Intelligent airport unmanned vehicle optimization control method based on rain and fog image enhancement

Abstract

The invention discloses an intelligent airport unmanned vehicle optimization control method based on rain and fog image enhancement, and the method comprises the steps: obtaining scene environment data, constructing an image rain and fog removing model based on a Gaussian encoder, building a vehicle dynamics model, and carrying out the optimization control of an unmanned vehicle. And performing collision avoidance path planning on the unmanned vehicle according to a fusion algorithm of the dynamic window and the bidirectional rapid expansion random tree. Through obstacle classification and trajectory prediction, an improved bidirectional fast expansion random tree, a scene adaptive weighted evaluation function, and a multi-vehicle cooperative collision avoidance and fault tolerance mechanism, efficient dynamic collision avoidance path planning of unmanned vehicles in a gallery bridge scene is realized. Meanwhile, an optimized attention generative adversarial network is designed, the dependence of a background and rain lines is reduced by adding a Gaussian model, a rain removal module and a defogging module are fused, and enhancement processing is performed on an airport gallery bridge scene image in rainy and foggy weather.

Description

Technical Field

[0001] This application relates to the field of unmanned driving and image processing technology for smart airports, specifically to a method for optimizing the control of unmanned vehicles and enhancing rain and fog images for smart airports. Background Technology

[0002] In recent years, smart airports have developed rapidly, achieving remarkable results in technological innovation, policy support, market expansion, application scenario development, green transformation, and international cooperation, becoming a core engine for promoting the high-quality development of the aviation industry.

[0003] The construction of smart airports is a core direction for the intelligent upgrading of the air transport industry. The application of unmanned vehicles in the boarding bridge area can significantly improve ground handling efficiency and reduce labor costs. However, the boarding bridge scenario has significant technical pain points: First, the space is narrow and the pillars are dense, with complex dynamic obstacles (ground staff, temporary equipment, and multiple vehicles working simultaneously). Existing path planning algorithms suffer from poor real-time collision avoidance, insufficient dynamic adaptation, and inefficient handling of multi-vehicle conflicts. Second, rain and fog can cause image blurring and loss of detail, affecting the environmental perception accuracy of unmanned vehicles. Existing rain and fog removal algorithms are prone to missing background information or leaving rain streaks. These problems seriously restrict the unmanned progress of smart airports and urgently require a targeted optimization technology solution.

[0004] Therefore, it is necessary to provide an optimized control and rain / fog image enhancement method for autonomous vehicles in smart airports to address the problems of poor collision avoidance performance and poor image quality in rain and fog scenarios. Summary of the Invention

[0005] The intelligent airport unmanned vehicle optimization control method based on rain and fog image enhancement provided in this application can effectively improve...

[0006] The collision avoidance success rate of unmanned vehicles and the image quality in rain and fog provide reliable technical support for the unmanned operation of smart airports, demonstrating significant practicality and promotional value.

[0007] Firstly, an optimized control method for autonomous vehicles is provided, including:

[0008] By fusing LiDAR and visual sensors to acquire scene environment data, differentiated safety distance threshold settings are implemented.

[0009] A Gaussian encoder-based image deraining network model is constructed. This includes collecting image data of airport boarding bridge scenes under rainy and foggy weather, including a training set for training and a test set for testing, with no overlap between the training and test sets. A Gaussian module-based image deraining model is constructed, using an optimized attention-based generative adversarial network (IA-GAN) with a Gaussian module added to the generator input. The Gaussian module-based image deraining model is trained using the training and validation sets to obtain derained images. The output is fused with a defogging module to enhance the airport boarding bridge scene images under rainy and foggy weather, restoring image details, textures, and scene information. Finally, a dark channel prior image defogging model based on an atmospheric scattering model is constructed to obtain derained and foggy images.

[0010] Establish a vehicle dynamics model; including, establishing a kinematic model of the autonomous vehicle based on its motion to satisfy kinematic constraints during the process:

[0011]

[0012] The collision avoidance path planning for autonomous vehicles is performed based on a fusion algorithm of dynamic window and Bid-RRT, including scene adaptive weighted evaluation function, multi-vehicle cooperative collision avoidance and fault tolerance mechanism, improved bidirectional fast expanding random tree path search, and generation of dynamic collision avoidance path.

