An improved YOLOX-NANO model and an automotive electronic suspension control method based on this model

Abstract

The present invention discloses an improved YOLOX-NANO model and a vehicle electronic suspension control method based on the model. S1 constructs an obstacle dataset, labels samples, and divides them into a training set and a validation set. S2 selects the optimal model using transfer learning, deploys the optimal model to an embedded platform, and uses TensorRT to accelerate and optimize the model. S3 uses an external camera to detect the road surface in front of the vehicle in real time, and the model detects obstacles. The results are input into a SORT tracker to track the obstacles in a time series. S4 converts the coordinates of each independent tracking target in the image into the actual distance from the vehicle wheel to the obstacle. Combined with the real-time vehicle speed and the geometric dimensions of the obstacle, the start time and the end time of the vehicle wheel's impending contact with the obstacle are calculated. At the start time, the suspension damping is adjusted; at the end time, the suspension damping is adjusted to a normal state, thereby achieving precise optimization of suspension control and improving the driving comfort of the vehicle.

Description

Technical Field

[0001] The present invention belongs to the field of automobile electronic suspension control, and in particular relates to an improved YOLOX-NANO model and an automobile electronic suspension control method based on the model. Background Art

[0002] Manhole covers and speed bumps are common road obstacles on ordinary roads, but due to their specialized functions, they are often difficult to remove, resulting in reduced driving comfort. To improve driving comfort, vehicles need to sense the road surface in real time and adjust the suspension system based on the perceived signals. Existing road surface recognition technologies include direct measurement and inverse road surface recognition based on vehicle signals. However, both methods rely on high-precision sensors, which are generally expensive to manufacture, increasing the cost of actual deployment. Furthermore, maintaining high-precision measurements often requires regular adaptation and maintenance. Furthermore, these two road surface recognition methods cannot identify specific road obstacles, which limits their applicability to complex road conditions. Furthermore, they experience a certain degree of latency during use, failing to meet the accuracy and real-time performance requirements for road surface perception at higher speeds.

[0003] In recent years, with the rapid development of deep learning and convolutional neural networks, people have gradually used artificial intelligence to realize the perception of the car's road surface. However, it has high requirements for computing power. Achieving low-cost and efficient deployment solutions plays an important role in reducing manufacturing costs and improving performance. At the same time, the interaction logic between the existing vehicle perception system and the vehicle suspension control system still needs to be improved.

[0004] Therefore, the present invention proposes an improved YOLOX-NANO model and a low-cost real-time automobile electronic suspension control method based on the model, which is of great significance to improving the driving comfort of the automobile. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an improved YOLOX-NANO model and an electronic suspension control method based on this model, which can enhance the comfort of vehicles on the road. Compared with traditional electronic suspension control methods based on road surface recognition, this invention reduces costs and enhances real-time performance and accuracy, providing a new approach to improving the comfort of vehicles on the road.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] The present invention designs an improved YOLOX-NANO model, the specific structure of which is as follows:

[0008] The improved YOLOX-NANO network consists of three parts: the backbone feature extraction network, the feature enhancement and fusion network, and the detection head network;

[0009] The backbone feature extraction network performs preliminary feature extraction on the image fed into the network and generates feature maps of three different scales;

[0010] The feature enhancement fusion network performs multi-scale feature fusion on the three different scale feature maps generated by the backbone feature extraction network to generate three different scale fused feature maps;

[0011] The detection head network resizes the three fused feature maps of different scales generated by the feature enhancement fusion network and concatenates them along the feature channel dimension to generate a feature map. After passing through the convolution module and the linear layer, the feature map outputs the target type and target prediction box respectively.

[0012] As a preferred implementation scheme, the backbone feature extraction network is composed of eleven basic feature extraction blocks, each of which includes, in order: a 3×3 convolution layer, a batch normalization layer, a ReLU activation function layer, a 5×5 depth-separable convolution layer, a batch normalization layer, and a ReLU activation function layer. At the same time, each basic feature extraction block uses a residual structure to directly add the feature map that has not entered the basic feature extraction block for processing to the feature map obtained after processing by the basic feature extraction block. The eleven basic feature extraction blocks are superimposed to form the backbone feature extraction network of the improved YOLOX-NANO network.

