A camera parameter auto-optimization system, method, medium, and apparatus for binocular depth estimation

Abstract

A camera parameter automatic optimization system for binocular depth estimation, a true value acquisition module acquires a depth true value map in a simulation environment; an Auto-CAM learning module adopts a deep reinforcement learning framework based on an actor-critic, learns in reverse propagation according to a current camera configuration, a state vector generated by left and right view images and a depth error, so as to optimize camera parameters such as focal length, baseline distance and distortion coefficient; a deployment module outputs an optimal camera parameter configuration under a specific scene, a camera calibration module acquires accurate camera parameters through corner scanning, an image acquisition module acquires RGB binocular images, a depth map prediction module outputs a predicted disparity map by using a stereo matching network algorithm and converts the predicted disparity map into a depth map, and finally, a visual rendering module performs filtering and color processing. The application not only improves the accuracy of depth estimation and the adaptability of the system, but also improves the accuracy and efficiency of camera calibration, optimizes the user experience and visual effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a camera parameter automatic optimization system, method, medium and equipment for binocular depth estimation. BACKGROUND

[0002] In the field of binocular depth estimation, the selection and optimization of camera parameters are crucial for the accuracy and robustness of depth estimation. In-depth research on camera parameters is a key link to improve the performance of depth estimation algorithms and promote the development of computer vision applications.

[0003] Focal length, as one of the basic parameters of a camera, determines the size of the object projection on the imaging plane, and its optimization is crucial for improving the accuracy of depth estimation. Although researchers have explored the optimal focal length setting through experimental adjustment and multi-scene testing, this method is not only time-consuming and laborious, but also difficult to ensure optimal depth estimation results in all cases. In addition, the adjustment of focal length may also be restricted by camera hardware limitations and cost factors, further limiting its optimization space.

[0004] Baseline is the distance between the two cameras in a binocular camera system, and is one of the key factors affecting depth estimation. Different baseline settings directly affect the density and range of the depth map. Researchers explore the possibility of obtaining better depth estimation in specific scenarios by adjusting the length of the baseline. A too long baseline may result in too small parallax for distant objects, making it difficult to accurately estimate depth; while a too short baseline may reduce depth resolution, affecting the depth estimation accuracy of close-range objects. In addition, the adjustment of the baseline may also be restricted by factors such as device volume, weight and cost, making it difficult to find the best baseline configuration in practical applications.

[0005] Multi-camera systems can provide more depth information than binocular systems, but also face more technical challenges. By optimizing camera layout and configuration, researchers hope to achieve higher quality depth estimation. This research involves the spatial layout and angle selection between cameras, aiming to find the best configuration scheme. Although the rise of deep learning has brought new perspectives to camera parameter research, its application in binocular depth estimation still has limitations. Existing deep learning models often rely on large-scale datasets for training, and the acquisition and processing of these datasets are costly. In addition, the sensitivity and generalization ability of deep learning models to camera parameters still need to be improved, which limits their applicability in different scenarios and tasks. At the same time, the interpretability of deep learning models is poor, making it difficult for researchers to deeply understand the complex relationship between camera parameters and depth estimation.

[0006] Therefore, the selection and optimization of camera parameters in the current binocular depth estimation field still faces many challenges and deficiencies. In order to overcome these shortcomings, it is necessary to explore more efficient, universal and adaptive optimization strategies and methods SUMMARY

[0007] In order to overcome the deficiencies of the prior art, the present application provides a camera parameter automatic optimization system and method for binocular depth estimation.

