A robot turning angle prediction model construction method and system

Abstract

The application provides a robot steering angle prediction model construction method and system, the method comprises the following steps: S1, obtaining the first perspective image and the corresponding actual steering angle of the robot in different scenes to obtain an original data set; S2, preprocessing the original data set and forming an expanded data set; S3, lane line semantic guide labeling is performed on the expanded data set to obtain a labeled data set; S4, taking ResNet18 as a baseline network, embedding a non-local attention module and a lightweight module to construct a robot steering angle prediction network; S5, training the robot steering angle prediction network by using the labeled data set to obtain a robot steering angle prediction model. The application can improve the prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of robot autonomous navigation technology, specifically relating to a method and system for constructing a robot steering angle prediction model. Background Technology

[0002] Autonomous navigation is one of the core technologies for achieving intelligence in wheeled robots. During autonomous navigation, wheeled robots rely on high-precision environmental perception and path planning capabilities for movement and obstacle avoidance. Among these, steering angle prediction is a crucial task for autonomous navigation, directly impacting the robot's path planning and obstacle avoidance abilities. Traditional methods, which utilize the expertise and experience of domain specialists to manually select or construct features, and then encode or fuse these features to build object detection models, perform poorly in complex environments and struggle to adapt to dynamically changing conditions, limiting the performance of wheeled robots in practical applications. In recent years, the rapidly developing end-to-end autonomous driving technology utilizes convolutional neural networks to directly learn the mapping relationship of control commands from input images, reducing reliance on manual features.

[0003] However, these early models typically rely solely on single-frame images for prediction, neglecting the temporal continuity of driving behavior and failing to effectively utilize continuous frame information to understand dynamic scene changes. This results in insufficient prediction accuracy and stability in complex, dynamic environments. Introducing more complex model structures to address these shortcomings would undoubtedly increase computational overhead and training difficulty, hindering real-time inference on resource-constrained embedded platforms. Therefore, the challenge in robot steering angle prediction lies in achieving a balance between model prediction accuracy and computational efficiency. There is an urgent need in this field to develop a robot steering angle prediction method that features high prediction accuracy, low computational latency, and lightweight design. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for constructing a robot steering angle prediction model, which can improve prediction accuracy.

[0005] This invention is implemented as follows: A method for constructing a robot steering angle prediction model, comprising the following steps: S1. Obtain first-view images of the robot in different scenarios and the corresponding actual turning angles to obtain the raw dataset; S2. Preprocess the original dataset to form an expanded dataset; S3. Perform lane line semantic guidance annotation on the expanded dataset to obtain an annotated dataset; S4. Using ResNet18 as the baseline network, embedding a non-local attention module and a lightweight module to construct a robot corner prediction network; S5. Train the robot corner prediction network using the labeled dataset to obtain the robot corner prediction model.

[0006] Furthermore, in S1, the first perspective image is a turning image with the runway and sky as the background, and is named with a turning angle label. The scene includes scenes under different weather conditions and motion states.

[0007] Furthermore, in S2, the preprocessing includes increasing or decreasing brightness, increasing or decreasing sharpness, increasing or decreasing chroma, horizontal flipping, and Y-axis translation.

[0008] Furthermore, step S3 includes the following steps: S301. Use an edge detection algorithm to extract the complete lane edges in each frame of the expanded dataset; S302. Use Hough transform to detect straight lines in the lane edge; S303. A region-based filtering method is used to retain only the parts related to lane lines.

[0009] Furthermore, step S301 includes the following steps: S30101. Perform Gaussian filtering on the expanded dataset; S30102. Calculate the gradient intensity and gradient direction of each image using the gradient operator; S30103. Perform non-maximum suppression processing on the gradient intensity of the image to obtain a single-pixel wide edge line; S30104. Based on the gradient intensity and the preset size, perform dual threshold detection and edge connection to extract the complete lane edge.