[0013] The feasibility of the proposed algorithm was verified through simulation experiments.

[0014] Secondly, an optimized control method for autonomous vehicles is provided, including:

[0015] The environmental data acquisition module is used to acquire surrounding environmental data and set differentiated safety distance thresholds;

[0016] The model building module constructs an image de-raining and de-fogging model based on a Gaussian encoder.

[0017] The model training module is used to perform unsupervised training on the anomaly detection model.

[0018] The path planning module is used for path planning of autonomous vehicles.

[0019] The simulation verification module is used to perform simulation verification on the obtained model.

[0020] In the first or second aspect mentioned above, acquiring environmental data includes obtaining environmental data of the covered bridge scene by fusing LiDAR and visual sensors to acquire a dataset containing rain and fog images and clear images corresponding one-to-one with the rain and fog images, and dividing it into training set and test set.

[0021] In the first aspect mentioned above, constructing an image deraining model based on Gaussian modules includes:

[0022] An optimized Attention-Based Generative Adversarial Network (IA-GAN) is employed, adding a Gaussian module to the generator input and fusing a dehazing module at the output to enhance airport boarding bridge scene images in rainy or foggy weather, restoring image details, textures, and scene information. The output dehazing module further dehazes the de-rained image using a dark channel prior-based method.

[0023] In the first or second aspect mentioned above, the acquisition of environmental data includes the classification and trajectory prediction of obstacles in the acquired scene environment data.

[0024] In the first or second aspect mentioned above, the obstacle classification is defined as classifying obstacles into static fixed type, dynamic low-speed type, and sudden dynamic type.

[0025] In the first or second aspect mentioned above, the trajectory prediction is to predict the motion trajectory of dynamic obstacles using Kalman filtering, and to set a corresponding differentiated safety distance threshold.

[0026] In the first aspect mentioned above, constructing an image deraining model based on Gaussian modules includes:

[0027] An optimized Attention-Based Generative Adversarial Network (IA-GAN) is employed, adding a Gaussian module to the generator input and fusing a dehazing module at the output to enhance airport boarding bridge scene images in rainy or foggy weather, restoring image details, textures, and scene information. The output dehazing module further dehazes the de-rained image using a dark channel prior-based method.

[0028] Compared with the prior art, the advantages of the present invention are:

[0029] 1. By using obstacle classification and trajectory prediction, improved bidirectional fast expanding random tree (Bid-RRT), scene adaptive weighted evaluation function, multi-vehicle cooperative collision avoidance and fault tolerance mechanism, efficient dynamic collision avoidance path planning for unmanned vehicles in the corridor bridge scenario is achieved.

[0030] 2. Design an optimized attention-based generative adversarial network (IA-GAN), which reduces the dependence of background and rain lines by adding a Gaussian model, and integrates de-raining and de-fogging modules to enhance airport boarding bridge scene images under rainy and foggy weather.

[0031] 3. This invention can effectively improve the collision avoidance success rate of unmanned vehicles and the image quality in rain and fog, providing reliable technical support for the unmanned operation of smart airports. Its practicality and promotion value are significant. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0033] Figure 1 A schematic flowchart of an airport unmanned vehicle optimization control method based on rain and fog enhancement according to an embodiment of this application is shown.

[0034] Figure 2 The diagram shows the network architecture for image deraining based on Gaussian modules. Detailed Implementation

[0035] This application provides an optimized control method for autonomous vehicles based on rain and fog image enhancement. It utilizes a generative adversarial network (GAN) to sample from the latent variable distribution, then reconstructs the output using a decoder. The method optimizes the autonomous vehicle's collision avoidance path planning algorithm from multiple dimensions, introducing obstacle classification, heuristic search, adaptive evaluation, cooperative collision avoidance, and fault tolerance mechanisms. This addresses issues such as slow path search in narrow spaces, poor dynamic obstacle adaptation, multi-vehicle collisions, and the imbalance between safety and efficiency in smart airport boarding bridge scenarios. Simultaneously, it optimizes the attention-based GAN to achieve efficient rain and fog removal from rain and fog images, improving the reliability of the autonomous vehicle's environmental perception. The optimized scheme demonstrates significant image enhancement effects, adapts to the complex operating environment of smart airports, and exhibits outstanding technological advancement and practicality. It can promote the unmanned transformation of smart airports and has broad application prospects.

