Multi-task convolutional neural network architecture applied to advanced driving assistance system
Through the combination of batch processor, multi-task model and task allocator, the sharing characteristics and computing resource allocation problems of multi-task networks in advanced driving assistance systems are solved, and the parallel processing and performance improvement of multi-aware tasks are achieved.
Patent Information
- Application Number
- CN202311852390.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-29
- Publication Date
- 2025-07-01
AI Technical Summary
In the existing advanced driving assistance systems, multi-task convolutional neural networks are difficult to effectively share common features, task-specific heads fail to ensure task performance, computing resources are not fully shared, and multi-task training has problems such as inconsistent priority and unbalanced data distribution.
A batch processor is used to merge multiple batches of perceived task data, a multi-task model is used for feature extraction and shared feature fusion, and a task configuration and loss weight adjustment is performed through a task allocator. Combining a shared encoder, an attention mechanism decoder and a task-specific head, the task training process is dynamically adjusted.
The parallel processing of multiple perception tasks is realized, which improves computing efficiency and performance, ensures effective training and performance optimization of tasks, reduces task conflicts, and improves the computing efficiency and accuracy of the system.
Smart Images

Figure CN120235189A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to advanced driving assistance technology, and particularly to a multi-task convolutional neural network architecture applied to an advanced driving assistance system. Background Art
[0002] An advanced driving assistance system (ADAS) requires various perception tasks to implement a robust system to meet the requirements of customized service functions, such as the perception of front views, fish-eye views, and panoramic views. To improve accuracy, a multi-task convolutional neural network (abbreviated as CNN) is generally used to run these perception tasks.
[0003] Although the computing power in automotive embedded systems is constantly increasing, due to the continuous increase in the number of camera heads and perception tasks, more effective designs need to be proposed to address this. As a countermeasure, multi-task learning (MTL) has been proposed. However, there are still some difficulties in the design of multi-task network structures, such as: (1) it is difficult to combine object and pixel-level tasks for various ADAS applications, including moving object perception and surrounding environment perception; (2) the multi-task network cannot ensure that common features can be fully shared among multiple tasks, and the task-specific heads do not guarantee task performance; (3) shared computing or parameter percentages cannot be considered among tasks.
[0004] On the other hand, there are also some problems in the simultaneous training of these tasks, such as: different tasks may have different training priorities in applications; the data distributions among multi-tasks are unbalanced; in some cases, the optimization of one task will have a negative impact on the optimization of other tasks. Summary of the Invention
[0005] Based on the above problems in the prior art, the present invention aims to provide a multi-task convolutional neural network architecture applied to an advanced driving assistance system that can parallelly process multiple perception tasks and simultaneously maximize computational efficiency and performance.
[0006] The multi-task convolutional neural network architecture applied to an advanced driving assistance system according to one aspect of the present invention is characterized by including:
[0007] A batch processor for merging datasets of various perception tasks from an advanced driving assistance system and outputting a merged data batch;
[0008] A multi-task model for extracting shared features from the merged data batch through feature extraction and fusing the shared features based on an attention mechanism and then outputting multi-tasks; and
[0009] A task allocator for performing task configuration and loss weight adjustment on the multi-tasks. Brief Description of the Drawings
[0010] The above and other objects and advantages of the present application will become more fully apparent from the following detailed description taken in conjunction with the accompanying drawings, in which like or similar elements are denoted by like reference numerals.
[0011] Figure 1 It is a schematic diagram of a multi-task CNN (Convolutional Neural Network) system applied to an advanced driver assistance system according to an embodiment of the present invention.
[0012] Figure 2 It is a schematic diagram showing an example of a batch processor.
[0013] Figure 3 It is a schematic diagram showing an example of a partially shared decoder based on an attention mechanism. Detailed Embodiments
[0014] The following are some of the multiple embodiments of the present invention, aiming to provide a basic understanding of the present invention, and not aiming to identify the key or decisive elements of the present invention or limit the scope to be protected.
[0015] Before describing the multi-task convolutional neural network architecture of the present invention applied to an advanced driver assistance system, relevant technical terms will be explained first.
[0016] (1) CNN is short for Convolutional Neural Network, which is a neural network specifically used to process data with a similar grid structure.
