Robot visual navigation optimization system based on edge calculation

Abstract

The invention is suitable for the technical field of artificial intelligence, and provides a robot visual navigation optimization system based on edge computing, the system comprises a visual sensor module, an edge computing unit, a navigation algorithm module, a communication module and a robot executor, the visual sensor module is used for collecting high-definition images or video streams, and the edge computing unit is used for processing the high-definition images or video streams; the edge calculation unit is responsible for locally executing a visual navigation algorithm, the navigation algorithm module is deployed on the edge calculation unit, and the navigation algorithm module is used for executing a visual odometer, synchronous positioning and map construction, environment perception, obstacle detection, local path planning and motion control; and the communication module is used for carrying out model updating, task scheduling and non-real-time data uploading with a cloud. Complex visual data processing and navigation decision are sunk to the robot body or near-end edge equipment, so that the navigation precision and the response speed of the robot in complex and diversified environments are remarkably improved, and the energy consumption efficiency is optimized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a robot vision navigation optimization system based on edge computing. Background Technology

[0002] Currently, robot visual navigation mainly relies on two modes: pure local processing and cloud processing.

[0003] In the purely local processing mode, the robot itself carries all computing resources and algorithms, independently completing the acquisition, processing, and navigation decisions of visual data. This type of method has advantages in environments with limited or no network access and offers fast response times. However, the robot's computing resources and power consumption are typically severely limited, making it difficult to support complex deep learning models and high-resolution image processing. This results in limited navigation accuracy and environmental perception capabilities, especially when dealing with large-scale maps or complex scenes, leading to performance degradation.

[0004] In the cloud-based processing mode, the robot uploads raw image data collected by its visual sensors to a remote cloud server. The cloud then performs high-performance computing and complex navigation algorithms, before sending decision commands back to the robot. This mode leverages the powerful computing capabilities of the cloud to support more complex algorithms and larger-scale map construction. However, its main drawbacks are the huge demand for network bandwidth and the unavoidable network transmission latency. In navigation tasks with high real-time requirements (such as high-speed movement and dynamic obstacle avoidance), transmission latency can lead to decision lag, affecting the real-time performance and safety of navigation. Furthermore, continuous data uploading increases the robot's energy consumption and may raise data privacy and security issues.

[0005] While the two existing technologies mentioned above each have their own advantages in terms of performance, they still face significant challenges in handling complex dynamic environments, meeting high real-time requirements, and balancing computing resources and energy consumption. For example, when encountering unstable networks, limited computing resources, or the need to process large amounts of high-resolution visual data, these methods often exhibit poor robustness and generalization, failing to meet the stringent requirements for navigation accuracy and response speed in fields such as intelligent inspection.

[0006] In addition, network bandwidth bottlenecks and data transmission latency are another prominent issue. Real-time uploading of high-resolution raw image data consumes enormous network bandwidth and introduces unacceptable transmission latency. For real-time navigation tasks requiring millisecond-level response times, such as dynamic obstacle avoidance, this latency is fatal, potentially leading to decision lag and severely impacting the real-time performance and safety of navigation.

[0007] Furthermore, energy consumption and battery life are also pressing issues that need to be addressed. Continuously transmitting large amounts of data to the cloud for processing via wireless networks significantly increases the energy consumption of the robot's communication modules, thereby shortening the robot's battery life and limiting its ability to operate for extended periods. At the same time, raw visual data may contain sensitive information, and uploading it to the cloud poses a risk of data leakage and misuse, failing to meet increasingly stringent data privacy protection requirements, thus raising data privacy and security risks.

[0008] Finally, existing systems also suffer from insufficient robustness and independent operation capabilities. Cloud-dependent robots will be unable to navigate and make decisions if the network connection is unstable or interrupted, leading to system failure. Robots that rely solely on local processing may perform poorly in complex and unknown environments due to insufficient computing power, thus affecting the overall robustness of the system. Therefore, there is a need to provide a robot visual navigation optimization system based on edge computing, aiming to solve the above problems. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention aims to provide a robot visual navigation optimization system based on edge computing. This system overcomes the limitations of current methods in terms of computing resources, network bandwidth, and real-time performance. The invention leverages the low latency, high bandwidth, and data privacy advantages of edge computing by deploying the core visual navigation algorithm on the robot itself or near-edge devices. Simultaneously, through optimizing the algorithm framework, model pruning and quantization, and heterogeneous computing, it ensures efficient image processing and decision-making even with limited edge resources. Furthermore, by designing an efficient cloud-edge collaboration mechanism, it balances the advantages of local processing and cloud services, ensuring the system's robustness, accuracy, and energy efficiency in complex dynamic environments.

