An automatic driving method and device based on the Internet of Things and a vehicle-mounted processor

Abstract

The application relates to the technical field of Internet of Things automatic driving, and discloses an automatic driving method and device based on the Internet of Things and a vehicle-mounted processor, the steps of the method comprising the following steps: crawling a training data set from a network; training a deep learning network model by using the training data set, and constructing an automatic driving network model; the automatic driving network model is composed of a feature extraction module, a spatial feature pooling module and a detection module; the automatic driving network model is used to simultaneously detect a target and a lane line, and to segment a drivable area; a vehicle navigation path is calculated according to the detection and segmentation results; and the vehicle is automatically driven according to the vehicle navigation path. In the method, the improved automatic driving network model and the input proportional function can be used to simultaneously perform target detection, lane line detection and drivable area segmentation, so that a vehicle driving path is planned according to a model output result, and the vehicle is automatically driven, thereby effectively solving the technical problem of inaccurate environment sensing in the prior art.

Description

Technical Field

[0001] This invention relates to the field of IoT autonomous driving technology, and in particular to an IoT-based autonomous driving method, device, and on-board processor. Background Technology

[0002] The Internet of Things (IoT) refers to the interconnection of multiple devices via the internet, and autonomous vehicles utilize this connectivity technology when updating algorithms based on user data. Autonomous vehicles need to collect and process vast amounts of data; in this context, the IoT allows them to share road information, including actual routes, traffic conditions, and obstacle avoidance strategies. All this data is shared among IoT-connected vehicles and wirelessly uploaded to a cloud system for analysis and use, thereby increasing the level of automation.

[0003] To ensure driving safety, the vehicle first needs to capture information about the surrounding driving environment through sensors. The perception system correctly perceives and understands the complex scene information such as the surrounding environment, dynamic and static traffic participants, traffic lights, and traffic signs based on the sensor data. The planning and decision-making system then further solves the feasible strategies for the current vehicle based on the perception and understanding results. The control system then determines the best strategy and outputs vehicle control signals. Finally, without driver intervention, the vehicle performs automatic acceleration, deceleration, or steering operations based on the control signals to complete automatic driving.

[0004] In recent years, innovations in artificial intelligence technology have further propelled the development of the autonomous driving field. For single-task environmental perception methods, object detection algorithms are typically divided into two-stage object detection and single-stage object detection. Two-stage object detection algorithms first extract candidate boxes from the input image, and then fit the final detection result based on the candidate boxes selected in the first stage. This can better handle the problem of class imbalance and has higher detection accuracy. However, due to the need for two-stage calculations, the speed is relatively slow and cannot meet the needs of real-time detection scenarios. Driving area segmentation is generally also a semantic segmentation task, which performs binary classification on the input pixels to segment the background and the drivable area.

[0005] However, autonomous driving scenarios require the simultaneous acquisition of multiple types of information on the road. Single-task environmental perception methods cannot meet the requirements, and activation functions have poor scalability. Therefore, multi-task autonomous driving environmental perception algorithms and highly scalable activation functions have become urgent needs for developers. Summary of the Invention

[0006] The main objective of this invention is to provide an IoT-based autonomous driving method, device, and in-vehicle processor, aiming to solve the technical problem of inaccurate environmental perception in the prior art.