[0036] The following will combine Figure 1 and Figure 2 This paper provides a detailed explanation of the optimized control method for autonomous vehicles based on rain and fog image enhancement.

[0037] Figure 1 A schematic flowchart illustrating an embodiment of the autonomous vehicle optimization control method based on rain and fog image enhancement according to this application is shown. The method specifically includes:

[0038] Step S1: Obtain environmental data and set differentiated safety distance thresholds;

[0039] The system collects environmental data of the covered bridge scene by fusing LiDAR and visual sensors; the acquired environmental data includes obstacle classification and trajectory prediction.

[0040] The obstacles are classified into three categories: static fixed obstacles, dynamic low-speed obstacles, and sudden dynamic obstacles.

[0041] The trajectory prediction involves using Kalman filtering to predict the motion trajectory of dynamic obstacles, and setting corresponding differentiated safety distance thresholds. The differentiated safety distance thresholds are set as follows: 2.5m for personnel, 1.5m for equipment, and 3m for sudden dynamic obstacles.

[0042] Step S2: Construct an image de-raining and de-fogging model based on a Gaussian encoder.

[0043] Data preparation: 2000 images of the airport boarding bridge scene under rainy and foggy weather were collected. The image resolution was uniformly 1920×1080. 1500 images were used for training and 500 images were used for testing. The training set and the test set did not overlap.

[0044] A Gaussian module-based image deraining model is constructed. This model comprises a generator network and a discriminator network. The generator network aims to minimize the difference between the original image data and the generated image data, producing images that are as realistic as possible and unaffected by raindrops. The discriminator network analyzes whether the input image is real or generated and provides feedback to the generator network based on the output. The Gaussian module-based image deraining model is trained using training and validation sets to obtain derained images. A dark channel prior image dehazing model based on an atmospheric scattering model is then constructed to obtain derained fogged images.

[0045] Optimized training of Attention-Based Generative Adversarial Networks (IA-GAN):

[0046] Generator structure: A Gaussian module is added to the input end, and a 5×5 Gaussian filter is used to preprocess the input image with a standard deviation of 1.2 to reduce the adhesion between rain lines and the background; the intermediate layer adopts a channel attention mechanism to focus on the rain line and fog areas; the output end integrates a dehazing module, which uses a dark channel prior method to further dehaze the image after rain removal.

[0047] The generator takes a rainy image as input, processes it through hidden layers to obtain a de-rained image, and outputs it. The generator network consists of two sub-networks: an attention recurrent network and a context autoencoder. The attention recurrent network aims to find attention-grabbing regions in the input image to produce better local image reconstruction and to focus the discriminative network on evaluation.

[0048] The recursive network consists of a residual network, LSTM units, and convolutional layers. The structural parameters are set as follows: Figure 2 As shown. Each residual module in the ResNet network consists of 2 residual units (6 convolutional layers), using ReLU as the activation function. An LSTM unit includes an input gate. A forgotten door An input gate A unit state This solves the long-term dependency problem inherent in recurrent neural networks. The interaction between states and gates over time is defined as follows:

[0049]

[0050]

[0051]

[0052]

[0053]

[0054] In the formula: These are characteristics of residual network generation. Encode the state to be passed to the next LSTM. This represents the output characteristics of the LSTM unit. The output features of the LSTM are then fed into a convolutional layer, producing a 2D attention map. At each time step, the current attention map is concatenated with the input and fed into the next module of the recurrent network.

[0055] During network training, rain images and clean images with identical background scenes are used. Under the output mapping at a specific time step t, the loss function in each recurrent block is defined as:

[0056]

[0057] The multi-scale loss function is:

[0058]

[0059] The perceptual loss function is:

[0060]

[0061] Therefore, the loss of the generator network is:

[0062]

[0063] In the formula: For counter-loss, For cyclic network loss, For perceived loss, O is the output image of the autoencoder.