[0010] This invention is implemented as follows: a robot visual navigation optimization system based on edge computing. The system includes a visual sensor module, an edge computing unit, a navigation algorithm module, a communication module, and a robot actuator. The visual sensor module is used to acquire high-definition images or video streams. The edge computing unit is responsible for locally executing the visual navigation algorithm. The navigation algorithm module is deployed on the edge computing unit and is used to perform visual odometry, simultaneous localization and mapping (SLT), environmental perception, obstacle detection, local path planning, and motion control. The communication module is used for model updates, task scheduling, and non-real-time data uploads with the cloud. The communication module supports high-speed local communication with the robot actuator and the visual sensor module. The robot actuator is used for motion control according to navigation commands.

[0011] As a further aspect of the present invention, the edge computing unit integrates a high-performance embedded processor, a dedicated AI acceleration chip, memory, and storage.

[0012] As a further aspect of the present invention, the working steps of the system are as follows: The robot continuously acquires images or video streams of the environment through its vision sensor module to obtain raw images. ,in The timestamp is used to transmit the original image to the edge computing unit; The original image is processed by the edge computing unit. Local preprocessing is performed to obtain key visual features using lightweight vision algorithms. ; By combining data from the IMU, a visual-inertial odometry algorithm is run at the edge computing unit to obtain pose estimation. ; Based on key visual features pose estimation The edge computing unit runs synchronous localization and mapping algorithms in real time to build and update local environment maps. ; Edge AI models for local environment mapping and real-time images The system performs analysis, obstacle detection, identification, and classification to obtain environmental perception results. Based on pose estimation Local environment map Based on the environmental perception results, the edge computing unit runs a local path planning algorithm in real time. This algorithm, combined with the robot's kinematic constraints, obstacle avoidance requirements, and task objectives, generates an optimal local motion path. .

[0013] As a further aspect of the present invention, the lightweight vision algorithm includes a pruned and quantized YOLOvX series algorithm and a lightweight semantic segmentation network. The YOLOvX series algorithm is used for object detection, and the lightweight semantic segmentation network is used to extract key visual features from the image. Key visual features This includes feature points, line segments, planar information, identified object bounding boxes, and semantic regions in the environment.

[0014] As a further aspect of the present invention, the step of running the visual-inertial odometry calculation method in the edge computing unit specifically includes: fusing the relative pose obtained by visual feature matching with the high-frequency pose change obtained by IMU pre-integration through Kalman filtering or factor graph optimization methods, wherein the IMU pre-integration is expressed as: in , , They represent arrive Position, velocity, and rotational pre-integral quantities within the time period, and These are the acceleration and angular velocity measured by the IMU. and It's the bias of the accelerometer and gyroscope. and It's noise. It is the gravity vector. It is an antisymmetric matrix.

[0015] As a further aspect of the present invention, the edge computing unit runs a synchronous localization and mapping algorithm in real time, including a front-end visual odometry, which is responsible for estimating the relative motion between adjacent frames; and a back-end optimization, which uses graph optimization methods to adjust the pose and map for global consistency.

[0016] As a further aspect of the present invention, the graph optimization method is expressed as minimizing the following error function: in These are the state variables to be optimized, including camera pose and 3D points. It is a measurement error. It is the covariance matrix. It is a constraint set, a local environment map. It includes information on the environmental structure and obstacle locations within the robot's current perception range, and is represented in the form of point cloud, grid map, or semantic map.

[0017] As a further aspect of the present invention, the steps of obstacle detection, identification, and classification specifically include: distinguishing between static and dynamic obstacles, and using a lightweight convolutional neural network or a Transformer-based detection model. The lightweight convolutional neural network and the detection model need to undergo model pruning, quantization, and knowledge distillation to adapt to the computing power of edge devices.

[0018] As a further aspect of the present invention, the model pruning reduces the model size and computational load by removing unimportant connections or neurons in the network; quantization converts floating-point parameters into low-ratio specific-point numbers to accelerate inference; and local semantic segmentation enables the robot to understand the semantic information of different regions in the environment.