[0017] (2) ODD classification is a technology in an Observational Driving Assistance System (ADAS) for identifying and classifying objects on the road. It is based on the vehicle's sensor data, such as radar, LiDAR (Light Detection and Ranging), and camera heads, to detect and classify various objects on the road, such as vehicles, pedestrians, bicycles, animals, etc.
[0018] (3) FPN (Feature Pyramid Network) is a structure for constructing a multi-scale feature pyramid.
[0019] (4) PANet (Path Aggregation Network) is a network that enhances feature fusion, especially in tasks such as object detection.
[0020] (5) Bi-FPN is a variant of FPN, which may provide bidirectional or multi-path feature fusion.
[0021] (6) The NAS model is a neural architecture search model, which refers to a method of finding the optimal neural network structure in a given search space (such as different numbers of layers and different structures in a convolutional neural network) through automated search to improve the model performance.
[0022] (7) ImageNet is a large visual database for visual object recognition software research.
[0023] (8) Softmax is a mathematical function, usually used to convert a set of arbitrary real numbers into real numbers representing a probability distribution. In essence, it is a normalization function.
[0024] (9) NAS is Neural Architecture Search, which is an optimization-based algorithm aiming to find the best structure of a neural network for a specific task on a specific dataset.
[0025] Figure 1 It is a schematic diagram of a multi-task CNN (Convolutional Neural Network) system applied to an advanced driver assistance system according to an embodiment of the present invention.
[0026] As Figure 1 shown, a multi-task CNN (Convolutional Neural Network) system applied to an advanced driver assistance system according to an embodiment of the present invention includes: a perception dataset 100, a batch processor 200, a multi-task model 300, and a task allocator 400.
[0027] First, the perception dataset 100 will be described.
[0028] The perception dataset 100 includes: a 2D object detection set 101, a monocular 3D object detection set 102, a keypoint detection set 104, an ODD (Object Detection) classification set 105, a lane detection set 106, a depth estimation set 107, and so on.
[0029] The perception dataset 100 provides an interface for users to obtain the required data. All these datasets can be independently labeled without annotating all labels on one image. The task-specific head 330 in the following will only be responsible for backpropagating the gradient for the input of a specific dataset. In machine learning, backpropagation is a method of training a neural network. In backpropagation, there is an error between the output of the network and the expected output, and this error will be propagated backward layer by layer from the output layer to the input layer in the form of a gradient, and each layer's weights will be updated according to this gradient. Therefore, a specific task or layer is only responsible for processing the backward gradient from the input of a specific dataset, that is, only responsible for updating the weights related to this specific dataset.
[0030] Next, the batch processor 200 will be described.
[0031] The batch processor 200 is used to combine data batches of various perception data sets obtained from the perception data set 100. In Figure 2 , DF represents the data batch and CB represents the combined batch. The batch processor 200 is responsible for combining data sets of different perception tasks and providing the combined batch for subsequent processing. Since the batch processor 200 allows batch processing of data, the processing efficiency can be improved.
[0032] The sizes of these data batches are not fixed, but are dynamically determined according to the requirements of specific tasks and the characteristics of the data sets. For example, the priority of the task and the classification of the data set may affect the size of the data batch, because more important tasks may require larger data batches for more accurate processing. The sampling factor of each data batch or task (sampling_factor i ) is determined by the following formula (1):
[0033] sampling_factor i = normalize(α i × w i ) (1)
[0034] where α i is an additional weighted task priority factor used to enhance important tasks. Usually, α i ≥ 1, w i is inversely proportional to the class frequency and can penalize the main class (which means assigning less weight to instances of the majority class). The normalize operator normalizes all factors so that their sum is 1.0.
[0035] where w i can be set in two forms, as shown in formula (2) or formula (3) below:
[0036]
[0037] or
[0038]
[0039] where di is the number of images in a certain data set, and is the proportion relative to all data.
[0040] In this way, the batch size (BATCH i ) of a specific task can be defined by the following formula (4):
[0041] BATCH i = int(BATCH × sampling_factor i ) (4)
[0042] Where BATCH is the predefined total batch size, and the int function is used to convert the result to an integer, then the total batch size (BATCH) will be updated to the following formula (5):
[0043]
[0044] In other words, in a fixed training batch, tasks with higher priority or fewer quantities may tend to have a higher sampling factor.
[0045] Figure 2 is a schematic diagram showing an example of a batch processor.