[0010] Furthermore, the robot corner prediction network includes an input layer, an initial standard convolutional layer, a max pooling layer, four core backbone feature layers, a non-local attention module, a lightweight convolutional module, a global average pooling layer, and a fully connected output layer; the four core backbone feature layers integrate two residual sub-modules.

[0011] Furthermore, in S4, the algorithmic expression for the non-local attention module is: ; Where, f(x) i ,x j ) is the similarity function, g(x) i Let be the feature map at position j, and let x be the feature map at position x. i The calculated feature map, C x x is the normalization factor. i This is the input signal.

[0012] Furthermore, step S5 includes the following steps: S501. Determine SmoothL1Loss as the loss function and Adam as the network parameter optimizer for parameter optimization. S502. Input the labeled dataset into the constructed corner prediction training network, perform forward propagation calculation of the network, and obtain the predicted value of the robot's turning angle. S503. Calculate the training loss based on the predicted value and the true value, and after completing the backpropagation to solve the gradient, update the network weights through the Adam optimizer. S504: The ReduceLROnPlateau strategy is used to dynamically adjust the learning rate and continuously monitor the validation set loss. If the loss does not improve significantly for two consecutive training cycles, the current learning rate is automatically reduced to 50% of its original value. S505: During training, continuously monitor the validation set loss. If it does not improve significantly within 5 complete training cycles, immediately trigger the early stop mechanism to stop training. S506. After training, save the weights of the best-performing model on the validation set to obtain the trained robot corner prediction model.

[0013] Furthermore, step S504 includes the following steps: Before training, pre-specify the core monitoring metrics of the strategy and set the metric improvement threshold, the number of consecutive rounds of performance stagnation, and the learning rate decay coefficient. After completing a single training cycle, calculate and obtain the latest value of the specified monitoring metric; Compare the latest value of the monitoring indicator with the historical best value to determine whether the change in the indicator is greater than the set improvement threshold, i.e. whether there is a significant improvement. If the indicator does not meet the improvement standard, record a performance stagnation. If the cumulative number of stagnations reaches the set number of consecutive rounds, adjust the current learning rate according to the decay coefficient. If the indicator has a significant improvement or the number of stagnations has not reached the threshold, keep the current learning rate unchanged and reset the cumulative number of performance stagnations. Repeat the above single-cycle indicator monitoring, judgment, and learning rate adjustment operations until training ends.

[0014] This invention also provides a robot steering angle prediction model construction system, comprising: The acquisition module is used to acquire first-view images of the robot in different scenarios and the corresponding actual turning angles to obtain the raw dataset. The preprocessing module is used to preprocess the original dataset and form an expanded dataset; The annotation module is used to perform lane line semantic guidance annotation on the expanded dataset to obtain an annotated dataset; The building blocks are used to construct a robot corner prediction network by embedding non-local attention modules and lightweight modules with ResNet18 as the baseline network. The training module is used to train the robot corner prediction network using the labeled dataset to obtain the robot corner prediction model.

[0015] The beneficial effects of this invention are: This invention acquires first-person view image data of robots under multiple scene conditions, enabling the prediction model to effectively learn complex road conditions such as lighting changes, road types, and the diversity of the surrounding environment. Furthermore, it employs preprocessing techniques such as image rotation, scaling, brightness changes, and noise addition to significantly increase the diversity of sample data and effectively improve the performance of the prediction model under uncertain conditions in the real world.

[0016] This invention significantly improves the ability to capture global contextual information and feature extraction efficiency of robot operation scenarios by optimizing the ResNeT baseline network. By introducing a model learning rate scheduler, the model learning rate is adaptively optimized, achieving a good balance between image feature extraction capability and model complexity. With fewer parameters and less computation, it is more suitable for deployment on robot embedded platforms with limited computing resources.