[0007] To achieve the above objectives, embodiments of the present invention provide an autonomous driving method based on the Internet of Things, which includes: crawling a training dataset from the network; A deep learning network model is trained using a training dataset to construct an autonomous driving network model; the autonomous driving network model consists of a feature extraction module, a spatial feature pooling module, and a detection module. YOLOv5's CSPDarknet is used as the feature extraction module. A 3-channel image is taken as input, and the feature map obtained after multiple convolutions is 1 / 32 the size of the original image. The YOLOv5 feature extraction module replaces the focus layer with an equivalent convolutional layer, replaces BottleneckCSP with module C3, and removes the convolutional layer after the concat layer. The feature extraction module encapsulates three modules: convolution, batch normalization, and activation functions. The activation functions are shown below: , where x represents the input of the function; The spatial feature pooling module includes a fast spatial convolutional pooling layer and a feature pyramid layer. The fast spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPoll layers, and an additive layer, used to extract feature maps of different sizes and fuse different features. The input to the feature pyramid layer is the output of the fast spatial convolutional pooling layer, which is 1 / 32 of the original size. After passing through multiple convolutional layers and upsampling layers with a scale parameter of 2, and then through a Concact layer to fuse features of different scales from the encoder, the feature pyramid is finally obtained and used for downstream detection and segmentation tasks. The detection module contains three detection heads, each composed of a generative adversarial network (GAN). The generation method of the GAN includes: creating a random function to generate data following a distribution P. Noisedata Random data Nx; establish sampling rules to obtain P that follows the overall distribution of the true data. Realdata The actual sampled data Rx is input into the discriminator D to train the discriminator D and obtain the parameter set D. θ The random data Nx is input into the generator G, which outputs generator sample data Gx that follows the Pg distribution. The generator sample data Gx is then input into the discriminator D. It is determined whether the difference between the value P(D(x)) output by the discriminator D and the preset probability value is greater than the preset threshold. If so, the generator gradient obtained by the discriminator D is fed back to the generator G. If not, the discriminator continues to be trained. The autonomous driving network model is used to simultaneously detect targets and lane lines, and segment drivable areas. Calculate the vehicle navigation path based on the detection and segmentation results; then perform autonomous driving based on the vehicle navigation path.

[0008] The IoT-based autonomous driving method described in this invention further includes: a spatial convolutional pooling layer consisting of two standard convolutional layers, three MaxPool layers, and an additive layer, wherein the kernel_size of the MaxPool layer is uniformly set to 5, and the three MaxPool layers are connected in a cascading manner.

[0009] The IoT-based autonomous driving method described in this invention further includes: constructing a loss function when training the autonomous driving network model, and adjusting the parameters of the autonomous driving network model according to the loss function.

[0010] The IoT-based autonomous driving method described in this invention further includes: denoted as L for the loss of the target detection branch, which includes a classification loss L. class Regression loss L box And confidence loss L object The weighted average of the three loss classes constitutes the loss for the object detection part, as shown below: L=aL class +bL box +cL object Where a, b, and c represent the classification loss L, respectively. class Regression loss L box And confidence loss L object The corresponding weights.

[0011] The IoT-based autonomous driving method described in this invention further includes: using binary cross-entropy loss for classification loss, calculated as follows: ; , where x i y represents the predicted value for the current category, N represents the total number of categories, and y represents the predicted value for the current category. i y represents the probability of the current class after the activation function. i * This represents the true value of the current category.

[0012] The IoT-based autonomous driving method described in this invention also includes: calculating the vehicle navigation path using a shortest path search algorithm and a heuristic function search algorithm.

[0013] This invention also provides an IoT-based autonomous driving device, comprising: an acquisition module for crawling a training dataset from the network; a model building module for training a deep learning network model using the training dataset to build an autonomous driving network model; the autonomous driving network model consists of a feature extraction module, a spatial feature pooling module, and a detection module. YOLOv5's CSPDarknet is used as the feature extraction module. A 3-channel image is taken as input, and the feature map obtained after multiple convolutions is 1 / 32 the size of the original image. The YOLOv5 feature extraction module replaces the focus layer with an equivalent convolutional layer, replaces BottleneckCSP with module C3, and removes the convolutional layer after the concat layer. The feature extraction module encapsulates three modules: convolution, batch normalization, and activation functions. The activation functions are shown below: , where x represents the input of the function; The spatial feature pooling module includes a fast spatial convolutional pooling layer and a feature pyramid layer. The fast spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPoll layers, and an additive layer, used to extract feature maps of different sizes and fuse different features. The input to the feature pyramid layer is the output of the fast spatial convolutional pooling layer, which is 1 / 32 of the original size. After passing through multiple convolutional layers and upsampling layers with a scale parameter of 2, and then through a Concact layer to fuse features of different scales from the encoder, the feature pyramid is finally obtained and used for downstream detection and segmentation tasks. The detection module contains three detection heads, each composed of a generative adversarial network (GAN). The generation method of the GAN includes: creating a random function to generate data following a distribution P. Noisedata Random data Nx; establish sampling rules to obtain P that follows the overall distribution of the true data. Realdata The actual sampled data Rx is input into the discriminator D to train the discriminator D and obtain the parameter set D. θ The random data Nx is input into the generator G, which outputs generator sample data Gx that follows the Pg distribution. The generator sample data Gx is then input into the discriminator D. It is determined whether the difference between the value P(D(x)) output by the discriminator D and the preset probability value is greater than the preset threshold. If so, the generator gradient obtained by the discriminator D is fed back to the generator G. If not, the discriminator continues to be trained. The detection and segmentation module is used to simultaneously detect targets and lane lines and segment drivable areas using an autonomous driving network model. The calculation module is used to calculate the vehicle navigation path based on the detection and segmentation results; The control module is used to control the vehicle's autonomous driving based on the vehicle's navigation path.