[0046] In Figure 2 , DS0 to DSn-1 represent dataset 0 to dataset n-1, SB represents the sampling batch, BD represents achieving balance, RD represents repeated sampling, ML represents the maximum length, B0 to Bn-1 represent batch 0 to batch n-1, and CB represents the combined batch.
[0047] Figure 2 shows how the batch processor 200 provides a combined batch. The batch processor 200 extracts samples from different datasets and creates balanced batches for training. Among them, each dataset may have different tasks, such as dataset 0, dataset 1 to dataset n-1. Moreover, multiple batches will be sampled from each dataset, and these batches may have different sizes. In the batch processor 200, the batches from different datasets will be adjusted to ensure data balance throughout the training process. For example, some datasets with fewer samples may be repeatedly sampled to reach the largest dataset (i.e., the maximum length, Figure 2 in which ML), thereby ensuring data balance. D0, D1,... Di,... Dn-1 represent the corresponding batches output after sampling, balancing, repeating, etc. These batches will ultimately be combined into a total batch (i.e., the combined batch represented by CB) and input into the subsequent model for training.
[0048] Through the processing and combination of the batch processor 200, it is possible to ensure data balance and full utilization during the multi-task learning process, which is crucial for efficient multi-task learning, can ensure the effective training of tasks, and will not dominate the training process due to the large amount of data of some tasks.
[0049] Next, the multi-task model 300 will be described.
[0050] As Figure 1 shown, the input of the multi-task model 300 is the combined batch from the batch processor 200 (in Figure 1(in the Chinese text, the merged batch is represented by CB), the output of the multi-task model 300 is the result of the multi-task and a list of task losses (in Figure 1 (in the Chinese text, the list of task losses is represented by TLL), the list of task losses is the complete set of individual task losses in the multi-task model, and the subsequent task allocator 400 will weight these individual task losses and output the sum of the multi-task losses (in Figure 1 (in the Chinese text, the sum of the losses is represented by LS), the sum of the losses is the total loss for neural network training, and the collected total loss will undergo gradient reversal based on the total loss.
[0051] Specifically, the multi-task model 300 mainly includes: a shared encoder 310, a partially shared decoder 320 based on the attention mechanism, and a task-specific head 330.
[0052] Among them, the shared encoder 310, as the backbone feature extractor, can provide different downsampling rate features from 1 / 2 to 1 / 128 ( Figure 1 represented by D1 / 2~D1 / 128 in the Chinese text), while the partially shared decoder 320 based on the attention mechanism can provide different downsampling fusion features from 1 / 4 to 1 / 128 ( Figure 1 represented by F1 / 2~F1 / 128 in the Chinese text).
[0053] The shared encoder 310, as the backbone feature extractor, aims to obtain shared features among multiple tasks. It refers to ResNet (Residual Network) and RegNet (Regularized Network). As Figure 1 shown, the shared encoder 310 can provide different downsampling rate features from 1 / 2 to 1 / 128 to ensure that different types of tasks can obtain fine high-level features. The downsampling rate refers to the method of sampling a lower-resolution feature map from a higher-resolution feature map. This technique can help reduce computational and storage requirements, and may also help improve the performance of the network. The downsampling rate varies from 1 / 2 to 1 / 128, which means that the size or number of channels of the feature map may be reduced from half of the original size to 1 / 128 of the original size.
[0054] The shared encoder 310 is usually pre-trained on a common classification dataset such as ImageNet. In the present invention, the shared encoder 310 is improved through semi-supervised learning of large-scale self-data.
[0055] The partially shared decoder 320 based on the attention mechanism is a segmented shared decoder. It is not only shared by multiple tasks or models, but its structure may also be divided into several parts. The specific structure of the partially shared decoder 320 based on the attention mechanism will be described later with reference to Figure 3 this.
[0056] The task-specific head 330 refers to the "heads" that are designed with specific parameters and calculations exclusive to each task. Among them, in deep learning and multi-task learning, the "heads" usually refer to the last few layers of the model, which are responsible for converting the features extracted by the shared encoder or backbone network into task-specific outputs. When a task is said to be "task-specific", it means that these last layers or "heads" are customized for that specific task and do not share parameters or calculations with other tasks. In Figure 1 , the task-specific head 330 correspondingly includes, for the perception dataset 100: a 2D object detection head 331, a monocular 3D object detection head 332, a keypoint detection head 334, an ODD (object detection) classification head 335, a lane detection head 336, a depth estimation head 337, and so on.