[0017] 3. This invention uses the Canny edge detection and Hough transform lane marking method to enhance the semantic guidance of the input image of the robot corner prediction network. Compared with the shortcomings of traditional convolutional networks that focus more on local domain information, this invention significantly enhances the key feature expression ability and generalization performance of the prediction network, which is more conducive to the robot making more forward-looking and global decision-making in driving behavior and significantly improves the prediction accuracy in complex road conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method provided by the present invention; Figure 2 This is a schematic diagram of the original dataset in this invention; Figure 3 This is a schematic diagram of the expanded dataset in this invention; Figure 4 This is a schematic diagram of the lane markings in this invention; Figure 5 This is a schematic diagram of the improved ResNet18 network structure that integrates Non Local Block and Ghost Module in this invention; Figure 6 This is a schematic diagram of the Non-local Block module structure in this invention; Figure 7This is a schematic diagram of the Ghost Module structure in this invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0020] like Figures 1 to 7 As shown, a method for constructing a robot steering angle prediction model includes the following steps: S1. The wheeled robot's onboard binocular vision sensor acquires first-person perspective images and corresponding actual turning angles under different scenarios such as sunny, cloudy, and rainy weather, obtaining the raw dataset. The wheeled robot's hardware architecture consists of a coaxial pendulum suspension chassis, a low-level main controller, an M10P LiDAR, and a Gemini Pro binocular depth camera. The binocular camera has a field of view of 71.5° horizontally and 56.7° vertically, a depth of field range of 0.25m to infinity, a maximum frame rate of 30fps, a resolution of 1920×1080, and a maximum robot speed of 0.63m / s. The wheeled robot is controlled via a PS2 controller, and Orbbec Gemini Pro binocular depth camera is used to acquire robot turning images against a runway and sky background. During the acquisition process, commands are sent via the ROS1 system terminal to control the robot to acquire and save images at a frequency of 12 frames per second. All images are labeled with their corresponding turning angle values, and a robot motion space scene image dataset is constructed that includes various weather conditions (sunny, cloudy, rainy) and motion states (turning left, turning right, going straight).

[0021] S2. To improve the model's generalization ability and accuracy, the original dataset is preprocessed to form an expanded dataset. For example... Figure 2 As shown, specifically, in this step, preprocessing includes increasing or decreasing brightness, increasing or decreasing sharpness, increasing or decreasing chroma, horizontal flipping, and Y-axis translation. Figure 3 As shown, after preprocessing, we obtain the augmented dataset after data augmentation.

[0022] S3. Perform lane line semantic guidance annotation on the expanded dataset to obtain an annotated dataset. Specifically, S3 includes the following steps: S301. Use the Canny edge detection algorithm to extract the complete lane edges from each frame of the augmented dataset. This step includes the following steps: S30101. Perform Gaussian filtering on the expanded dataset to reduce image noise.

[0023] S30102. Calculate the gradient intensity and gradient direction for each image using the gradient operator. The formulas for calculating the gradient intensity and gradient direction are: ; ; Where G is the gradient intensity, G x and G y θ represents the gradient magnitudes of the first-view image in the x and y directions, respectively, where θ is the gradient direction.

[0024] S30103. Perform non-maximum suppression processing on the gradient intensity of the image to obtain a single-pixel wide edge line. In this step, the Sobel operator is first used to calculate the gradient intensity and gradient direction of each pixel. The gradient direction points to the direction of the fastest increase in gray level. The normal direction of the edge is the gradient direction, and the tangent direction of the edge is perpendicular to it. For each pixel, check two opposite pixels (neighbors in the front and back directions) along its gradient direction. If the gradient magnitude of the current pixel is less than that of either neighbor, it is considered not a local maximum and is suppressed to 0; otherwise, it is a possible edge point. To achieve efficient comparison, the gradient direction is usually quantized into four directions (0°, 45°, 90°, 135°), corresponding to horizontal, upper right to lower left, vertical, and lower right to upper left, respectively, and then compared with neighboring pixels in these two directions.

[0025] S30104. Based on the gradient intensity and the preset size, perform dual-threshold detection and edge connection to extract the complete lane edge. In this step, if the gradient intensity of a pixel is greater than the high threshold, it is marked as a strong edge point; if the gradient intensity is less than the low threshold, its grayscale value is set to 0; for pixels with gradient intensities between the high and low thresholds, they are considered weak edge points. By searching for breakpoints of strong edge points, weak edge points are connected to extract the complete lane edge.