[0008] The technical solution of the present application is as follows:

[0009] A camera parameter automatic optimization system for binocular depth estimation, characterized in that it comprises:

[0010] A true value acquisition module for acquiring a depth true value map in a simulation environment;

[0011] An Auto-CAM learning module using a deep reinforcement learning framework based on actor-critic, which learns in reverse according to the current camera configuration, state vector and depth error, wherein the state vector is generated by the encoder of the left view and right view images through shared weights, the depth error is calculated from the depth map generated by the pre-trained stereo matching model and the depth true value map, and the camera configuration includes focal length, baseline distance and distortion coefficient;

[0012] An Auto-CAM deployment module for outputting the optimal camera parameter configuration according to the trained reinforcement learning agent in a specific scene;

[0013] A camera calibration module for obtaining accurate camera intrinsic and extrinsic parameters through corner scanning after changing the camera parameters;

[0014] An image acquisition module for acquiring RGB binocular images;

[0015] A depth map prediction module for outputting a predicted disparity map according to the deployed stereo matching network algorithm, and converting the disparity map into a depth map using the intrinsic and extrinsic parameters obtained by the camera calibration module;

[0016] A visual rendering module for filtering and color processing the obtained depth map to present the binocular depth map.

[0017] Further, the Auto-CAM learning module uses a reward function, which includes:

[0018] A first reward term for calculating the difference between the depth estimation error under the current camera configuration and the target depth error set by humans;

[0019] A second reward term for analyzing the error estimated by the model; wherein the analysis model is as follows:

[0020]

[0021] where, ΔD is the accuracy of binocular depth, Δd px is the accuracy of disparity map in pixels, b and f are the baseline and focal length of the camera, respectively, D gt is the ground truth of binocular depth, r hor is the pixel density in the horizontal direction, w sen is the width of the sensor.

[0022] The third reward term, RoI exclusion penalty, is used to ensure that the pixels of the RoI are always within the field of view (FoV) of the binocular camera view during the update of the camera configuration.

[0023] The specific process of the Auto-CAM deployment module includes:

[0024] The binocular camera captures left and right views under the current camera configuration;

[0025] The left and right images are fed into the encoder with shared weights, respectively, to generate two feature vectors, and are integrated into a state vector through a specific fusion strategy;

[0026] The reinforcement learning agent takes the state vector as input to generate an action vector;

[0027] The binocular camera adjusts its configuration according to the action vector and captures new image data.

[0028] Further, it further includes an embedded system including a deep learning edge computing device and a variable baseline binocular camera system, which can obtain high-precision RGB binocular images according to specific scenes, input the images into the encoder to obtain a state vector, then the reinforcement learning agent generates an action vector, finally the output of the more optimal configuration changes the parameters of the binocular camera and performs camera calibration, and then the embedded system performs depth map prediction and visual rendering according to the embedded deep neural network.

[0029] On the other hand, the present application also provides a method for automatically optimizing binocular depth estimation camera parameters, characterized in that it comprises the following steps:

[0030] S1. Obtain the depth ground truth map in the simulation environment;

[0031] S2. Use the actor-critic based deep reinforcement learning framework to perform backpropagation learning according to the current camera configuration, state vector and depth error, and output the more optimal camera configuration;

[0032] S3. Use the trained reinforcement learning agent to output the optimal camera parameter configuration in a specific scene;

[0033] S4. After changing the camera parameters, obtain accurate camera intrinsic and extrinsic parameters through corner scanning.

[0034] S5. Acquire RGB binocular images and output the predicted disparity map based on the deployed stereo matching network algorithm;

[0035] S6. Use the intrinsic and extrinsic parameters obtained from the camera calibration module to convert the disparity map into a depth map;

[0036] S7. Filter and color process the obtained depth map to finally present a higher quality binocular depth map.

[0037] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed, it implements the steps of the method described herein.

[0038] A computer device includes a processor, a memory, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the method described above.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1) For exploration scenarios with ground truth assistance, this invention outperforms all existing technologies in terms of performance and speed; while for deployment scenarios without ground truth assistance, this invention, as the only existing method, can improve the accuracy of the binocular depth map by outputting better camera parameters based solely on the input binocular images. This invention provides a unified solution for two different scenarios.

[0041] 2) This invention applies deep reinforcement learning to the automatic optimization of camera parameters, breaking through the limitations of traditional manual adjustment or experience-based presets. It can automatically output the optimal camera parameter configuration in specific scenarios, and has broad application prospects and market demand. Attached Figure Description

[0042] Figure 1 Flowchart of Automatic Camera Parameter Configuration Optimization System

[0043] Figure 2 Auto-CAM Framework Learning Module Flowchart

[0044] Figure 3 Auto-CAM Framework Deployment Module Flowchart

[0045] Figure 4 Example diagram of the results of deep optimization during the exploration phase of the Auto-CAM framework.

