Deep Learning-Based Lane Detection Method in Low-Light Environments

Abstract

This invention discloses a deep learning-based lane line detection method for low-light environments, comprising: acquiring the lane line dataset TuSimple and preprocessing it, dividing it into training and validation sets; training a lane line detection network using the training set, and selecting the optimal lane line detection network model using the validation set; acquiring the multi-exposure image dataset SCIE and using it to train a low-light enhancement network, selecting the optimal low-light enhancement network model; quantizing and deploying the lane line detection network model and the low-light enhancement network model; using the deployed low-light enhancement network model to enhance the illumination of the input image from an onboard front-facing camera to obtain an enhanced image; using the lane line detection network model to detect lane lines in the enhanced image, and outputting the detection results. This invention can detect lane lines more accurately and quickly in low-light environments, further improving the safety and reliability of autonomous driving.

Description

Technical Field

[0001] This invention relates to the technical fields of computer vision, lane detection, and autonomous driving environment perception, and in particular to a deep learning-based lane detection method for low-light environments. Background Technology

[0002] Environmental perception is a crucial component of autonomous driving perception technology. In the three main stages of autonomous driving—perception, decision-making, and control—perception plays a vital role in providing accurate and effective road information to the decision-making module, directly impacting decision-making and control performance. Lane detection is an important part of autonomous driving environmental perception.

[0003] There are three main types of sensors commonly used in autonomous driving: cameras, lidar, and millimeter-wave radar. Due to the high cost of radar and the fact that radar point clouds are more sensitive to material and height features, the visual features of lane lines are more significant. They are usually white or yellow straight lines or curves. Therefore, the research focus of lane line detection is mainly on camera-based visual algorithms.

[0004] With the continuous development of computer vision, researchers have proposed many methods for camera-based lane detection, which can be broadly categorized into several types: lane detection methods based on geometry and machine learning, lane detection methods that treat lane detection as an object detection task, lane detection methods that treat lane detection as a semantic segmentation task, and lane detection methods that treat lane detection as a keypoint detection and regression task. While geometry and machine learning-based lane detection methods are fast, their accuracy is low; methods that treat lane detection as a semantic segmentation or object detection task have high accuracy but are too slow. For the real-time and stability requirements of autonomous driving scenarios, a more suitable method is the lane detection method that treats lane detection as a keypoint detection and regression task. Furthermore, because cameras are sensitive to lighting conditions, low visibility affects the accuracy of lane detection; therefore, regardless of the method, the accuracy is low under dark conditions.

[0005] In summary, while various lane detection methods exist, their performance is not fully adequate to meet the real-time stability requirements of autonomous driving and the demands of varying lighting conditions. Detection accuracy in dark conditions also needs improvement. Therefore, it is urgent to improve existing lane detection networks and propose a deep learning-based lane detection method for low-light environments to address these issues. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of the imbalance between real-time performance and accuracy of lane line detection in autonomous driving scenarios, as well as the problem of decreased lane line detection accuracy caused by changes in lighting conditions. It provides a lane line detection method based on deep learning in low-light environments. By using deep learning methods, the speed and accuracy of lane line detection are improved, and lane lines can be detected more accurately and quickly in low-light environments, thereby further improving the safety and reliability of autonomous driving.

[0007] To achieve the above objectives, the technical solution provided by this invention is: a lane line detection method based on deep learning in low-light environments, comprising the following steps:

[0008] 1) Obtain the lane line dataset TuSimple;

[0009] 2) Preprocess TuSimple and divide it into training and validation sets proportionally;

[0010] 3) The training set is fed into the constructed deep learning-based lane detection network for training, and the training effect is verified using the validation set. The optimal lane detection network model weight parameters are determined based on the performance of the lane detection network on the validation set, and the optimal lane detection network model is constructed using these parameters. The deep learning-based lane detection network consists of a ResNet50 backbone network, a CoT attention module, and two MLP (Multilayer Perceptron) modules. The ResNet50 backbone network is used to extract features from the input lane image, and the CoT attention module is used to filter key features from the features extracted by the ResNet50 backbone network and input them into the two MLP modules. One MLP module is used for lane classification, and the other MLP module is used for lane localization.