[0026] S302. The Hough transform is used to detect straight lines along the lane edge. The Hough transform is mainly used to detect specific geometric shapes in an image. Its core idea is to map edge points in the image space to the parameter space through a parameter space voting mechanism, and to locate the target shape by statistically analyzing the peak values ​​in the parameter space. A straight line in the image space can be represented in polar coordinates, as shown in the following formula: ; Where: r represents the distance from the line to the origin, α represents the angle between the line and the x-axis, and (x, y) represents a point on the line in image space. For each edge point (x, y) in the image, it corresponds to a curve in the parameter space (r, α), which is the combination of all parameter space (r, α) that satisfy the condition that the point lies on the line. If multiple edge points belong to the same line, their parameter curves will intersect at the same point (r0, α0) in the parameter space, and the number of intersections at this point will be significantly higher than at other locations. By statistically analyzing the peak values ​​in the parameter space, the line in the image can be located. Substituting the detected parameters (r, α) into the line equation: By converting the coordinates back to the image coordinate system and drawing straight lines, straight lines in the lane edge can be detected.

[0027] S303 employs a region-based filtering method to retain only the portions relevant to lane lines. For example... Figure 4 As shown, in this step, a region-based filtering method is used to retain only the parts related to lane lines. Specifically, the image region is divided into upper and lower parts based on its height, focusing only on the lower half. The width of the lower half is the same as the original image, and its vertical length is half that of the original image, extracted starting from the bottom of the image. Then, each line obtained from the Hough transform is checked, and its position is determined by calculating the coordinates of its endpoints. If at least one endpoint of a line is located within the lower half region, the line is retained; otherwise, it is discarded. This region-based filtering method effectively reduces lines generated by upper interfering objects, allowing the annotation process to focus more on the actual lane lines at the bottom of the image, thereby improving the accuracy and robustness of lane line detection.

[0028] S4. Using ResNet18 as the baseline network, a non-local attention module and a lightweight module are embedded to construct a robot corner prediction network. For example... Figure 5 As shown, the robot corner prediction network includes an input layer, an initial standard convolutional layer, a max pooling layer, four core backbone feature layers, a non-local attention module, a lightweight convolutional module, a global average pooling layer, and a fully connected output layer; the four core backbone feature layers integrate two residual sub-modules.

[0029] The robot turning angle prediction network's input layer is an augmented and semantically guided labeling dataset of the target lane. A 7×7 initial standard convolutional layer performs preliminary large-scale feature extraction, followed by a max-pooling layer to reduce the dimensionality of the feature map while preserving salient features, laying the foundation for subsequent deep feature extraction. Data flows into the core feature extraction backbone, consisting of four layers (Layer 1 to Layer 4), integrating two residual modules. A Non-local Block module is integrated after the backbone network. This module enables the network to capture long-range dependencies globally, enhancing the model's understanding of complex scene context. A Ghost Module replaces some standard convolutional layers, achieving feature representation capabilities similar to standard convolutions with fewer parameters and computational cost. The rich features extracted by the above deep layers are compressed into a one-dimensional feature vector by a global average pooling layer. This vector is then passed through a fully connected output layer, mapping the high-dimensional features to the final output value—the predicted robot turning angle.

[0030] like Figure 6 As shown, in step S4, the algorithm expression for the Non-Local Block attention module is: ; Where, f(P) i ,P j ) is the similarity function, g(P) i Let be the feature map at position i, and let P be the feature map at position i. i The calculated feature map, C P P is the normalization factor. i For the input signal, Q i The output signal is obtained by converting each characteristic position P in the input signal. i Other positions P j The result obtained by weighting and averaging the features is the input signal P. i The updated value obtained after operations by the NonLocal Block module.