[0046] Figure 5Example depth diagrams before and after Auto-CAM optimization for the simulator deployment scenario in the experiment.

[0047] Figure 6 : Schematic diagram of the optimization path in the exploration phase of the Auto-CAM framework

[0048] Figure 7 For the actual setup and deployment scenario of the experiment, here are some depth example diagrams before and after Auto-CAM optimization. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the scope of protection of the present invention.

[0050] An analytical model was constructed that considers the impact of factors such as focal length, baseline, pixel density, and stereo matching model accuracy on the accuracy of binocular depth estimation. The specific formula is as follows:

[0051]

[0052] Where ΔD is the depth accuracy of the binoculars, Δd px This refers to the resolution of the disparity map (in pixels), where b and f are the camera's baseline and focal length, respectively, and D... gt It is the true value of binocular depth, r hor It is the pixel density in the horizontal direction, w sen It is the width of the sensor.

[0053] The mathematical model in the formula provides the following trends for improving the accuracy of BDE: 1) Camera parameters: Depth error can be reduced by increasing the product of baseline and focal length (b·f); 2) Camera specifications: Choosing a higher resolution image sensor does not reduce depth error, but a higher horizontal pixel density (r) hor / w sen This will be helpful. 3) Stereo matching algorithm: Optimize the stereo matching algorithm (i.e., reduce Δd). px 4) Target distance: Depth error increases with the distance from the target to the binocular vision system.

[0054] Nevertheless, the trends summarized above do not always apply to simulation scenario deployments because of the parallax error (Δd) px This is not a fixed value; it varies with the camera parameters (i.e., b and f). Figure 1 In the figure, the top subplot shows how the disparity error varies with b and FoV. Since FoV is inversely proportional to f, in Figure 2In the intermediate subplot, we can see that the optimal setting (minimum log|ΔDAM|) is achieved when FoV is 60° and b is 2 (i.e., maximum b·f). However, the actual depth error obtained from experimental measurements confirms that optimizing the camera configuration is a more complex problem, where the error distribution differs from the mathematical model.

[0055] Therefore, the Auto-CAM framework is proposed, which uses deep reinforcement learning to automatically and quickly obtain ideal camera parameters for different scenarios. Auto-CAM consists of three core components: actor commentary agent, encoder, and environment, and is divided into a learning phase and a deployment phase.

[0056] Actor-Critic Agent: Responsible for generating motion vectors with camera parameter configurations and evaluating the quality of these motion vectors. It is trained and optimized using deep reinforcement learning algorithms (such as the actor-critic algorithm).

[0057] Encoder: A pre-trained Transformer model processes the input stereo image to generate a high-dimensional state vector. This state vector reflects the geometric and texture information in the image, providing input data for the reinforcement learning agent.

[0058] Environment: Provides input interfaces for stereo images and interface for adjusting camera parameters. It receives action vectors (camera parameter configuration) generated by the reinforcement learning agent and returns the corresponding reward signal and a new state vector. The environment part is also responsible for camera calibration and depth map prediction and visualization rendering.

[0059] During the learning phase, Auto-CAM optimizes its reinforcement learning agent through continuous trial and error and reward feedback. In the deployment phase, the trained agent is used to quickly obtain ideal camera parameter configurations and apply them to real-world scenarios.

[0060] Finally, an embedded system implementing the above framework was built, including a deep learning edge computing device from Jetson Orin Nano and a stereo camera system with variable baseline. High-precision RGB stereo images can be obtained for specific scenes and input into an encoder to obtain state vectors. A reinforcement learning agent then generates action vectors. Finally, the parameters of the stereo camera are adjusted and calibrated based on the optimized output configuration. The embedded system then uses its built-in deep neural network to predict depth maps and render visualizations.