[0013] As a preferred implementation scheme, the three different scale feature maps generated by the backbone feature extraction network are 1 / 4, 1 / 8 and 1 / 16 of the original image size, and are named S3, S4 and S5 respectively.

[0014] As a preferred embodiment, the feature enhanced fusion network structure is as follows Figure 4 As shown in the figure, the inputs are S3, S4, and S5 output by the backbone feature extraction network, respectively. Among them, S5 first performs a two-fold upsampling operation to increase the size of the feature map, and is added to the input S4 to update the value of S4. After the value is updated, S4 has three branch flows: first, S4 passes through a parallel 3×3 convolution layer, a ReLU activation function layer, and a 1×1 convolution layer, a ReLU activation function layer. The two values ​​obtained are added and then upsampled twice and concatenated with S3 along the feature channel dimension, which is recorded as F3; secondly, S4 itself directly passes the value to F4; finally, S4 passes through a parallel 3×3 convolution layer, a ReLU activation function layer, and a 1×1 convolution layer, a ReLU activation function layer. The two values ​​obtained are added and then downsampled twice and concatenated with S5 along the feature channel dimension, which is recorded as F5.

[0015] As a preferred implementation scheme, the detection head network downsamples the feature map F3 output by the feature enhancement fusion network to the size of F4 by two times, upsamples the feature map F5 to the size of F4 by two times, and then splices F3, F4, and F5 along the feature channel dimension. The spliced ​​feature map passes through a 1×1 convolution layer and a ReLU activation function layer in sequence to obtain a feature map P. After passing through two parallel linear layers, the feature map P outputs the type of target and the target prediction box respectively.

[0016] A vehicle electronic suspension control method based on an improved YOLOX-NANO model includes the following steps:

[0017] (1) A dataset of 1262 images was constructed by taking photos of two different obstacles on the road: manhole covers and speed bumps. The dataset was annotated using annotation software, and a training set and a validation set were created based on the annotations. The training set was used to train the model, while the validation set was used to validate the trained model.

[0018] (2) Using transfer learning, the improved YOLOX-NANO model was trained multiple times using the training set, and the trained model was tested using the validation set. The model with the best validation effect was selected, and the selected improved YOLOX-NANO model was deployed on NVIDIA's embedded platform Jetson Nano, and the model was accelerated and optimized using TensorRT;

[0019] (3) The external camera on the Jetson Nano is used to detect the road ahead of the car in real time. The improved YOLOX-NANO model detects two obstacles, manhole covers and speed bumps, respectively. The detection results are input into the SORT tracker to track the manhole covers and speed bumps in the time series.

[0020] (4) The embedded platform Jetson Nano is deployed on the car. When the car is driving, the embedded platform obtains the car tire angle and the current speed of the car from the car electronic control unit, and calculates the real-time distance between the actual car wheel and the tracked manhole cover or speed bump based on the relative position of the camera in the car body coordinate system and the relative position of the tracking target in the current video frame. At the same time, combined with the real-time speed of the car and the geometric dimensions of the manhole cover or speed bump, the starting time when the car wheel is about to contact the manhole cover or speed bump and the end time when the car wheel ends contact with the manhole cover or speed bump are calculated respectively. At the starting time node of the contact, the embedded platform sends a signal to the electronic suspension controller to adjust the damping of the suspension; at the end time node of the end of the contact, the embedded platform sends a signal to the electronic suspension controller again to adjust the damping of the suspension to a normal state, thereby achieving precise optimization of the suspension control and improving the driving comfort of the car.

[0021] As a preferred implementation scheme, in step (1), the constructed dataset includes two common obstacles on the road: manhole covers and speed bumps. All images in the dataset are in JPG format. In order to enhance the versatility of the trained model, photos of manhole covers and speed bumps taken from different angles, different distances, different lighting conditions, and in a variety of road scenes are required as dataset samples.