[0031] like Figure 7 As shown, the Ghost Module lightweight module processing flow is as follows: Network complexity is calculated using FLOPs (floating point operations, used to measure the complexity of an algorithm or model):

[0032] For traditional convolutional layer structures: ; Regarding the intrinsic part of the Ghost Module: ; Regarding the ghost feature map portion of the Ghost Module: ; For the entire Ghost Module structure: ; Then we have: ; In the formula: C out It is the number of output channels, H out and W out These are the height and width of the output feature map, respectively, C in This is the number of input channels, n corresponds to C. out , is the number of output channels of a traditional convolutional layer, and h' and w' correspond to H respectively. out and W out , which are the height and width of the output feature map of a traditional convolutional layer, and c corresponds to C. in , where k×k is the number of input channels in a traditional convolutional layer, d1 is the number of input channels for the linear transformation, d2 is the number of output channels for the linear transformation, s is the number of Ghost feature maps generated (generally small), and c is the number of input channels (generally large). As the formula shows, using the Ghost Module reduces FLOPs to 1 / s compared to a traditional convolutional structure, indicating that the Ghost Module significantly reduces model complexity.

[0033] S5. Train the robot corner prediction network using the labeled dataset to obtain the robot corner prediction model. Further, S5 includes the following steps: S501. Determine SmoothL1Loss as the loss function and Adam as the network parameter optimizer for parameter optimization.

[0034] S502. Input the labeled dataset into the constructed corner prediction training network, perform forward propagation calculation of the network, and obtain the predicted value of the robot's turning angle.

[0035] S503. Calculate the training loss based on the predicted value and the true value, and after completing the backpropagation to solve the gradient, update the network weights through the Adam optimizer.

[0036] S504. The ReduceLROnPlateau strategy is used to dynamically adjust the learning rate, continuously monitoring the validation set loss. If the loss does not significantly improve after two consecutive training epochs, the current learning rate is automatically reduced to 50% of its original value. Specifically, this step includes the following: Before training, pre-specify the core monitoring metrics of the strategy (such as val_loss) and set the metric improvement threshold, the number of consecutive rounds of performance stagnation, and the learning rate decay coefficient.

[0037] After completing a single training cycle, calculate and obtain the latest value of the specified monitoring metric.

[0038] Compare the latest value of the monitoring indicator with the historical best value to determine whether the change in the indicator is greater than the set improvement threshold, i.e. whether there is a significant improvement. If the indicator does not meet the improvement standard, record a performance stagnation. If the cumulative number of stagnations reaches the set number of consecutive rounds, adjust the current learning rate according to the decay coefficient. If the indicator has a significant improvement or the number of stagnations has not reached the threshold, keep the current learning rate unchanged and reset the cumulative number of performance stagnations.

[0039] Repeat the above steps of monitoring, judging, and adjusting the learning rate for single-cycle metrics until training ends. The working principle of this step is as follows: Let the validation set loss in round t be λ. t The current learning rate is η t The adjustment strategy is as follows: ; Where γ∈(0,1) is the learning rate reduction factor, which is 0.5 in this example; p is the patience parameter, i.e., the number of consecutive epochs without improvement, which is 2 in this example; η t+1 This will be used as the learning rate for the next round.

[0040] S505: During training, continuously monitor the validation set loss. If it does not improve significantly within 5 complete training cycles, immediately trigger the early stop mechanism to stop training.

[0041] S506. After training, save the weights of the best-performing model on the validation set to obtain the trained robot corner prediction model.