[0061] Camera parameter optimization in binocular depth estimation plays a crucial role in several high-precision fields, especially in applications like autonomous driving and defect detection, where accuracy requirements are extremely high. Binocular depth estimation generates a disparity map using a pair of simultaneously captured cameras to calculate the depth information of a target scene. Accurate optimization of camera parameters (such as intrinsic and extrinsic parameters, distortion coefficients, and baseline distance) directly impacts the accuracy and robustness of the depth estimation results. In autonomous driving, vehicles need to perceive the distance and position of obstacles, road structures, and dynamic targets in the 3D environment in real time to complete core tasks such as path planning, obstacle avoidance, and lane keeping. Unoptimized camera parameters can lead to biased depth estimation results, resulting in serious safety hazards. For example, a small error in the baseline distance can cause significant deviations in depth measurement, especially in detecting distant targets. Through camera parameter optimization, autonomous driving systems can significantly improve the accuracy of depth estimation, ensuring reliable decisions in diverse scenarios such as urban congestion, rural roads, and highways. Particularly in complex dynamic environments, such as detecting pedestrians, cyclists, or other vehicles, accurate depth information can significantly improve vehicle reaction speed and safety. In the field of industrial defect detection, binocular depth estimation generates high-resolution 3D point cloud data to identify minute defects (such as cracks, dents, and wear) on workpiece surfaces. Optimized camera parameters ensure the stability of parallax calculation and the accuracy of depth maps, thus providing more reliable detection results. In some critical manufacturing sectors (such as semiconductor, aerospace, and automotive manufacturing), the accuracy of detection directly determines product quality and safety. For the detection of irregular surfaces or highly reflective materials, the optimization of camera intrinsic parameters (such as focal length and principal point offset) and extrinsic parameters (such as baseline distance and relative angle between cameras) is particularly important. Furthermore, for harsh conditions in certain industrial scenarios (such as uneven lighting and dusty environments), camera parameter optimization combined with advanced calibration algorithms can effectively improve the system's anti-interference capability and adaptability. Given the importance of binocular depth estimation camera parameter optimization in autonomous driving and industrial defect detection, this invention has been implemented in two edge applications. By integrating multiple modules such as ground truth acquisition, Auto-CAM learning, deployment, camera calibration, image acquisition, depth map prediction, and visualization rendering, this invention achieves automatic optimization of camera parameters and accurate calculation of depth estimation. This not only improves the overall performance and efficiency of the system, but also provides more reliable technical support and innovative solutions for fields such as autonomous driving and industrial defect detection.

[0062] like Figure 1The diagram shows a flowchart of an automatic camera parameter configuration optimization system applied to a simulator. It includes a ground truth acquisition module, an Auto-CAM learning module, an Auto-CAM deployment module, a camera calibration module, an image acquisition module, a depth map prediction module, and a visualization rendering module. Specifically: the ground truth acquisition module uses the simulator to obtain a depth ground truth map; the Auto-CAM learning module performs backpropagation learning of reinforcement learning based on the current camera configuration, state vector, and depth error; the Auto-CAM deployment module stops learning and uses the current state vector to output a better camera configuration; the camera calibration module obtains accurate camera intrinsic and extrinsic parameters through corner scanning after changing camera parameters; the image acquisition module obtains high-precision RGB stereo images; the depth map prediction module outputs a predicted disparity map based on the deployed stereo matching network algorithm, and then uses the intrinsic and extrinsic parameters obtained from the camera calibration module to transform the depth map; the visualization rendering module filters and processes the obtained depth map to ultimately present a higher-quality stereo depth map.

[0063] Example 1 establishes a closed-loop automated optimization process. From ground truth acquisition to depth map prediction and visualization rendering, each module plays an indispensable role. Through advanced technologies such as reinforcement learning and camera calibration, the system can automatically optimize camera parameter configuration, thereby improving the accuracy and efficiency of depth estimation. This automated optimization system has broad application prospects and significant practical value in high-precision fields such as autonomous driving and industrial defect detection.