[0011] 4) Obtain the multi-exposure image dataset SICE to train the constructed low-light enhancement network and obtain the optimal low-light enhancement network model. The low-light enhancement network has 6 layers. The first 3 layers each consist of a 32-channel 3*3 convolutional layer, and the last 3 layers each consist of two 32-channel 3*3 convolutional layers fused together. A loss function is designed to optimize it.

[0012] 5) Quantize and deploy the optimal lane detection network model obtained in step 3) and the optimal low-light enhancement network model obtained in step 4).

[0013] 6) Input the road images captured by the vehicle's front camera frame by frame into the deployed low-light enhancement network model to obtain enhanced images. Then, input the enhanced images into the deployed lane detection network model to obtain lane detection results. Finally, plot the results on the input image to visualize the detection results.

[0014] Furthermore, in step 1), the obtained lane line dataset TuSimple includes lane line images and lane line image annotation files.

[0015] Furthermore, in step 2), the preprocessing includes dividing the TuSimple image into several rows, recording the coordinate range of the region containing lane lines, and writing it into the TuSimple annotation file; the division ratio of the training set and the validation set is adjusted as needed, but the size of the training set should be greater than the size of the validation set.

[0016] Further, in step 3), when an image is input, the lane detection network first extracts deep features from the segmented rows of images obtained in step 2) using the ResNet50 backbone network. Then, it filters key features using the CoT attention module, converts them into high-dimensional feature vectors, flattens them into one-dimensional row vectors, and inputs them into two MLP modules for processing. One MLP module is responsible for lane line localization, finding the coordinates of each pixel on the lane line. The output of this MLP module is sent to a softmax module, which outputs the probability that each coordinate in each input row corresponds to a lane line. The formula for predicting the lane line for that row is:

[0017]

[0018] In the formula, Let Pb be the predicted location value for the i-th row, j be the traversal of all x-coordinates in the input row, J be the set of x-coordinates of all pixels in the input row, and Pb be the predicted location value for the i-th row. i,j It is the probability value at the i-th row and j-th horizontal coordinate, that is, the output value of the softmax module at that position;

[0019] Define a lane line structure loss function to optimize this MLP module:

[0020] L str =λL shape +L sim (2)

[0021] In the formula, L str Let L be the lane line structure loss function. shape Let L be the lane shape loss function. sim Let λ be the lane line pixel similarity loss function, and λ be the lane line pixel similarity loss weight coefficient.

[0022] Define the lane line shape loss function L shape as follows:

[0023]

[0024] In the formula, h is the number of rows after image segmentation, and k is the number of iterations over h. This represents the predicted location value for the k-th row. This represents the predicted location value for the (k+1)th row. ω represents the predicted location value for the (k+2)th row. shape For random disturbances that affect lane line shape loss;

[0025] Define the lane line pixel similarity loss function L sim as follows:

[0026]

[0027] In the formula, a is the number of rows after image segmentation, b is the number of iterations over a, and P b For the output value of the MLP module corresponding to line b, P b+1 For the output value of the MLP module corresponding to line b+1, ω sim For random perturbations to address the loss of similarity between lane line pixels;

[0028] Another MLP module performs classification, determining whether a segmented row from step 2) contains lane line pixels, thus aiding in the localization result. The classification result is then fed into a softmax module, which outputs the predicted classification value. The lane line classification loss function is defined as follows:

[0029]

[0030] In the formula, L cls Let N be the total number of pixels input to the MLP module, n be the number of iterations over N, and Pc represent the output of the softmax module. n The output of the softmax module for the nth pixel;

[0031] The lane detection network combines the outputs of two MLP modules for analysis. When the MLP module used for classification determines that a certain image row does not contain lane lines, if there are corresponding localization prediction values, it considers these prediction values ​​to lack credibility and therefore discards all localization prediction results for that row. Conversely, if the MLP module used for classification confirms that a certain image row contains lane lines, then the lane line prediction results for that row will be calculated using Equation (1).

[0032] After obtaining the final output lane coordinates, the lane detection result is obtained by least-squares fitting using a third-order spiral. The equation to be fitted is as follows:

[0033] y = m0 + m1x + m2x 2 +m3x 3 (6)

[0034] In the formula, y is the vertical coordinate of a point on the lane line in the image, x is the horizontal coordinate of a point on the lane line in the image, and m0, m1, m2, and m3 are the coefficients of the spiral line to be fitted.