[0042] The pseudocode for the model training method used in S5 is shown in Table 1: Table 1. Pseudocode of Model Training Algorithm 01. Set up the training device (GPU / CPU), initialize the AutoDriveNet model and TensorBoard logger. 02 Define data preprocessing transformations: resizing, tensor transformation, and normalization. 03 Load the training and validation datasets and create a data loader. 04 Define the SmoothL1 Loss function, Adam optimizer, and ReduceLROnPlateau learning rate scheduler. 05. Set the maximum number of training epochs T, the early stopping patience value, and initialize the best validation loss (best_val_loss). 06 While (current epoch < maximum number of epochs T) 07. Set the model to training mode and initialize the training loss and prediction result set. 08 for each training batch 09. Transfer image data and steering angle labels to the training device. 10. Forward Propagation: Obtaining the predicted output through the AutoDriveNet model. 11. Calculate the SmoothL1 Loss between the predicted output and the true label. 12 Backpropagation: Reset gradients, calculate gradients, and update model parameters 13. Accumulate the training loss and save the prediction results and the true labels. 14 end for 15. Calculate the average loss, mean absolute error, mean squared error, and root mean square error of the training set. 16. Set the model to evaluation mode and initialize the validation loss and prediction result set. 17 for each verification batch do 18. Move the validation image data and labels to the training device. 19. Forward propagation yields the predicted output (gradient not calculated). 20. Calculate the validation loss and save the validation prediction results. 21 end for 22. Calculate the validation set's average loss, mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE). 23. Record all metrics to TensorBoard: training / validation loss, MAE, MSE, RMSE, learning rate. 24. The learning rate update scheduler is based on validation loss. 25. If the current validation loss is less than the optimal validation loss, then... 26. Update the optimal validation loss, reset the early stopping counter, and save the optimal model parameters. 27 Output model saving prompt message 28 else 29. Early stop counter increments by 1, outputting no improvement message. 30 end if 31 if the early stop counter is greater than or equal to the patience value then 32. Trigger the early stop mechanism to exit the training loop. 33 end if 34-cycle counter epoch = epoch + 1 35 end while 36. Turn off the TensorBoard logger 37. Output training completion information and return the optimal model parameters. For the robot corner prediction model obtained after training with S5, the mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) are used as evaluation metrics to evaluate the performance of the model on the test set.

[0043] Specifically, it includes: MAE measures the mean absolute difference between the predicted steering angle and the actual steering angle, and is calculated using the following formula: ; MSE measures the squared mean of the differences between predicted and actual values, and is calculated using the following formula: ; RMSE is the square root of MSE, and the formula is as follows: ; In the formula, y i It is the actual steering angle. is the predicted steering angle, and n is the number of samples. The smaller these three values ​​are, the more significant the model accuracy; however, MSE is more sensitive to outliers than RMSE.

[0044] Based on the same inventive concept, the present invention also provides a robot steering angle prediction model construction system, comprising: The acquisition module is used to acquire first-view images of the robot in different scenarios and the corresponding actual turning angles to obtain the raw dataset. The preprocessing module is used to preprocess the original dataset and form an expanded dataset; The annotation module is used to perform lane line semantic guidance annotation on the expanded dataset to obtain an annotated dataset; The building blocks are used to construct a robot corner prediction network by embedding non-local attention modules and lightweight modules with ResNet18 as the baseline network. The training module is used to train the robot corner prediction network using the labeled dataset to obtain the robot corner prediction model.

[0045] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a robot steering angle prediction model, characterized in that, Includes the following steps: S1. Obtain first-view images of the robot in different scenarios and the corresponding actual turning angles to obtain the raw dataset; S2. Preprocess the original dataset to form an expanded dataset; S3. Perform lane line semantic guidance annotation on the expanded dataset to obtain an annotated dataset; S4. Using ResNet18 as the baseline network, embedding a non-local attention module and a lightweight module to construct a robot corner prediction network; S5. Train the robot corner prediction network using the labeled dataset to obtain the robot corner prediction model.

2. The method for constructing a robot steering angle prediction model according to claim 1, characterized in that, In S1, the first perspective image is a turning image with the runway and sky as the background, and is named with a turning angle label. The scene includes scenes under different weather conditions and motion states.

3. The method for constructing a robot steering angle prediction model according to claim 1, characterized in that, In S2, the preprocessing includes increasing or decreasing brightness, increasing or decreasing sharpness, increasing or decreasing chroma, horizontal flipping, and Y-axis translation.