[0064] The discriminator is a neural network designed to distinguish between real and fake images. It takes the attention map output from the generator and feeds it into nine convolutional layers with 5×5 kernels, using the ReLU activation function. The final layer of the discriminator applies the sigmoid function to the feature map to facilitate the determination of authenticity.

[0065] A loss function is defined based on the output and attention map of the CNN in the autoencoder:

[0066]

[0067] In the formula: The mapping is the loss between the features extracted from the discriminator's inner layers and the final mapping:

[0068]

[0069] Training process: The batch size was set to 16, the number of iterations was 50,000, the Adam optimizer was used, the learning rate was initialized to 0.0002, and the learning rate decayed by 50% every 10,000 iterations.

[0070] Image enhancement test: 500 test set images were input into the trained model. The results showed that the enhanced images achieved a rain line removal rate of over 95%, the fog dissipation improved image contrast by over 30%, the average peak signal-to-noise ratio (PSNR) reached 38.5dB, and the average structural similarity (SSIM) reached 0.92. All of these results are superior to existing traditional algorithms and basic GAN algorithms, effectively preserving key scene information such as bridge pillars and ground markings.

[0071] Step S3: Establish the vehicle dynamics model

[0072] A kinematic model of the autonomous vehicle is established based on its motion to satisfy the kinematic constraints during the process:

[0073]

[0074] Where w is the vehicle's angular velocity in the heading direction. For the vehicle's heading angle, , These are the coordinates of the rear axle center of the vehicle. The velocity is the center velocity of the following week.

[0075] Step S4: Plan collision avoidance paths for the autonomous vehicle based on the fusion algorithm of dynamic window and Bid-RRT.

[0076] Improved Bid-RRT path search: Introduce a corridor bridge heuristic guiding factor to restrict random tree sampling along the channel length direction (x-axis range ±5m). In areas with dense obstacles and pillars, set the node connectivity threshold to 1.2m. The original path after connectivity is pruned by Bézier curves to remove 6 redundant inflection points. The path curvature is ≤0.3rad / m, which meets the vehicle steering constraints.

[0077] Adaptive weighted evaluation function selection: The current operation is in its peak period and there are many obstacles. We set the weights W1=0.5, W2=0.3, and W3=0.2. Combined with the constraint of a speed limit of 3m / s at the entrance of the corridor bridge, we selected the optimal solution (linear velocity 2.8m / s, angular velocity 0.2rad / s) from 20 speed combinations.

[0078] Multi-vehicle cooperative collision avoidance: The vehicle-to-vehicle communication module detects a collision between the target path and a vehicle with lower priority (overlap time 1.8s). The vehicle maintains its original path, and the other vehicle avoids the collision by decelerating to 1.5m / s. The reserved emergency lane (1.5m wide) is in standby mode.

[0079] Fault tolerance test: Simulating a LiDAR malfunction, the vehicle automatically switches to vision + inertial navigation fusion positioning, and the safe distance threshold is expanded to 1.8m; when a new obstacle suddenly enters the safe area (position change of 1.2m), it triggers rapid replanning, and the path adjustment is completed in 0.25s.

[0080] Tests showed that the optimized path planning method achieved a collision avoidance success rate of 99.2% in this scenario, with a path planning time of ≤0.4s, a replanning response time of ≤0.3s, and a 30% reduction in path detour distance.

[0081] Step S5: Simulation verification of the feasibility of the proposed experimental algorithm.

[0082] The above text combines Figure 1 This paper describes in detail an optimized control method for unmanned vehicles at smart airports based on rain and fog image enhancement, according to embodiments of this application. The following section describes in detail a device for detecting abnormal ship trajectories based on a variational autoencoder, according to embodiments of this application. This device includes: an environmental data acquisition module, a model building module, a model training module, a path planning module, and a simulation verification module.

[0083] The environmental data acquisition module is used to acquire surrounding environmental data and set...