[0035] When validating the performance of the lane detection network using a validation set, accuracy is used as the metric. The formula for calculating accuracy (Acc) is as follows:

[0036]

[0037] In the formula, C region S represents the total number of points correctly predicted as lane lines. region This represents the total number of points on the lane line. The lane line detection network model weight parameter that achieves the maximum Acc value on the validation set is selected as the optimal lane line detection network model weight parameter, and the optimal lane line detection network model is constructed using this weight parameter.

[0038] Furthermore, in step 4), SICE contains images of the same scene under different exposure conditions; the loss function L of the low-light enhancement network... a for:

[0039] L a =W1L s +W2L e +W3L c (8)

[0040] In the formula, W1, W2, and W3 are all weights;

[0041] Spatial consistency loss L s The calculation formula is as follows:

[0042]

[0043] In the formula, K is the number of pixels, c and d are the traversal of pixels, R(c) is the 4-neighborhood of the c-th pixel, Y is the enhanced image, and X is the input image;

[0044] Exposure control loss L e The calculation formula is as follows:

[0045]

[0046] Using 16*16 as a large pixel area, N in the formula r Let F be the number of large pixel regions in the input image, k be the number of large pixel regions traversed, and F be the number of large pixel regions traversed. k It is the average brightness of the k-th large pixel region, and E is a constant located in (0,1);

[0047] Color constant loss L cThe calculation formula is as follows:

[0048]

[0049] In the formula, (p,q) iterates through all pairwise combinations η in the three color channels R, G, and B, and J p J represents the average brightness of color channel p. q This represents the average brightness of color channel q;

[0050] Multiple low-light enhancement network models were obtained by training the low-light enhancement network using SICE. The low-light enhancement network model that minimizes the calculated value of equation (8) was selected as the optimal low-light enhancement network model.

[0051] Further, in step 5), the optimal lane detection network model obtained in step 3) and the optimal low-light enhancement network model obtained in step 4) are quantized using INT8 and converted into ONNX format models. Then, the ONNX format models are converted into TensorRT format models to complete the deployment of the models.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] 1. This invention constructs a lane detection network based on deep learning. The ResNet50 backbone network is used to fully extract the features of the lane line image. The addition of the CoT attention module further improves the network's performance and anti-interference ability. Finally, two MLP modules are used to complete the classification and localization tasks respectively. The lane detection network constructed by this method has high detection accuracy and real-time performance.

[0054] 2. This invention uses a low-light enhancement network to enhance the light intensity of the input image. Images with normal brightness will not be affected, and images with darker brightness will be brightened, thus improving the detection performance of the lane line detection network in low-light environments.

[0055] 3. This invention deploys the lane detection network model and the low-light enhancement network model using TensorRT, which can make fuller use of the GPU's performance and improve the real-time performance of lane detection. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention.

[0057] Figure 2 The structure diagram of the network for enhancing low-light images.

[0058] Figure 3 This is a diagram showing the detection results in one embodiment. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0060] like Figure 1 As shown, this embodiment provides a lane line detection method based on deep learning in low-light environments; including the following steps:

[0061] 1) Obtain the lane line dataset TuSimple, which includes lane line images and their annotation files.

[0062] 2) Preprocess the TuSimple image and divide it into training and validation sets proportionally. Preprocessing includes dividing the TuSimple image into several rows, recording the coordinate range of the regions containing lane lines, and writing them into the TuSimple annotation file. The ratio of the training and validation sets can be adjusted as needed, but the size of the training set should be larger than that of the validation set.

[0063] 3) The training set is fed into the constructed deep learning-based lane detection network for training, and the training effect is verified using the validation set. The optimal lane detection network model weight parameters are determined based on the performance of the lane detection network on the validation set, and the optimal lane detection network model is constructed using these parameters. The deep learning-based lane detection network consists of a ResNet50 backbone network, a CoT attention module, and two MLP (Multilayer Perceptron) modules. The ResNet50 backbone network is used to extract features from the input lane image, and the CoT attention module is used to filter key features from the features extracted by the ResNet50 backbone network and input them into the two MLP modules. One MLP module is used for lane classification, and the other MLP module is used for lane localization.