[0014] The IoT-based autonomous driving device described in this invention also includes: The spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPool layers, and an additive layer. The kernel_size of the MaxPool layer is uniformly set to 5, and the three MaxPool layers are connected in a cascade manner.

[0015] The IoT-based autonomous driving device described in this invention further includes: constructing a loss function when training the autonomous driving network model, and adjusting the parameters of the autonomous driving network model according to the loss function.

[0016] The IoT-based autonomous driving device of the present invention further includes: the loss of the target detection branch is denoted as L, and the loss of the target detection branch includes the classification loss L. class Regression loss L box And confidence loss L object The weighted average of the three loss classes constitutes the loss for the object detection part, as shown below: L=aL class +bL box +cL object Where a, b, and c represent the classification loss L, respectively. class Regression loss L box And confidence loss L object The corresponding weights.

[0017] The IoT-based autonomous driving device described in this invention further includes: a classification loss using binary cross-entropy loss, calculated as follows: ; , where x i y represents the predicted value for the current category, N represents the total number of categories, and y represents the predicted value for the current category. i y represents the probability of the current class after the activation function. i * This represents the true value of the current category.

[0018] The IoT-based autonomous driving device described in this invention also includes: calculating the vehicle navigation path using a shortest path search algorithm and a heuristic function search algorithm.

[0019] The present invention also provides an in-vehicle processor, wherein the in-vehicle processor executes a program to implement the steps of the above-described Internet of Things-based autonomous driving method.

[0020] The present invention also provides a storage medium storing processor-executable instructions, which are loaded by one or more vehicle-mounted processors to execute the above-described IoT-based autonomous driving method.

[0021] The IoT-based autonomous driving method of this invention, through an improved autonomous driving network model, can simultaneously perform target detection, lane detection, and drivable area segmentation, thereby planning the vehicle's driving path based on the model output and controlling the vehicle to drive autonomously. Furthermore, the activation function used in this application incorporates sine calculation, and further, based on the sine calculation, an input scaling function is added. This enhances the accuracy of feature extraction. By periodically adjusting the input through the sine of the function's input, the scalarity and periodicity of the activation function are greatly enhanced, thus enabling the operability of autonomous driving. Attached Figure Description