[0022] As a preferred implementation scheme, in step (1), the two categories of road manhole covers and speed bumps in the data set are labeled separately using LabelImg labeling software, and the label names are Manhole cover and Deceleration zone respectively. After the labeling is completed, the corresponding XML format labeling file is generated. Each XML labeling file records the category information and location information of the manhole cover or speed bump in the corresponding image sample. Finally, the data set is randomly divided into a training set and a validation set in a ratio of 9:1.

[0023] As a preferred embodiment, in step (2), before using the dataset to train the model, the dataset needs to be preprocessed, including changing the original length and width of the image and filling the edges of the image to reduce the amount of computation required for model inference.

[0024] As a preferred implementation scheme, in step (2), the weight file obtained by training the improved YOLOX-NANO model on the COCO dataset is used as the initial training weight to achieve transfer learning.

[0025] As a preferred implementation scheme, in step (2), the improved YOLOX-NANO model first preliminarily extracts features through the backbone network, and then fuses the extracted multi-scale features to improve the network's detection effect on objects of different sizes. The fused features of different scales are sent to the detection head, and the detection head adopts a decoupled form. The decoupled branches respectively realize the classification and regression tasks.

[0026] As a preferred implementation scheme, in the step (2), during the training of the improved YOLOX-NANO model, due to the small model depth and the limited amount of data in the data set, in order to obtain a model with optimal detection performance, it is necessary to avoid false detection caused by overfitting of the model. For this purpose, different training strategies are adopted and different numbers of training rounds are set to train the improved YOLOX-NANO model. The specific training steps are as follows: 70 rounds, 100 rounds, 120 rounds, and 130 rounds of training rounds are selected respectively. For each training process with a different number of training rounds, control experiments with and without Mixup data enhancement are carried out respectively, and the model with the highest accuracy verified by the validation set is selected from all experimental results as the final detection model.

[0027] As a preferred embodiment, in step (2), Jetson Nano is an embedded platform with low computing power. In order to achieve real-time detection, the improved YOLOX-NANO model deployed on Jetson Nano needs to be accelerated and optimized for forward reasoning using TensorRT. First, the improved YOLOX-NANO model needs to be converted into ONXX format, and then converted from ONXX format to TensorRT format.

[0028] As a preferred embodiment, in step (3), the working principle of the external camera is:

[0029] The camera is connected to the Jetson Nano via the CSI interface;

[0030] The camera reads the input road scene video stream in real time through OpenCV and sends it to the improved YOLOX-NANO model for inference.

[0031] As a preferred implementation scheme, in step (3), the SORT tracker determines whether the objects are the same object based on the IOU values ​​between the objects in the video frames at different times and realizes target tracking.

[0032] As a preferred embodiment, in step (4), in order to enable the embedded platform Jetson Nano to obtain a relatively accurate positional relationship between the vehicle body coordinate system and the ground coordinate system, it is necessary to deploy the embedded platform at the front end of the vehicle chassis and keep the camera shooting angle at a fixed angle with the horizontal cross-section of the vehicle body.

[0033] As a preferred embodiment, in step (4), the embedded platform obtains the vehicle tire angle and the current speed of the vehicle from the vehicle electronic control unit, and calculates the real-time distance between the actual vehicle wheel and the tracked manhole cover or speed bump based on the relative position of the camera in the vehicle body coordinate system and the relative position of the tracking target in the current video frame, and on this basis predicts the starting time when the vehicle wheel is about to contact the manhole cover or speed bump.

[0034] As a preferred embodiment, in step (4), the actual dimensions of the manhole cover and the speed bump are set based on experience, the embedded platform obtains the vehicle tire angle and the current speed of the vehicle from the vehicle electronic control unit, and brings the empirical dimensions of the manhole cover or speed bump into the calculation to obtain the time when the vehicle wheel passes through the manhole cover or speed bump, and the start time when the wheel is about to contact the manhole cover or speed bump is added to the time when the vehicle wheel passes through the manhole cover or speed bump to obtain the end time when the vehicle wheel ends contact with the manhole cover or speed bump.