[0064] When an image is input, the lane detection network first extracts deep features from the segmented rows of the image in step 2) using the ResNet50 backbone network. Then, it filters key features using the CoT attention module, transforms them into high-dimensional feature vectors, flattens them into one-dimensional row vectors, and inputs them into two MLP modules for processing. One MLP module is responsible for lane line localization, finding the coordinates of each pixel on the lane line. The output of this MLP module is sent to a softmax module, which outputs the probability that each coordinate in each input row corresponds to a lane line. The formula for predicting the lane line for that row is:

[0065]

[0066] In the formula, Let Pb be the predicted location value for the i-th row, j be the traversal of all x-coordinates in the input row, J be the set of x-coordinates of all pixels in the input row, and Pb be the predicted location value for the i-th row. i,j It is the probability value at the i-th row and j-th horizontal coordinate, that is, the output value of the softmax module at that position;

[0067] Define a lane line structure loss function to optimize this MLP module:

[0068] L str =λL shape +L sim (2)

[0069] In the formula, L str Let L be the lane line structure loss function. shape Let L be the lane shape loss function. sim Let λ be the lane line pixel similarity loss function, and λ be the lane line pixel similarity loss weight coefficient.

[0070] Define the lane line shape loss function L shape as follows:

[0071]

[0072] In the formula, h is the number of rows after image segmentation, and k is the number of iterations over h. This represents the predicted location value for the k-th row. This represents the predicted location value for the (k+1)th row. ω represents the predicted location value for the (k+2)th row. shape For random disturbances that affect lane line shape loss;

[0073] Define the lane line pixel similarity loss function L sim as follows:

[0074]

[0075] In the formula, a is the number of rows after image segmentation, b is the number of iterations over a, and P b For the output value of the MLP module corresponding to line b, P b+1 For the output value of the MLP module corresponding to line b+1, ω sim For random perturbations to address the loss of similarity between lane line pixels;

[0076] Another MLP module performs classification, determining whether a segmented row from step 2) contains lane line pixels, thus aiding in the localization result. The classification result is then fed into a softmax module, which outputs the predicted classification value. The lane line classification loss function is defined as follows:

[0077]

[0078] In the formula, L cls Let N be the total number of pixels input to the MLP module, n be the number of iterations over N, and Pc represent the output of the softmax module. n The output of the softmax module for the nth pixel;

[0079] The lane detection network combines the outputs of two MLP modules for analysis. When the MLP module used for classification determines that a certain image row does not contain lane lines, if there are corresponding localization prediction values, it considers these prediction values ​​to lack credibility and therefore discards all localization prediction results for that row. Conversely, if the MLP module used for classification confirms that a certain image row contains lane lines, then the lane line prediction results for that row will be calculated using Equation (1).

[0080] After obtaining the final output lane coordinates, the lane detection result is obtained by least-squares fitting using a third-order spiral. The equation to be fitted is as follows:

[0081] y = m0 + m1x + m2x 2 +m3x 3 (6)

[0082] In the formula, y is the vertical coordinate of a point on the lane line in the image, x is the horizontal coordinate of a point on the lane line in the image, and m0, m1, m2, and m3 are the coefficients of the spiral line to be fitted.

[0083] When validating the performance of the lane detection network using a validation set, accuracy is used as the metric. The formula for calculating accuracy (Acc) is as follows:

[0084]

[0085] In the formula, C region S represents the total number of points correctly predicted as lane lines. region This represents the total number of points on the lane line. The lane line detection network model weight parameter that achieves the maximum Acc value on the validation set is selected as the optimal lane line detection network model weight parameter, and the optimal lane line detection network model is constructed using this weight parameter.

[0086] 4) Obtain the multi-exposure image dataset SICE (containing images of the same scene under different exposure conditions) to train the constructed low-light enhancement network and obtain the optimal low-light enhancement network model; such as Figure 2 As shown, the low-light enhancement network has 6 layers. The first 3 layers each consist of a 3x3 convolutional layer with 32 channels, and the last 3 layers each consist of two 3x3 convolutional layers with 32 channels fused together. A loss function is designed to optimize them.