[0057] This design of the task-specific head 330 allows the model to maintain flexibility when processing multiple tasks because each task can have its own output layer and specific calculation method, while the shared encoder 310 is responsible for extracting general features useful for all tasks. Thus, this structure can help the model achieve better performance on multiple tasks while reducing the overall number of parameters and computational complexity.
[0058] Figure 3 is a schematic diagram showing an example of a partially shared decoder based on the attention mechanism.
[0059] The partially shared decoder 320 based on the attention mechanism is a split shared decoder that is not only shared by multiple tasks or models, but its structure may be divided into several parts, providing different downsampling rate fusion features from 1 / 4 to 1 / 128 (as shown by F1 / 4 to F1 / 128 in Figure 3 ), that is, this decoder can generate or process features of multiple scales, and the resolution of these features varies from 1 / 4 of the original size to 1 / 128.
[0060] In Figure 3 's example, the partially shared decoder 320 based on the attention mechanism includes: a NAS model 321 for automatically selecting and optimizing parameters in the network structure and an attention model 322 that uses the attention mechanism to further enhance the selection and fusion of features.
[0061] Here, the partially shared decoder 320 based on the attention mechanism provides multi-scale feature fusion, aiming to aggregate features of different resolutions. Through multi-scale feature fusion, features of different resolutions can be fused together so that the model can capture information from multiple scales. Among them, structures such as FPN, PANet, or Bi-FPN are usually selected for the design of the "neck" or decoder. In deep learning, the "neck" refers to the part connecting the encoder and the decoder, and the decoder is responsible for converting the encoded features into the final output. Here, FPN, PANet, and Bi-FPN can be used as part of the "neck" or decoder, indicating that these structures can help fuse and transform features more effectively.
[0062] Here, as an example, according to the design pattern of Bi-FPN, a cross-stage aggregated decoder is designed in the present invention, that is, Figure 3 the NAS model 321 in [reference] has M Bi-FPN basic modules repeated N times (where M and N are natural numbers), which can achieve more advanced feature fusion. Compared with the prior art, by expanding the basic blocks of Bi-FPN, downsampling rate layers of 1 / 4 and 1 / 2 are added, thereby enabling adaptation to pixel-level tasks (such as segmentation).
[0063] Furthermore, in the attention model 322, additional weights are used to adjust the importance of each input feature (the attention model 322 adjusts the feature importance of different levels through the attention weight parameters α j and β j ) and fuse these features with features from different layers and different resolutions. The output of the fused features can be defined as follows in Equation (6):
[0064] 0 = ∑ j f(β j )(2 i f(α j ) × I ij ) (6)
[0065] where I ij is the pre-output of the high-level features, α j is the weighted cross-layer fusion factor, β j is the weighted factor of different resolutions, α j and β j are learnable attention weights, and f(·) is a fusion function based on softmax, which is defined as follows in Equation (7):
[0066]
[0067] where e represents the base of the natural logarithm, and w i is a factor of the dynamic weight.
[0068] To optimize accuracy and efficiency, the neural architecture search (NAS) technique is adopted to search for reasonable parameter combinations through the number of repetitions of M and N, and the basic blocks of Bi-FPN are improved to optimize cross-scale connections to find the best neural network architecture.
[0069] As described above, through Figure 3 the partially shared decoder based on the attention mechanism shown, by combining the automated structure selection of the NAS model and the directional feature enhancement of the attention model, can improve the efficiency and accuracy of multi-task learning. With this design, the model can more precisely process and fuse features from different tasks, thus providing more accurate and reliable performance in complex systems such as ADAS.
[0070] Finally, the task allocator 400 is described.
[0071] The task allocator 400 includes a task configuration module 410 for implementing task training configuration and a dynamic loss weight adjustment module 420 for implementing task training loss weight adjustment. Among them, the dynamic loss weight adjustment module 420 balances multi-task training by using dynamic loss weights. This is because, in multi-task learning, different tasks may have different importance or difficulty. To ensure that each task is treated fairly and to ensure that the model can learn all tasks in a balanced manner, the task allocator 400 dynamically adjusts the loss weights of each task according to the characteristics or importance of the task.