[0035] As a preferred embodiment, in step (4), when the embedded platform calculates the starting time point when the wheel is about to contact the manhole cover or speed bump, it reserves a certain time according to the adjustment speed of the electronically controlled suspension damping, and the electronically controlled suspension completes the damping adjustment before the starting time point when the wheel contacts the manhole cover or speed bump; and at the end time point when the vehicle wheel ends contacting the manhole cover or speed bump, the electronically controlled suspension adjusts the damping to a normal state.

[0036] Beneficial effects of the present invention:

[0037] The present invention adopts an improved YOLOX-NANO model and an automotive electronic suspension control method based on this model. By using a low-cost visual camera and the Jetson Nano platform, it overcomes the traditional road recognition-based electronic suspension control method's reliance on high-precision sensors, reduces deployment costs, and overcomes the defect of poor real-time performance. It can track road manhole covers and speed bumps during actual driving, and accurately adjust the suspension damping in real time during the tracking process, providing a new method for improving vehicle driving comfort. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the workflow for inputting road images into an embedded system and optimizing the control of the vehicle's electronic suspension controller.

[0039] Figure 2 Schematic diagram of the process for improving YOLOX-NANO model training, validation, and model selection.

[0040] Figure 3 The following are examples of results using the improved YOLOX-NANO model detection on an embedded platform.

[0041] Figure 4 Schematic diagram of the feature enhancement fusion network for improving the YOLOX-NANO model. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0043] Example 1 Preparation and annotation of road manhole cover and speed bump dataset and improvement of YOLOX-NANO network training Figure 2 The overall process of data set preparation and network training includes the following steps:

[0044] (1) Dataset collection: A dataset is constructed by taking photos of two different obstacles on the road, manhole covers and speed bumps, using a camera. The total number of images in the dataset is 1,262, and all images are in JPG format. In order to enhance the generality of the trained model, the dataset is required to include samples of photos of manhole covers and speed bumps taken at different angles, distances, lighting conditions, and in a variety of road scenarios.

[0045] (2) Labeling dataset: The LabelImg labeling software is used to label the two categories of road manhole covers and speed bumps in the dataset. The label names are Manhole cover and Deceleration zone respectively. After the labeling is completed, the corresponding XML format labeling files are generated. Each XML labeling file records the category information and location information of the manhole cover or speed bump in the corresponding image sample.

[0046] (3) Dataset division: The dataset is randomly divided into a training set and a validation set in a ratio of 9:1.

[0047] (4) Training the improved YOLOX-NANO model: Before using the dataset to train the model, the dataset needs to be preprocessed, including changing the original length and width of the image and filling the edges of the image to reduce the amount of computational complexity of model inference; the weight file obtained by training the improved YOLOX-NANO model on the COCO dataset is used as the initial training weight to achieve transfer learning; during the training process of the improved YOLOX-NANO model, due to the small model depth and the limited amount of data in the dataset, in order to obtain the model with the best detection performance, it is necessary to avoid false detection caused by overfitting of the model. For this purpose, different training strategies and different numbers of training rounds are used to train the improved YOLOX-NANO model. The specific training steps are as follows: 70 rounds, 100 rounds, 120 rounds, and 130 rounds of training rounds are selected respectively. For each training process with different numbers of training rounds, control experiments are carried out with and without Mixup data augmentation. (5) Model verification and selection: Among all the trained models, each model is individually verified for model accuracy using a validation set. After all models are verified, the model with the highest verification accuracy is selected as the final prediction model.

[0048] Example 2: Improved deployment and optimization acceleration of the YOLOX-NANO model and deployment of the embedded platform Jetson Nano in a car

[0049] The selected improved YOLOX-NANO model is deployed on NVIDIA's Jetson Nano embedded platform. To address the problem of slow model inference speed caused by the low computing power of the Jetson Nano platform, it is necessary to first convert the improved YOLOX-NANO model into ONXX format, and then convert the ONXX format into TensorRT format to achieve inference acceleration optimization of the improved YOLOX-NANO model to meet the needs of real-time detection; the deployed model is used to perform prediction tests on road manhole covers and speed bumps. The prediction results are as follows: Figure 3 As shown in (a) and (b), the manhole cover is outlined in a blue prediction box, while the speed bump is outlined in a yellow prediction box.