[0022] Figure 1 This is the implementation process of the embodiment of the present invention; Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Reference Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the document generation method of the present invention. The IoT-based autonomous driving method of the present invention may include: To achieve the above objectives, embodiments of the present invention provide an autonomous driving method based on the Internet of Things, which includes: crawling a training dataset from the network; A deep learning network model is trained using a training dataset to construct an autonomous driving network model. The autonomous driving network model consists of a feature extraction module, a spatial feature pooling module, and a detection module. YOLOv5's CSPDarknet is used as the feature extraction module. A 3-channel image is taken as input, and the feature map obtained after multiple convolutions is 1 / 32 the size of the original image. The YOLOv5 feature extraction module replaces the focus layer with an equivalent convolutional layer, replaces BottleneckCSP with module C3, and removes the convolutional layer after the concat layer. The feature extraction module encapsulates three modules: convolution, batch normalization, and activation functions. The activation functions are shown below: , where x represents the input of the function; The spatial feature pooling module includes a fast spatial convolutional pooling layer and a feature pyramid layer. The fast spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPoll layers, and an additive layer, used to extract feature maps of different sizes and fuse different features. The input to the feature pyramid layer is the output of the fast spatial convolutional pooling layer, which is 1 / 32 of the original size. After passing through multiple convolutional layers and upsampling layers with a scale parameter of 2, and then through a Concact layer to fuse features of different scales from the encoder, the final feature pyramid is obtained and used for downstream detection and segmentation tasks. The detection module contains three detection heads, each composed of a generative adversarial network (GAN). The generation method of the GAN includes: creating a random function to generate data following a distribution P. Noisedata Random data Nx; establish sampling rules to obtain P that follows the overall distribution of the true data. Realdata The actual sampled data Rx is input into the discriminator D to train the discriminator D and obtain the parameter set D. θ The random data Nx is input into the generator G, which outputs generator sampled data Gx that follows a Pg distribution. The generator sampled data Gx is then input into the discriminator D. It is determined whether the difference between the output value P(D(x)) of the discriminator D and the preset probability value is greater than a preset threshold. If so, the generator gradient obtained by the discriminator D is fed back to the generator G; otherwise, the discriminator continues to be trained. The spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPool layers, and an additive layer. The kernel_size of the MaxPool layer is uniformly set to 5, and the three MaxPool layers are connected in a cascade manner. When training the autonomous driving network model, a loss function is constructed, and the parameters of the autonomous driving network model are adjusted according to the loss function. The loss of the object detection branch is denoted as L, and the loss of this part includes the classification loss L. class Regression loss L box And confidence loss L object The weighted average of the three loss classes constitutes the loss for the object detection part, as shown below: L=aL class +bL box +cL object Where a, b, and c represent the classification loss L, respectively. class Regression loss L box And confidence loss L object The corresponding weights; the classification loss uses binary cross-entropy loss, calculated as follows: ; , where x i y represents the predicted value for the current category, N represents the total number of categories, and y represents the predicted value for the current category. i y represents the probability of the current class after the activation function.i * This represents the true value of the current category.

[0026] The autonomous driving network model is used to simultaneously detect targets and lane lines, and segment drivable areas. Calculate the vehicle navigation path based on the detection and segmentation results; then perform autonomous driving based on the vehicle navigation path.

[0027] In addition, the present invention also provides an autonomous driving device based on the Internet of Things.

[0028] The present invention provides an IoT-based autonomous driving device, comprising: an acquisition module for crawling a training dataset from the network; a model building module for training a deep learning network model using the training dataset to build an autonomous driving network model; the autonomous driving network model consists of a feature extraction module, a spatial feature pooling module, and a detection module. YOLOv5's CSPDarknet is used as the feature extraction module. A 3-channel image is taken as input, and the feature map obtained after multiple convolutions is 1 / 32 the size of the original image. The YOLOv5 feature extraction module replaces the focus layer with an equivalent convolutional layer, replaces BottleneckCSP with module C3, and removes the convolutional layer after the concat layer. The feature extraction module encapsulates three modules: convolution, batch normalization, and activation functions. The activation functions are shown below: , where x represents the input of the function; the spatial feature pooling module includes a fast spatial convolutional pooling layer and a feature pyramid layer. The fast spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPoll layers, and an additive layer, used to extract feature maps of different sizes and fuse different features; the input of the feature pyramid layer is the output of the fast spatial convolutional pooling layer, which is 1 / 32 of the original size. After passing through multiple convolutions and upsampling layers with a scale parameter of 2, and after passing through the Concact layer, it fuses features of different scales with the encoder to finally obtain the feature pyramid, which is used for downstream detection and segmentation tasks; The detection module contains three detection heads, each composed of a generative adversarial network (GAN). The generation method of the GAN includes: creating a random function to generate data following a distribution P. Noisedata Random data Nx; establish sampling rules to obtain P that follows the overall distribution of the true data. Realdata The actual sampled data Rx is input into the discriminator D to train the discriminator D and obtain the parameter set D. θThe random data Nx is input into the generator G, which outputs generator sample data Gx that follows the Pg distribution. The generator sample data Gx is then input into the discriminator D. It is determined whether the difference between the value P(D(x)) output by the discriminator D and the preset probability value is greater than the preset threshold. If so, the generator gradient obtained by the discriminator D is fed back to the generator G. If not, the discriminator continues to be trained. The detection and segmentation module is used to simultaneously detect targets and lane lines and segment drivable areas using an autonomous driving network model. The calculation module is used to calculate the vehicle navigation path based on the detection and segmentation results; The control module is used to control the vehicle's autonomous driving based on the vehicle's navigation path.