[0072] Here, the loss function of the task configuration module 410 is defined by the following equations (8) and (9):
[0073] loss = Σ i gate i × w i × loss(i) (8)
[0074] subject to: Σ i gate i × w i = 1 (9)
[0075] where loss represents the loss function, gate i represents the effective training period when the task is activated. When the task is activated, gate i = 1, otherwise gate i = 0, w i is a factor of the dynamic weight, satisfying equation (9), and "subject to" means "satisfying".
[0076] The task configuration module 410 fully considers the training requirements in the task allocator and defines tasks that can be trained separately in different training stages. As shown in Equation (10) below, through parameterized gates i to implement this function:
[0077]
[0078] where, epoch represents the cycle, and start_point represents the starting point. For example, it can be set that the object detection task will start training from the 10th cycle (epoch), and the segmentation task will start training from the beginning.
[0079] For a specific task, fine-tuning is required and it starts from the shared frozen features. In this way, the parameters of the shared features are frozen, and only the valid task gates i = 1. Thus, it can adapt to the requirements of specific tasks.
[0080] Next, the dynamic loss weight adjustment module 420 is described.
[0081] The dynamic loss weight adjustment module 420 considers the dynamic weights between tasks, including the magnitude of the loss, the learning speed and difficulty, the imbalanced data distribution, the importance and priority of tasks, and the uncertainty estimation, especially to alleviate task conflicts during the training process. Task conflicts occur because there is competition or mutual interference between different tasks, resulting in the model being difficult to converge or having poor performance during training. In the present invention, by considering these factors, the relationship between different tasks can be better balanced, thereby improving the training effect of the model.
[0082] Here, we define w i as the factor of the dynamic weight and make the following definitions in Equations (11) and (12):
[0083]
[0084]
[0085] where, γ i (t - 1) represents the learning speed, L k (t) represents the loss of the task at time t, then L k (t - 1), L k (t - 2) represent the loss of the previous frame and the frame before the previous frame at time t respectively, which are used here to evaluate the change trend of the loss.
[0086] where, α i is the additional weight that balances the importance of the task and the magnitude of the loss, and β iis the coefficient of multitask balance, which is set by the dynamic loss weight adjustment module 420. The temperature T is usually used to adjust the smoothness of task weights. When the temperature T is large, the weight distribution between tasks is more uniform, that is, different tasks are treated more equally without favoring one or several tasks.
[0087] To alleviate task conflicts during training to some extent, the gradient g of the task on the shared parameters will be considered k , as shown in the following equations (13) and (14):
[0088]
[0089]
[0090] Here, θ represents the training parameters of the neural network, and g i represents the gradient of task i. Set the important task as k. If the conflict i is detected as true, it is determined that the i-th task conflicts with the important task k. In this way, a lower β i factor can be used to alleviate task conflicts.
[0091] In this way, through task configuration and dynamic loss weight adjustment, the task allocator 400 can optimize the multitask training effect, alleviate conflicts between tasks, and ensure that the model has good performance on all tasks.
[0092] As described above, the multitask convolutional neural network architecture applied to the advanced driver assistance system according to the present invention can allow the advanced driver assistance system to process multiple perception tasks in parallel, and at the same time can maximize the computational efficiency and performance. Specifically, the following technical effects can be obtained:
[0093] Propose a multitask CNN architecture for visual perception in ADAS applications, including: object-level tasks, such as 2D object detection, monocular 3D object detection, key point detection; and pixel-level tasks, such as segmentation, lane detection, depth estimation;
[0094] Create a batch processor 200 according to the priority and dataset distribution of tasks, and determine the priority and sampling rate of each task by sampling factors in a fixed training batch;
[0095] A new multitask model 300 can be proposed, which combines a shared encoder 310, a partially shared decoder 320 based on the attention mechanism in the attention part, and a task-specific head 330;
[0096] The task configuration module 410 can flexibly allocate tasks. For example, it can control the independent training of tasks in different training stages, or use shared frozen features to fine-tune each specific task;
[0097] The dynamic loss weight adjustment module 420 can balance multi-task training, especially can reduce task conflicts during training to a certain extent.
[0098] The above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Those skilled in the art can think of other feasible changes or substitutions according to the technical scope disclosed in the present application, and such changes or substitutions are all covered by the protection scope of the present application. Without conflict, the embodiments of the present application and the features in the embodiments can also be combined with each other. The protection scope of the present application is subject to the claims.