[0064] Figure 2 This is a flowchart of the Auto-CAM framework's learning module, as shown in the figure. It aims to automatically optimize the parameter configuration of the stereo camera using a reinforcement learning framework. The following are the steps of the iterative learning phase:

[0065] (1) Initialize camera configuration: The stereo camera captures the left and right views under the current camera configuration. For initialization, the baseline and focal length are randomly selected from the value range.

[0066] (2) Image capture and feature extraction: The stereo camera captures the left and right views in the current camera configuration. The left and right images are fed into an encoder with shared weights and output two feature vectors.

[0067] (3) Feature fusion: The feature vectors of the left and right images are fused into a single state vector by element-wise addition.

[0068] (4) Reinforcement learning agent decision-making: The reinforcement learning agent (RL agent) processes the state vector and generates action vectors, which describe how to adjust the camera parameters to optimize the accuracy of depth estimation.

[0069] (5) Camera configuration update and image recapture: Set up the stereo camera with the new configuration and recapture the left and right view images and ground truth depth maps.

[0070] (6) Depth map estimation and reward calculation:

[0071] Estimated depth maps are generated using a pre-trained stereo matching model (such as HITNet). During reinforcement learning, the parameters of the stereo matching model are frozen to ensure the stability and consistency of reward calculation.

[0072] A reward is calculated based on the difference between the estimated depth map and the ground truth depth map. The reward function consists of three components: a reward for depth estimation error, a reward for model estimation error, and a penalty for RoI exclusion.

[0073] (7) Update the model parameters of the actor and critic networks: The encoder converts the pixels of the left and right view images into state vectors (embeddings). The encoder inherits the encoder part of a mask autoencoder (MAE) pre-trained on the ImageNet dataset. The optimization objective can be refactored into our reward function to maximize:

[0074]

[0075] The first reward term calculates the depth estimation error under the current camera configuration, which is a manually set target depth error. The second reward term represents the error estimated by the analytical model. The third reward term is an RoI exclusion penalty to ensure that the pixels of the RoI are always within the FoV of the stereo camera view during camera configuration updates. alpha and beta are reward scaling factors; it returns 1 if the condition is met, and 0 otherwise. The second reward term greatly aids DRL convergence, inspired by reward shaping. It uses the AM estimation error as a potential energy function to avoid sparse rewards during spatial exploration.

[0076] The specific process during the Auto-CAM deployment phase is as follows: Figure 3 As shown. The deployment phase is the core application phase of the Auto-CAM system, directly utilizing a pre-trained reinforcement learning agent to output the optimal camera parameter configuration based on the specific scene. This dynamic adjustment method can significantly improve the accuracy of binocular depth estimation, providing reliable support for various practical application scenarios. The specific process is as follows:

[0077] (1) First, the stereo camera captures the left and right views under the current camera configuration. These images record the basic visual information of the current scene, laying the data foundation for subsequent depth estimation. (2) Next, the left and right images are fed into a shared-weight encoder. The encoder extracts important features from the images through a convolutional network, generating two feature vectors respectively. These feature vectors are integrated into a state vector through a specific fusion strategy, representing the current camera configuration and environmental conditions. (3) Subsequently, the reinforcement learning agent takes the state vector as input, performs analysis and decision-making, and generates action vectors. The action vectors describe how to adjust camera parameters to optimize the configuration, including focal length, baseline distance, and distortion coefficients. (4) Finally, the stereo camera adjusts its own configuration according to the action vectors and captures new image data to achieve closed-loop optimization.

[0078] To evaluate the practical effectiveness of Auto-CAM, it was deployed and validated in two edge applications: autonomous driving and defect detection. On CARLA, a high-efficiency simulator capable of acquiring real-time depth ground truth maps, we comprehensively evaluated Auto-CAM's performance through experiments. In the experimental design, we selected several classic stereo matching models (such as PSMNet and HITNet) as benchmark tools and compared them with currently widely used fixed camera parameter configurations and other optimization methods. Table 1 shows the specific experimental results. The data clearly shows that Auto-CAM performs excellently in optimizing camera configuration, significantly improving the accuracy of depth estimation. The depth maps generated by Auto-CAM are more accurate in estimating target distances, reducing errors caused by improper camera parameters. This result demonstrates that the reinforcement learning-based dynamic camera configuration method can more flexibly adapt to different scenarios, providing high-quality depth information support for the perception module in autonomous driving.

