global context information, enabling the model to better understand spatial relationships in complex scenarios; in the loss function part, adopt the PIoU loss function, which further improves the localization accuracy and robustness by combining the target size adaptive penalty factor and the gradient adjustment strategy based on the anchor box quality.

7. The traffic scene object detection method based on feature fusion and attention mechanism according to claim 2, wherein: Training the network model includes the following steps: Select the PyTorch framework to implement the network model; Configure the network according to the parameters of the original model, set the training image resolution to 640×640, the number of training epochs to 100, and keep the other parameters the same as those of the original model; Conduct experiments on a server equipped with Linux = 3.10.0 - 1127.el7.x8664 and GPU = A100; Adjust the parameters of the network model according to the experimental results to maximize the detection accuracy and speed; Repeat the above steps and continuously adjust the parameters of the network model to achieve the best detection effect.

Images