[0050] In order for the embedded platform Jetson Nano to obtain a relatively accurate positional relationship between the vehicle body coordinate system and the ground coordinate system, it is necessary to deploy the embedded platform at the front end of the vehicle chassis and maintain a fixed angle between the camera's shooting angle and the horizontal cross-section of the vehicle body.

[0051] Example 3 uses an embedded platform for real-time detection and tracking. The embedded platform sends different control signals to the electronic suspension controller based on the tracking results and the different time points of contact between the car wheel and the manhole cover or speed bump.

[0052] Figure 1 Schematic diagram of the workflow for inputting road images into the embedded system and optimizing the control of the vehicle's electronic suspension controller:

[0053] (1) A visual camera with a CSI interface is connected to the embedded platform Jetson Nano. The visual camera uses OpenCV to read the real-time image of the road ahead of the car and sends it to the improved YOLOX-NANO model for inference calculation.

[0054] (2) Use the improved YOLOX-NANO model accelerated by TensorRT to perform forward inference calculations on the input real-time video stream and generate prediction results.

[0055] (3) The detection results are input into the SORT tracker to track the road manhole covers and speed bumps in the time series.

[0056] (4) If the embedded platform does not continue to track the manhole cover or speed bump, it returns to step (1), and the embedded platform continues to perform real-time reading of the camera and real-time inference calculation of the model.

[0057] (5) If the embedded platform continues to track the manhole cover or speed bump, the embedded platform obtains the vehicle tire angle and the current speed of the vehicle from the vehicle electronic control unit, and calculates the real-time distance between the actual vehicle wheel and the tracked manhole cover or speed bump based on the relative position of the camera in the vehicle body coordinate system and the relative position of the tracking target in the current video frame, and predicts the starting time when the vehicle wheel is about to contact the manhole cover or speed bump; the actual size of the manhole cover and speed bump is set according to experience, and the embedded platform again obtains the vehicle tire angle and the current speed of the vehicle from the vehicle electronic control unit, and brings the empirical size of the manhole cover or speed bump into the calculation to obtain the time when the vehicle wheel passes through the manhole cover or speed bump, and adds the starting time when the wheel is about to contact the manhole cover or speed bump to the time when the vehicle wheel passes through the manhole cover or speed bump to obtain the end time when the vehicle wheel ends contact with the manhole cover or speed bump.

[0058] (6) When the embedded platform is at the starting time node where the vehicle wheel is about to come into contact with the manhole cover or speed bump, the embedded platform sends a signal to the electronic suspension controller in advance according to the adjustment completion speed of the electronic suspension damping, and the electronic suspension controller adjusts the suspension damping from ζ1 to ζ2, where ζ1 < ζ2, and the electronic suspension damping value increases. When the wheel passes through the manhole cover or speed bump, the vehicle driving stability is enhanced while the vibration filtering ability is enhanced. When the embedded platform is at the ending time node where the vehicle wheel ends contacting the manhole cover or speed bump, the embedded platform sends a signal to the electronic suspension controller again, and the electronic suspension controller adjusts the suspension damping from ζ2 to ζ1, and the electronic suspension damping value decreases to meet the comfort requirements of the vehicle driving on normal roads, realize the precise optimization of the suspension control, and improve the driving comfort of the vehicle. Finally, the embedded platform returns to step (1), and continues to perform real-time reading of the camera and real-time inference calculation of the model.

[0059] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent methods or changes that do not deviate from the technology of the present invention should be included in the scope of protection of the present invention.