[0079] Furthermore, we conducted a quantitative analysis of the exploration efficiency of Auto-CAM. In reinforcement learning methods, the exploration convergence speed is an important indicator for measuring algorithm efficiency. Figure 4 The results of six methods, including random search (RS), simulated annealing (SA), and Bayesian optimization (BO), are presented. It can be concluded that Auto-CAM achieves significant performance improvements while also substantially optimizing exploration efficiency. In practical applications, Auto-CAM can quickly converge to the optimal parameter configuration with fewer iterations, thereby reducing time costs.

[0080] To further analyze the exploration path of Auto-CAM in optimizing camera parameter configuration, we recorded its exploration process on multiple stereo matching models and visualized it using heatmaps, such as... Figure 6As shown in the heatmap, red areas represent larger depth errors caused by configuration, while blue areas represent smaller depth errors. The heatmap clearly shows that the Auto-CAM method exhibits a stable and efficient optimization process. Unlike other methods that may wander or detour in the parameter space, Auto-CAM can directly and accurately approach the optimal region, demonstrating the efficiency and reliability of its decision path. This intuitive representation of the optimization process further proves the powerful adaptability of reinforcement learning agents in complex parameter spaces.

[0081] To evaluate the applicability of Auto-CAM in practical deployments, we designed a special experimental scenario to verify its accuracy level even without a ground truth depth map. The experimental results are shown in Table 2. In this scenario, traditional optimization methods (such as RS, SA, BO, etc.) become infeasible due to the lack of a true ground truth depth map, as the objective function cannot be defined. Auto-CAM, however, does not rely on a ground truth depth map and dynamically adjusts itself entirely through reinforcement learning, thus functioning correctly. In this experiment, we compared Auto-CAM with the Expert Setting (ES) method. The results show that even without a ground truth map, Auto-CAM significantly outperforms ES. For example, in the PSMNet model test, the camera configuration optimized by Auto-CAM improved performance by 2.18 times compared to the fixed configuration of ES. This result demonstrates that Auto-CAM has higher robustness and wider applicability, providing strong support for the promotion of depth estimation in real-world scenarios.

[0082] In addition to quantitative analysis, we also conducted a qualitative evaluation of Auto-CAM, and... Figure 5 and Figure 7 The document provides visualizations of depth maps in these two edge applications. A comparison reveals a significant improvement in the quality of depth maps generated by the Auto-CAM-optimized camera configuration. Specifically, the optimized depth maps not only reduce noise and artifacts but also demonstrate greater clarity in detail, particularly in scenes with sparse textures or complex lighting. This improvement is especially crucial for scenarios such as autonomous driving and defect detection. For instance, in autonomous driving, accurate depth maps help vehicles correctly determine the position and distance of obstacles ahead, improving path planning and obstacle avoidance capabilities. In defect detection, clear depth maps help detection systems better identify the details of surface defects, enhancing the accuracy of product quality control.

[0083] In summary, Auto-CAM, through reinforcement learning agents that dynamically adjust camera parameters, demonstrates superior performance across multiple edge computing scenarios. Experimental data shows that Auto-CAM significantly outperforms existing methods in depth estimation accuracy, exploration efficiency, and adaptability. Furthermore, its independence from ground-value depth maps enhances its practicality and feasibility for real-world deployment. Whether for real-time perception requirements in autonomous driving or high-precision demands in industrial defect detection, Auto-CAM provides a reliable solution for depth estimation and camera configuration optimization.

[0084]

[0085] Table 1: Depth error ΔD of Auto CAM and different methods during the learning phase of optimization.

[0086]

[0087] Table 2: Depth error ΔD for Auto CAM and different methods during the deployment phase.