Claims
1. A multi-task convolutional neural network architecture applied to an advanced driver assistance system, characterized in that, Including: A batch processor for merging datasets of multiple perception tasks from the advanced driver assistance system and outputting a merged data batch; A multi-task model for extracting shared features from the merged data batch, fusing the shared features based on an attention mechanism, and then outputting multiple tasks; And A task allocator for task configuration and loss weight adjustment of the multiple tasks.
2. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 1, wherein The batch processor separately samples and balances the datasets of the multiple perception tasks, and then outputs batches corresponding to the datasets of the multiple perception tasks. After merging the corresponding batches, a total data batch is output.
3. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 2, wherein The batch processor separately samples the datasets of the multiple perception tasks using a sampling factor, where the sampling factor is determined based on the priority of the task and the classification of the dataset.
4. The multi-task convolutional neural network architecture applied to the advanced driver assistance system according to claim 1, characterized in that, The multi-task model includes: A shared encoder for extracting multi-resolution shared features of multiple perception tasks from the merged data batch; A partially shared decoder based on an attention mechanism for fusing shared features of different resolutions based on the attention mechanism; and Task-specific heads, each task of the multiple perception tasks having specific parameters and calculation methods.
5. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 4, wherein The shared encoder provides different downsampling rate features from 1 / 2 to 1 / 128, and the partially shared decoder based on an attention mechanism provides different downsampling fusion features from 1 / 4 to 1 / 128.
6. The multi-task convolutional neural network architecture applied to the advanced driver assistance system according to claim 5, wherein The partially shared decoder based on an attention mechanism includes: A NAS model for automatically selecting and optimizing parameters in the network structure; and An attention model for further enhancing feature selection and fusion using the attention mechanism.
7. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 6, wherein The NAS model has M basic modules repeated N times, where M and N are natural numbers, and the basic module is any one of FPN, PANet, or Bi-FPN.
8. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 7, wherein The attention model uses additional weights to adjust the importance of each input feature and fuses these features with features from different layers and different resolutions. The output of the fused features is defined by the following formula (6): 0 = ∑ j f(β j )(∑ i f(α j ) × I ij ) (6) Among them, I ij is the pre-output of the high-level features, α j is the weighted cross-layer fusion factor, β j is the weighted factor of different resolutions, α j and β j are learnable attention weights, and f(·) is a softmax-based fusion function, which is defined as follows in Equation (7): where e represents the base of the natural logarithm, and w i is a factor of the dynamic weight.
9. The multi-task convolutional neural network architecture applied to the advanced driver assistance system according to claim 8, characterized in that, The task allocator includes: A task configuration module for implementing task training configuration; and A dynamic loss weight adjustment module for implementing task training loss weight adjustment.
10. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 9, wherein The loss function of the task configuration module is defined by the following formulas (8) and (9): loss = ∑ i gate i × w i × loss(i) (8) subject to: ∑ i gate i × w i = 1 (9) where loss represents the loss function, and subjectto means compliance with Among them, gate i represents the effective training cycle when the task is activated. When the task is activated, gate i = 1, otherwise gate i = 0, w i is the factor of the dynamic weight and conforms to Equation (9). Moreover, in the task configuration module 410, the tasks separately trained in different training phases are defined as in the following formula (10) where epoch represents the cycle and start_point represents the starting point 11. The multi-task convolutional neural network architecture applied to an advanced driver assistance system according to claim 10, characterized in that the following formulas (11) and (12) are defined in the dynamic loss weight adjustment module Among them, γ i (t - 1) represents the learning rate, and L k (t) represents the loss of the task at time t. Then L k (t - 1) 、 L k (t - 2) represent the losses of the previous frame and the frame before the previous frame at time t respectively. where α i is an additional weight for balancing task importance and loss magnitude, β i is a coefficient for multi-task balance, T is used to adjust the smoothness of task weights, and, to alleviate task conflicts during training, the gradients g of tasks on shared parameters are defined as in the following equations (13) and (14) k : Among them, θ represents the training parameters of the neural network, k represents the important task, and g i represents the gradient of task i, conflict represents conflict, true represents true, and flase represents false.
12. A computer-readable medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the multi-task convolutional neural network architecture applied to an advanced driver assistance system according to any one of claims 1 to 11 13. A computer device, comprising a storage module, a processor, and a computer program stored on the storage module and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-task convolutional neural network architecture applied to an advanced driver assistance system according to any one of claims 1 to 11