Claims

1. An improved YOLOX-NANO model, characterized in that: The specific structure of the model is as follows: It consists of three parts: backbone feature extraction network, feature enhancement fusion network and detection head network; The backbone feature extraction network performs preliminary feature extraction on the image fed into the network, and generates three feature maps of different scales, namely 1 / 4, 1 / 8, and 1 / 16 of the original image, named S3, S4, and S5 respectively. The backbone feature extraction network consists of eleven basic feature extraction blocks, each of which contains, in order: a 3×3 convolution layer, a batch normalization layer, a ReLU activation function layer, a 5×5 depth-separable convolution layer, a batch normalization layer, and a ReLU activation function layer. At the same time, each basic feature extraction block uses a residual structure to directly add the feature map that has not entered the basic feature extraction block to the feature map obtained after processing by the basic feature extraction block. The eleven basic feature extraction blocks are superimposed to form the backbone feature extraction network of the improved YOLOX-NANO network. The feature enhancement fusion network performs multi-scale feature fusion on the three different scale feature maps generated by the backbone feature extraction network to generate three different scale fusion feature maps. Its inputs are S3, S4, and S5 output by the backbone feature extraction network, respectively. Among them, S5 first performs a two-fold upsampling operation to increase the size of the feature map, and is added to the input S4 to update the value of S4. After the updated value, S4 has three branch flows: first, S4 passes through a parallel 3×3 convolution layer, a ReLU activation function layer, and a 1×1 convolution layer, a ReLU activation function layer. The two values ​​obtained are added and then upsampled twice and concatenated with S3 along the feature channel dimension, which is recorded as F3; secondly, S4 itself directly passes the value to F4; finally, S4 passes through a parallel 3×3 convolution layer, a ReLU activation function layer, and a 1×1 convolution layer, a ReLU activation function layer. The two values ​​obtained are added and then downsampled twice and concatenated with S5 along the feature channel dimension, which is recorded as F5; The detection head network resizes the three fused feature maps of different scales generated by the feature enhancement fusion network and splices them along the feature channel dimension to generate a feature map. After passing through the convolution module and two parallel linear layers, the feature map outputs the type of target and the target prediction box respectively. The detection head network downsamples the feature map F3 output by the feature enhancement fusion network to the size of F4 by twice, and upsamples the feature map F5 to the size of F4 by twice, and then splices F3, F4, and F5 along the feature channel dimension. The spliced ​​feature map passes through a 1×1 convolution layer and a ReLU activation function layer in sequence to obtain a feature map P. After passing through two parallel linear layers, the feature map P outputs the type of target and the target prediction box respectively.

2. A vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 1, characterized in that: The steps include: S1 samples two different obstacles on the road: manhole covers and speed bumps, and constructs a dataset. It uses annotation software to annotate the manhole covers and speed bumps in the dataset, and then divides the dataset into a training set and a validation set. The training set is used to train the model, and the validation set is used to verify the trained model. S2 uses transfer learning to train the improved YOLOX-NANO model multiple times using the training set. The trained model is then tested using the validation set. The model with the best validation results is selected and deployed on NVIDIA's embedded platform, Jetson Nano, where it is accelerated and optimized using TensorRT. S3 uses the external camera on the Jetson Nano to monitor the road ahead in real time. It then uses the improved YOLOX-NANO model to detect two obstacles, manhole covers and speed bumps. The detection results are then fed into the SORT tracker to track the manhole covers and speed bumps in a time series. S4 deploys the Jetson Nano embedded platform on a car. While the car is driving, the embedded platform obtains the tire angle and current speed from the car's electronic control unit. It then calculates the real-time distance between the actual car wheel and the tracked manhole cover or speed bump based on the camera's relative position in the vehicle's coordinate system and the relative position of the tracked target in the current video frame. Combined with the car's real-time speed and the geometry of the manhole cover or speed bump, it calculates the start time of impending contact between the car wheel and the manhole cover or speed bump and the end time of contact. At the start time of contact, the embedded platform sends a signal to the electronic suspension controller to adjust the suspension damping. At the end time of contact, the embedded platform sends another signal to the electronic suspension controller to adjust the suspension damping to its normal state, achieving precise optimization of suspension control and improving driving comfort.

3. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S1, the constructed dataset includes two types of obstacles on the road: manhole covers and speed bumps. All images in the dataset are in JPG format. In order to enhance the generality of the trained model, photos of manhole covers and speed bumps taken from different angles, different distances, different lighting conditions and in various road scenes are used as dataset samples.

4. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S1, the LabelImg annotation software labels the two categories of road manhole covers and speed bumps in the dataset respectively, with the label names Manhole cover and Deceleration zone respectively. After the annotation is completed, the corresponding XML format annotation files are generated. Each XML annotation file records the category information and location information of the manhole cover or speed bump in the corresponding image sample. Finally, the dataset is randomly divided into a training set and a validation set in a ratio of 9:

1.

5. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S2, before using the dataset to train the model, the dataset needs to be preprocessed, including changing the original length and width of the image and filling the edges of the image to reduce the amount of computation required for model inference.

6. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S2, the weight file obtained by training the improved YOLOX-NANO model on the COCO dataset is used as the initial training weight to achieve transfer learning.

7. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S2, the improved YOLOX-NANO model first uses the backbone network to preliminarily extract features, and then fuses the extracted multi-scale features to improve the network's detection effect on objects of different sizes. The fused features of different scales are sent to the detection head. The detection head adopts a decoupled form, and the decoupled branches respectively realize classification and regression tasks.

8. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S2, during the training of the improved YOLOX-NANO model, the model depth is small and the amount of data in the data set is limited. In order to obtain a model with optimal detection performance, it is necessary to avoid false detection caused by overfitting of the model. Different training strategies and different numbers of training rounds are used to train the improved YOLOX-NANO model. The specific training steps are as follows: 70, 100, 120, and 130 training rounds are selected respectively. For each training process with a different number of training rounds, control experiments are carried out with and without Mixup data enhancement. Among the results obtained from all experiments, the model with the highest accuracy verified by the validation set is selected as the final detection model.

9. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In the S3, the camera is connected to the Jetson Nano via the CSI interface. The camera reads the input road scene video stream in real time through OpenCV and feeds it into the improved YOLOX-NANO model for inference. The SORT tracker determines whether the objects in the video frames at different times are the same object based on the IOU values ​​between them and implements target tracking.

10. The vehicle electronic suspension control method based on the improved YOLOX-NANO model according to claim 2, characterized in that: In S4, in order to enable the embedded platform Jetson Nano to obtain the precise positional relationship between the vehicle body coordinate system and the ground coordinate system, the embedded platform is deployed at the front end of the vehicle chassis, and the camera shooting angle is kept at a fixed angle with the horizontal cross section of the vehicle body; The embedded platform obtains the vehicle's tire angle and current speed from the vehicle's electronic control unit. Based on the camera's relative position in the vehicle's coordinate system and the relative position of the tracked target in the current video frame, it calculates the real-time distance between the actual vehicle wheel and the tracked manhole cover or speed bump. Based on this, it predicts the start time of the vehicle wheel's impending contact with the manhole cover or speed bump. The actual dimensions of the manhole cover and speed bump are set based on experience. The embedded platform obtains the vehicle's tire angle and current speed from the vehicle's electronic control unit. It then uses the empirical dimensions of the manhole cover or speed bump into its calculations to determine the time it takes for the vehicle's wheels to pass through the manhole cover or speed bump. The final time it takes for the vehicle's wheels to stop contact with the manhole cover or speed bump is calculated by adding the start time of the wheel's approach to contact with the manhole cover or speed bump to the time it takes the vehicle's wheels to pass through the manhole cover or speed bump. When the embedded platform calculates the starting time point when the wheel is about to contact the manhole cover or speed bump, it reserves a certain time according to the adjustment speed of the electronic suspension damping and sends a signal to the electronic suspension controller in advance. The electronic suspension controller adjusts the damping of the suspension from ζ1 to ζ2, where ζ1<ζ2. The damping value of the electronic suspension increases. When the wheel passes through the manhole cover or speed bump, the driving stability of the vehicle is enhanced while the vibration filtering ability is enhanced. When the embedded platform reaches the end time node when the vehicle wheel ends contact with the manhole cover or speed bump, the signal is sent to the electronic suspension controller again. The electronic suspension controller adjusts the damping of the suspension from ζ2 to ζ1, and the damping value of the electronic suspension is reduced to meet the comfort requirements of the vehicle on normal roads.

Images