[0029] This invention also provides an in-vehicle processor, wherein the in-vehicle processor executes a program to implement the steps of the above-described IoT-based autonomous driving method.

[0030] This invention also provides a storage medium storing processor-executable instructions, which are loaded by one or more vehicle-mounted processors to execute the IoT-based autonomous driving method described above.

[0031] The specific working principle of the IoT-based autonomous driving device of the present invention is the same as or similar to that described in the specific embodiments of the IoT-based autonomous driving method above. For details, please refer to the relevant descriptions of the specific embodiments of the IoT-based autonomous driving method above.

[0032] The IoT-based autonomous driving method of the present invention can simultaneously perform target detection, lane line detection and drivable area segmentation through an improved autonomous driving network model, thereby planning the vehicle driving path based on the model output and controlling the vehicle to drive automatically.

[0033] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0034] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0035] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal to execute the methods described in the various embodiments of the present invention.

[0036] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An autonomous driving method based on the Internet of Things, characterized in that: Crawl the training dataset from the web; A deep learning network model is trained using a training dataset to construct an autonomous driving network model. The autonomous driving network model includes a feature extraction module, a spatial feature pooling module, and a detection module. YOLOv5's CSPDarknet is used as the feature extraction module. A 3-channel image is taken as input, and the feature map obtained after multiple convolutions is 1 / 32 the size of the original image. The YOLOv5 feature extraction module replaces the focus layer with an equivalent convolutional layer, replaces BottleneckCSP with module C3, and removes the convolutional layer after the concat layer. The feature extraction module encapsulates three modules: convolution, batch normalization, and activation functions. The activation functions are shown below: , where x represents the input of the function; The spatial feature pooling module includes a fast spatial convolutional pooling layer and a feature pyramid layer. The fast spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPoll layers, and an additive layer, used to extract feature maps of different sizes and fuse different features. The input to the feature pyramid layer is the output of the fast spatial convolutional pooling layer, which is 1 / 32 of the original size. After passing through multiple convolutional layers and upsampling layers with a scale parameter of 2, and then through a Concact layer to fuse features of different scales from the encoder, the feature pyramid is finally obtained and used for downstream detection and segmentation tasks. The detection module contains three detection heads, each composed of a generative adversarial network (GAN). The generation method of the GAN includes: creating a random function to generate data following a distribution P. Noisedata Random data Nx; establish sampling rules to obtain P that follows the overall distribution of the true data. Realdata The actual sampled data Rx is input into the discriminator D to train the discriminator D and obtain the parameter set D. θ The random data Nx is input into the generator G, which outputs generator sample data Gx that follows the Pg distribution. The generator sample data Gx is then input into the discriminator D. It is determined whether the difference between the value P(D(x)) output by the discriminator D and the preset probability value is greater than the preset threshold. If so, the generator gradient obtained by the discriminator D is fed back to the generator G. If not, the discriminator continues to be trained. The autonomous driving network model is used to simultaneously detect targets, lane lines, and segment drivable areas; Calculate the vehicle navigation path based on the detection and segmentation results; Automated driving of vehicles based on vehicle navigation routes.

2. The method according to claim 1, characterized in that, The spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPool layers, and an additive layer. The kernel_size of the MaxPool layer is uniformly set to 5, and the three MaxPool layers are connected in a cascade manner.

3. The method according to claim 2, characterized in that, When training an autonomous driving network model, a loss function is constructed, and the parameters of the autonomous driving network model are adjusted based on the loss function.

4. The method according to claim 3, characterized in that, The loss of the object detection branch is denoted as L, and the loss of the object detection branch includes the classification loss L. class Regression loss L box And confidence loss L object The weighted average of the three loss classes constitutes the loss for the object detection part, as shown below: L=aL class +bL box +cL object Where a, b, and c represent the classification loss L, respectively. class Regression loss L box And confidence loss L object The corresponding weights.