4. The method for constructing a robot steering angle prediction model according to claim 1, characterized in that, S3 includes the following steps: S301. Use an edge detection algorithm to extract the complete lane edges in each frame of the expanded dataset; S302. Use Hough transform to detect straight lines in the lane edge; S303, a region-based filtering method is used to retain only the parts related to lane lines.

5. The method for constructing a robot steering angle prediction model according to claim 4, characterized in that, S301 includes the following steps: S30101. Perform Gaussian filtering on the expanded dataset; S30102. Calculate the gradient intensity and gradient direction of each image using the gradient operator; S30103. Perform non-maximum suppression processing on the gradient intensity of the image to obtain a single-pixel wide edge line; S30104. Based on the gradient intensity and the preset size, perform dual threshold detection and edge connection to extract the complete lane edge.

6. The method for constructing a robot steering angle prediction model according to claim 1, characterized in that, The robot corner prediction network includes an input layer, an initial standard convolutional layer, a max pooling layer, four core backbone feature layers, a non-local attention module, a lightweight convolutional module, a global average pooling layer, and a fully connected output layer; the four core backbone feature layers integrate two residual sub-modules.

7. The method for constructing a robot steering angle prediction model according to claim 1, characterized in that, In step S4, the algorithmic expression for the non-local attention module is: ； Where, f(x) i ,x j ) is the similarity function, g(x) i Let be the feature map at position j, and let x be the feature map at position x. i The calculated feature map, C x x is the normalization factor. i This is the input signal.

8. The method for constructing a robot steering angle prediction model according to claim 1, characterized in that, S5 includes the following steps: S501. Determine SmoothL1Loss as the loss function and Adam as the network parameter optimizer for parameter optimization. S502. Input the labeled dataset into the constructed corner prediction training network, perform forward propagation calculation of the network, and obtain the predicted value of the robot's turning angle. S503. Calculate the training loss based on the predicted value and the true value, and after completing the backpropagation to solve the gradient, update the network weights through the Adam optimizer. S504: The ReduceLROnPlateau strategy is used to dynamically adjust the learning rate and continuously monitor the validation set loss. If the loss does not improve significantly for two consecutive training cycles, the current learning rate is automatically reduced to 50% of its original value. S505: During training, continuously monitor the validation set loss. If it does not improve significantly within 5 complete training cycles, immediately trigger the early stop mechanism to stop training. S506. After training, save the weights of the best-performing model on the validation set to obtain the trained robot corner prediction model.

9. The method for constructing a robot steering angle prediction model according to claim 8, characterized in that, S504 includes the following steps: Before training, pre-specify the core monitoring metrics of the strategy and set the metric improvement threshold, the number of consecutive rounds of performance stagnation, and the learning rate decay coefficient. After completing a single training cycle, calculate and obtain the latest value of the specified monitoring metric; Compare the latest value of the monitoring indicator with the historical best value to determine whether the change in the indicator is greater than the set improvement threshold, i.e. whether there is a significant improvement. If the indicator does not meet the improvement standard, record a performance stagnation. If the cumulative number of stagnations reaches the set number of consecutive rounds, adjust the current learning rate according to the decay coefficient. If the indicator has a significant improvement or the number of stagnations has not reached the threshold, keep the current learning rate unchanged and reset the cumulative number of performance stagnations. Repeat the above single-cycle indicator monitoring, judgment, and learning rate adjustment operations until training ends.

10. A robot steering angle prediction model construction system, characterized in that, include: The acquisition module is used to acquire first-view images of the robot in different scenarios and the corresponding actual turning angles to obtain the raw dataset. The preprocessing module is used to preprocess the original dataset and form an expanded dataset; The annotation module is used to perform lane line semantic guidance annotation on the expanded dataset to obtain an annotated dataset; The building blocks are used to construct a robot corner prediction network by embedding non-local attention modules and lightweight modules with ResNet18 as the baseline network. The training module is used to train the robot corner prediction network using the labeled dataset to obtain the robot corner prediction model.

Images