[0084] The model building module is used to build an image de-raining and de-fogging network based on Gaussian modules;

[0085] The model training module is used to perform unsupervised training on the anomaly detection model;

[0086] The path planning module is used to plan the path for the autonomous vehicle;

[0087] The simulation verification module is used to perform simulation verification on the obtained model.

[0088] It should be noted that when a component is said to be "set on" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component.

[0089] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0090] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0091] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement, characterized in that, Includes the following steps: S1 acquires scene environment data by fusing LiDAR and visual sensors, and sets differentiated safety distance thresholds; S2, Construct an image de-raining and fogging network model based on a Gaussian encoder; The process includes collecting image data of airport boarding bridge scenes under rainy and foggy weather. This image data includes a training set for training and a test set for testing, with no overlap between the training and test sets. A Gaussian module-based image deraining model is constructed. An optimized attention-based generative adversarial network is used, with a Gaussian module added to the generator input. The Gaussian module-based image deraining model is trained using the training and validation sets to obtain derained images. The output is fused with a defogging module to enhance the airport boarding bridge scene images under rainy and foggy weather, restoring image details, textures, and scene information. A dark channel prior image defogging model based on an atmospheric scattering model is also constructed to obtain derained and foggy images. S3, Establish the vehicle dynamics model; This includes establishing a kinematic model of the autonomous vehicle based on its motion to satisfy the kinematic constraints during the process: . = + w Where w is the vehicle's angular velocity in the heading direction. For the vehicle's heading angle, , These are the coordinates of the rear axle center of the vehicle. For the center velocity of the following week; S4. Collision avoidance path planning for autonomous vehicles is performed based on the fusion algorithm of dynamic window and Bid-RRT; including scene adaptive weighted evaluation function, multi-vehicle cooperative collision avoidance and fault tolerance mechanism, improved bidirectional fast expanding random tree path search, and generation of dynamic collision avoidance path; S5. The feasibility of the proposed algorithm is verified through simulation experiments.

2. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, Step S1 involves acquiring scene environment data, including obstacle classification and trajectory prediction. The obstacle classification is divided into static fixed type, dynamic low speed type, and sudden dynamic type; the trajectory prediction is to use Kalman filter to predict the motion trajectory of dynamic obstacle and set a corresponding differentiated safety distance threshold.

3. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, The differentiated safety distance thresholds are set as follows: 2.5m for personnel, 1.5m for equipment, and 3m for sudden dynamic obstacles.

4. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, The optimized attention-based generative adversarial network described in step S2 includes a generator and a discriminator. The Gaussian model at the generator input end uses Gaussian filtering to preprocess the input image, the intermediate layer uses an attention mechanism, and the dehazing module at the output end uses a method based on dark channel priors.

5. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, The training parameters for the generative adversarial network in step S2 include: batch size of 16, number of iterations of 50,000, Adam optimizer, and learning rate initialized to 0.0002.

6. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, Step S4 also includes an improved bidirectional fast-expanding random tree path search. The expansion process of the improved bidirectional fast-expanding random tree includes: introducing a heuristic guiding factor for the corridor scene to restrict the random tree from sampling along the length of the channel; pruning the original path after connection using Bézier curves and curvature constraints, and dynamically adjusting the bidirectional tree node connectivity threshold according to the obstacle density.

7. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, The fault tolerance mechanism described in step S4 includes: automatically switching the fusion positioning mode and increasing the safety distance by 1.2 times when the sensor fails; triggering rapid local replanning when the obstacle change is ≥0.8m; and planning a path back to the nearest safe parking point when the passage is completely blocked.

8. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, Scene Adaptive Weighted Evaluation Function For weight The speed limit is dynamically adjusted based on obstacle density, vehicle speed, and peak operating hours, and is also integrated with differentiated speed limit constraints for the covered bridge area.

9. The method for optimized control of unmanned vehicles in smart airports based on rain and fog image enhancement according to claim 1, characterized in that, The multi-vehicle cooperative collision avoidance in step S4 includes: sharing vehicle status and planned paths through vehicle-to-vehicle communication modules, detecting path conflicts and implementing coordination strategies based on vehicle priorities, and predefining emergency avoidance channels for bridge scenarios for emergency avoidance in the event of multi-vehicle conflicts.

Images