Claims

1. An automatic camera parameter optimization system for binocular depth estimation, characterized in that, include: The truth acquisition module is used to acquire depth truth maps in a simulation environment; The Auto-CAM learning module employs an actor-critic deep reinforcement learning framework. It performs backpropagation learning based on the current camera configuration, state vector, and depth error. The state vector is generated by an encoder with shared weights from the left and right view images. The depth error is calculated from the depth map and the ground truth depth map generated by the pre-trained stereo matching model. The camera configuration includes focal length, baseline distance, and distortion coefficient. The Auto-CAM deployment module is used to output the optimal camera parameter configuration based on the trained reinforcement learning agent in a specific scenario. The camera calibration module is used to obtain accurate camera intrinsic and extrinsic parameters by corner scanning after changing camera parameters; The image acquisition module is used to acquire RGB binocular images; The depth map prediction module outputs the predicted disparity map based on the deployed stereo matching network algorithm, and converts the disparity map into a depth map using the intrinsic and extrinsic parameters obtained from the camera calibration module. The visualization rendering module is used to filter and process the color of the obtained depth map to present a binocular depth map; The specific process of the Auto-CAM deployment module includes: - The binocular camera captures the left and right views under the current camera configuration; - The left and right images are fed into a shared-weight encoder to generate two feature vectors, which are then integrated into a single state vector using a specific fusion strategy. - Reinforcement learning agents take state vectors as input and generate action vectors; - The binocular camera adjusts its configuration based on motion vectors and captures new image data.

2. The automatic camera parameter optimization system for binocular depth estimation according to claim 1, characterized in that, The Auto-CAM learning module employs a reward function, including: The first reward item is to calculate the difference between the depth estimation error under the current camera configuration and the manually set target depth error. The second reward item is based on the error estimated by the analysis model; wherein the analysis model is as follows: Where ΔD is the depth accuracy of the binoculars, Δd px This refers to the accuracy of the disparity map, expressed in pixels. b and f represent the camera's baseline and focal length, respectively, and D... gt It is the true value of binocular depth, r hor It is the pixel density in the horizontal direction, w sen It is the width of the sensor; The third bonus, RoI Exclusion Penalty, is designed to ensure that RoI pixels are always within the field of view (FoV) of the binocular camera view during camera configuration updates.

3. The automatic camera parameter optimization system for binocular depth estimation according to claim 1, characterized in that, It also includes an embedded system comprising a deep learning edge computing device and a variable baseline binocular camera system capable of obtaining high-precision RGB binocular images for specific scenes and inputting the images into an encoder to obtain state vectors. Then, a reinforcement learning agent generates action vectors. Finally, the parameters of the binocular camera are changed and the camera is calibrated based on the optimal configuration of the output. Subsequently, the embedded system performs depth map prediction and visualization rendering based on the onboard deep neural network.

4. A method for automatic optimization of parameters of a binocular depth estimation camera, characterized in that, Includes the following steps: S1. Obtain the depth truth map in the simulation environment; S2. Employ an actor-critic deep reinforcement learning framework, perform backpropagation learning based on the current camera configuration, state vector, and depth error, and output a better camera configuration; S3. In a specific scenario, use the trained reinforcement learning agent to output the optimal camera parameter configuration; S4. After changing the camera parameters, obtain accurate camera intrinsic and extrinsic parameters through corner scanning. S5. Acquire RGB binocular images and output the predicted disparity map based on the deployed stereo matching network algorithm; S6. Use the intrinsic and extrinsic parameters obtained from the camera calibration module to convert the disparity map into a depth map; S7. Filter and color process the obtained depth map to finally present a higher quality binocular depth map; The specific process of step S3 includes: - The binocular camera captures the left and right views under the current camera configuration; - The left and right images are fed into a shared-weight encoder to generate two feature vectors, which are then integrated into a single state vector using a specific fusion strategy. - Reinforcement learning agents take state vectors as input and generate action vectors; - The binocular camera adjusts its configuration based on motion vectors and captures new image data.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the steps of the method as described in claim 4.

6. A computer device comprising a processor, a memory, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in claim 4.

Images