5. The method according to claim 4, characterized in that, The classification loss uses the binary cross-entropy loss, calculated as follows: ; , where x i y represents the predicted value for the current category, N represents the total number of categories, and y represents the predicted value for the current category. i y represents the probability of the current class after the activation function. i * This represents the true value of the current category.

6. The method according to claim 4, characterized in that, The vehicle navigation path is calculated using the shortest path search algorithm or the search algorithm based on heuristic functions.

7. An Internet of Things-based autonomous driving device, comprising: The acquisition module is used to crawl the training dataset from the network; The model building module is used to train a deep learning network model using the training dataset and build an autonomous driving network model. The autonomous driving network model includes a feature extraction module, a spatial feature pooling module, and a detection module. YOLOv5's CSPDarknet is used as the feature extraction module. A 3-channel image is taken as input, and the feature map obtained after multiple convolutions is 1 / 32 the size of the original image. The YOLOv5 feature extraction module replaces the focus layer with an equivalent convolutional layer, replaces BottleneckCSP with module C3, and removes the convolutional layer after the concat layer. The feature extraction module encapsulates three modules: convolution, batch normalization, and activation functions. The activation functions are shown below: , where x represents the input of the function; The spatial feature pooling module includes a fast spatial convolutional pooling layer and a feature pyramid layer. The fast spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPoll layers, and an additive layer, used to extract feature maps of different sizes and fuse different features. The input to the feature pyramid layer is the output of the fast spatial convolutional pooling layer, which is 1 / 32 of the original size. After passing through multiple convolutional layers and upsampling layers with a scale parameter of 2, and then through a Concact layer to fuse features of different scales from the encoder, the feature pyramid is finally obtained and used for downstream detection and segmentation tasks. The detection module contains three detection heads, each composed of a generative adversarial network (GAN). The generation method of the GAN includes: creating a random function to generate data following a distribution P. Noisedata Random data Nx; establish sampling rules to obtain P that follows the overall distribution of the true data. Realdata The actual sampled data Rx is input into the discriminator D to train the discriminator D and obtain the parameter set D. θ The random data Nx is input into the generator G, which outputs generator sample data Gx that follows the Pg distribution. The generator sample data Gx is then input into the discriminator D. It is determined whether the difference between the value P(D(x)) output by the discriminator D and the preset probability value is greater than the preset threshold. If so, the generator gradient obtained by the discriminator D is fed back to the generator G. If not, the discriminator continues to be trained. The detection and segmentation module is used to simultaneously detect targets and lane lines and segment drivable areas using an autonomous driving network model. The calculation module is used to calculate the vehicle navigation path based on the detection and segmentation results; The control module is used to control the vehicle's autonomous driving based on the vehicle's navigation path.

8. The apparatus according to claim 7, characterized in that, The spatial convolutional pooling layer consists of two standard convolutional layers, three MaxPool layers, and an additive layer. The kernel_size of the MaxPool layer is uniformly set to 5, and the three MaxPool layers are connected in a cascade manner.

9. The apparatus according to claim 8, characterized in that, When training an autonomous driving network model, a loss function is constructed, and the parameters of the autonomous driving network model are adjusted based on the loss function.

10. The apparatus according to claim 9, characterized in that, The loss of the object detection branch is denoted as L, and the loss of the object detection branch includes the classification loss L. class Regression loss L box And confidence loss L object The weighted average of the three loss classes constitutes the loss for the object detection part, as shown below: L=aL class +bL box +cL object Where a, b, and c represent the classification loss L, respectively. class Regression loss L box And confidence loss L object The corresponding weights.

11. The apparatus according to claim 10, characterized in that, The classification loss uses the binary cross-entropy loss, calculated as follows: ; , where x i y represents the predicted value for the current category, N represents the total number of categories, and y represents the predicted value for the current category. i y represents the probability of the current class after the activation function. i * This represents the true value of the current category.

12. The apparatus according to claim 11, characterized in that, The vehicle navigation path is calculated using the shortest path search algorithm or the search algorithm based on heuristic functions.

13. An in-vehicle processor, characterized in that, When the on-board processor executes the program, it implements the steps of the IoT-based autonomous driving method as described in any one of claims 1-6.

Images