[0087] Loss function L of low-light enhancement network a for:

[0088] L a =W1L s +W2L e +W3L c (8)

[0089] In the formula, W1, W2, and W3 are all weights;

[0090] Spatial consistency loss L s The calculation formula is as follows:

[0091]

[0092] In the formula, K is the number of pixels, c and d are the traversal of pixels, R(c) is the 4-neighborhood of the c-th pixel, Y is the enhanced image, and X is the input image;

[0093] Exposure control loss L e The calculation formula is as follows:

[0094]

[0095] Using 16*16 as a large pixel area, N in the formula r Let F be the number of large pixel regions in the input image, k be the number of large pixel regions traversed, and F be the number of large pixel regions traversed. k It is the average brightness of the k-th large pixel region, and E is a constant located in (0,1);

[0096] Color constant loss L c The calculation formula is as follows:

[0097]

[0098] In the formula, (p,q) iterates through all pairwise combinations η in the three color channels R, G, and B, and J p J represents the average brightness of color channel p. q This represents the average brightness of color channel q;

[0099] Multiple low-light enhancement network models were obtained by training the low-light enhancement network using SICE. The low-light enhancement network model that minimizes the calculated value of equation (8) was selected as the optimal low-light enhancement network model.

[0100] 5) Perform INT8 quantization on the optimal lane detection network model obtained in step 3) and the optimal low-light enhancement network model obtained in step 4), convert them into ONNX format models, and then convert the ONNX format models into TensorRT format models to complete the model deployment.

[0101] 6) Input the road images captured frame by frame by the vehicle's front-facing camera into the deployed low-light enhancement network model to obtain enhanced images. Then, input the enhanced images into the deployed lane detection network model to obtain lane detection results. Plot the results on the input image to visualize the detection results. An example of the detection results is shown below. Figure 3 As shown.

[0102] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A lane line detection method based on deep learning in low-light environments, characterized in that, Includes the following steps: 1) Obtain the lane line dataset TuSimple; 2) Preprocess TuSimple and divide it into training and validation sets proportionally; 3) The training set is fed into the constructed deep learning-based lane detection network for training, and the training effect is verified using the validation set. The optimal lane detection network model weight parameters are determined based on the performance of the lane detection network on the validation set, and the optimal lane detection network model is constructed using these parameters. The deep learning-based lane detection network consists of a ResNet50 backbone network, a CoT attention module, and two MLP modules. The ResNet50 backbone network is used to extract features from the input lane image. The CoT attention module is used to filter the key features extracted by the ResNet50 backbone network and input them into the two MLP modules. One MLP module is used for lane classification, and the other MLP module is used for lane localization. When an image is input, the lane detection network first extracts deep features from the segmented rows of the image in step 2) using the ResNet50 backbone network. Then, it filters key features using the CoT attention module, transforms them into high-dimensional feature vectors, flattens them into one-dimensional row vectors, and inputs them into two MLP modules for processing. One MLP module is responsible for lane line localization, finding the coordinates of each pixel on the lane line. The output of this MLP module is sent to a softmax module, which outputs the probability that each coordinate in each input row corresponds to a lane line. The formula for predicting the lane line for that row is: (1)； In the formula, For the first The predicted value of the line location, It iterates through all the x-coordinates of the input line. It is the set of x-coordinates of all pixels in the input line. It is the first line, number The probability value at each horizontal axis, that is, the output value of the softmax module at that point; 4) Obtain the multi-exposure image dataset SICE to train the constructed low-light enhancement network and obtain the optimal low-light enhancement network model. The low-light enhancement network has 6 layers. The first 3 layers each consist of a 32-channel 3*3 convolutional layer, and the last 3 layers each consist of two 32-channel 3*3 convolutional layers fused together. A loss function is designed to optimize it. 5) Quantize and deploy the optimal lane detection network model obtained in step 3) and the optimal low-light enhancement network model obtained in step 4); 6) Input the road images captured by the vehicle's front camera frame by frame into the deployed low-light enhancement network model to obtain enhanced images. Then, input the enhanced images into the deployed lane detection network model to obtain lane detection results. Finally, plot the results on the input image to visualize the detection results.

2. The lane line detection method based on deep learning in low-light environments according to claim 1, characterized in that, In step 1), the obtained lane line dataset TuSimple includes lane line images and lane line image annotation files.

3. The lane line detection method based on deep learning in low-light environments according to claim 2, characterized in that, In step 2), preprocessing includes dividing the TuSimple image into several rows, recording the coordinate range of the region containing lane lines, and writing it into the TuSimple annotation file; the ratio of the training set and the validation set is adjusted as needed, but the size of the training set should be larger than the size of the validation set.

4. The lane line detection method based on deep learning in low-light environments according to claim 3, characterized in that, In step 3), the lane line structure loss function is defined to optimize the MLP module: (2)； In the formula, The lane line structure loss function is... Let lane shape loss function be used. Let be the loss function for the similarity of lane line pixels. The weighting coefficients for the similarity loss of lane line pixels; Define the lane line shape loss function as follows: (3)； In the formula, The number of rows after image segmentation. To Traversal, Indicates the first The predicted value of the line location, Indicates the first +1 row's location prediction value, Indicates the first +2 rows of location prediction values, For random disturbances that affect lane line shape loss; Define the lane line pixel similarity loss function as follows: (4)； In the formula, The number of rows after image segmentation. To Traversal, For the corresponding number The output value of the MLP module in the row, For the corresponding number The output value of the MLP module in line +1. For random perturbations to address the loss of similarity between lane line pixels; Another MLP module performs classification, determining whether a segmented row from step 2) contains lane line pixels, thus aiding in the localization result. The classification result is then fed into a softmax module, which outputs the predicted classification value. The lane line classification loss function is defined as follows: (5)； In the formula, For lane line classification loss function, The input is the total number of pixels in this MLP module. To Traversal, This represents the output of the softmax module. For the first Each pixel is output through the softmax module; The lane detection network combines the outputs of two MLP modules for analysis. When the MLP module used for classification determines that a certain image row does not contain lane lines, if there are corresponding localization prediction values, it considers these prediction values ​​to lack credibility and therefore discards all localization prediction results for that row. Conversely, if the MLP module used for classification confirms that a certain image row contains lane lines, then the lane line prediction results for that row will be calculated using Equation (1). After obtaining the final output lane coordinates, the lane detection result is obtained by least-squares fitting using a third-order spiral. The equation to be fitted is as follows: (6)； In the formula, Let be the ordinate of a point on the lane line in the image. Let x be the x-coordinate of a point on the lane line in the image. , , , All of these are the coefficients of the spiral curve to be fitted; When validating the performance of the lane detection network using a validation set, accuracy is used as the metric. The calculation formula is: (7)； In the formula, This represents the total number of points correctly predicted as lane lines. This represents the total number of points along the lane line; Select the maximum value on the validation set. The optimal lane detection network model is constructed using the lane detection network model weight parameters.

5. The lane line detection method based on deep learning in low-light environments according to claim 4, characterized in that, In step 4), SICE contains images of the same scene under different exposure conditions; the loss function of the low-light enhancement network. for: (8)； In the formula, , , All are weights; Spatial consistency loss The calculation formula is as follows: (9)； In the formula, It's the number of pixels. , It is a traversal of pixels. It is the first A 4-neighborhood of 1 pixel It is image enhancement. It is the input image; Exposure control loss The calculation formula is as follows: (10)； Using 16*16 as a large pixel area, in the formula... The number of large pixel regions in the input image. It involves traversing large pixel regions. It is the first The average brightness in a large pixel region. For located The constant; Color constant loss The calculation formula is as follows: (11)； In the formula, It iterates through all pairwise combinations in the R, G, and B color channels. , Indicates color channel The average brightness Indicates color channel The average brightness; Multiple low-light enhancement network models were obtained by training the low-light enhancement network using SICE. The low-light enhancement network model that minimizes the calculated value of equation (8) was selected as the optimal low-light enhancement network model.

6. The lane line detection method based on deep learning in low-light environments according to claim 5, characterized in that, In step 5), the optimal lane detection network model obtained in step 3) and the optimal low-light enhancement network model obtained in step 4) are quantized using INT8 and converted into ONNX format models. Then, the ONNX format models are converted into TensorRT format models to complete the deployment of the models.

Images