[0019] As a further embodiment of the present invention, the local path planning algorithm is DWA, TEB, or a variant of sampling-based RRT / PRM. DWA predicts the robot's trajectory in a short time by sampling multiple sets of linear and angular velocities within the robot's current velocity window, and selects the optimal trajectory based on target proximity, obstacle distance, and velocity evaluation functions. TEB, on the other hand, represents the robot's trajectory as a series of time-optimized discrete poses and minimizes the energy function through an optimizer to meet dynamic constraints and obstacle avoidance requirements.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention significantly improves the robot's navigation accuracy and response speed in complex and diverse environments by offloading complex visual data processing and navigation decisions to the robot body or near-edge devices, while also optimizing energy efficiency. Furthermore, this invention enhances system robustness; even with unstable or interrupted network connections, the robot can still rely on its edge computing capabilities for autonomous navigation and decision-making, thus strengthening the system's independent operation and reliability. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the architecture of a robot vision navigation optimization system based on edge computing.

[0022] Figure 2 This is a flowchart of a robot vision navigation optimization system based on edge computing. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0025] like Figure 1As shown, this embodiment of the invention provides a robot visual navigation optimization system based on edge computing. The system includes a visual sensor module, an edge computing unit, a navigation algorithm module, a communication module, and a robot actuator. The visual sensor module is used to acquire high-definition images or video streams, such as RGB-D cameras, stereo cameras, or fisheye cameras. The edge computing unit is responsible for locally executing complex visual navigation algorithms. The edge computing unit integrates a high-performance embedded processor (such as NVIDIA Jetson series, Qualcomm Snapdragon, etc.), a dedicated AI acceleration chip (such as NPU, FPGA), memory, and storage. The navigation algorithm module is deployed on the edge computing unit and is used to execute algorithms such as visual odometry (VO), simultaneous localization and mapping (SLAM), environmental perception, obstacle detection, local path planning, and motion control. The communication module is used for low-frequency model updates, advanced task scheduling, and non-real-time data uploads with the cloud (remote server). The communication module supports high-speed local communication with the robot actuator and the visual sensor module. The robot actuator is used for motion control according to navigation instructions.

[0026] It should be noted that the embodiments of this invention mainly involve the fields of artificial intelligence, robotics, and edge computing, specifically including robot visual navigation, real-time image processing, distributed computing, and optimization methods. These technologies are not only important directions in current research on intelligent robots and the Internet of Things, but also key to solving the problem of autonomous robot navigation in complex dynamic environments. Edge computing refers to pushing computing and data storage to the network edge, i.e., near the data source, to reduce latency, save bandwidth, and improve response speed. In the embodiments of this invention, edge computing is used to deploy core algorithms in robot visual navigation, such as image data processing, feature extraction, local map building, real-time localization, and path planning, to the robot body or near-edge devices, thereby improving the real-time performance, accuracy, and energy efficiency of navigation. Robot visual navigation aims to achieve autonomous robot localization, map building, environmental perception, and path planning by analyzing environmental information acquired by visual sensors (such as cameras). The embodiments of this invention innovatively introduce an edge computing architecture, solving the limitations of traditional cloud processing or purely local processing in terms of computing resources, network bandwidth, and real-time performance, and improving the robot's navigation performance in complex environments. Real-time image processing is one of the key technical focuses of this proposal. To address the low-latency requirements of image processing in robot visual navigation, this invention employs lightweight visual algorithm deployment (specifically, breakthroughs lie in model pruning, quantization, knowledge distillation, and heterogeneous computing optimization for edge hardware) to achieve efficient local processing of image data. This significantly reduces data transmission and processing time, improving navigation response speed and accuracy. Through the comprehensive application of these technologies, this invention achieves higher real-time performance, accuracy, and robustness in robot visual navigation tasks.

[0027] like Figure 2 As shown in the embodiment of the present invention, the working steps of the system are as follows: S100 continuously acquires images or video streams of the environment through the robot's vision sensor module to obtain raw images. ,in The timestamp is used to transmit the original image to the edge computing unit; S200, through the edge computing unit, processes the original image Perform local preprocessing (such as denoising and distortion correction) and obtain key visual features using lightweight vision algorithms. ; To further optimize processing efficiency and accuracy, the S300 also incorporates data from the IMU (Inertial Measurement Unit) and runs a Visual-Inertial Odometry (VIO) algorithm on the edge computing unit to obtain pose estimation. ; S400, based on key visual features pose estimation The edge computing unit runs the Simultaneous Localization and Mapping (SLAM) algorithm in real time to build and update the local environment map. ; S500, edge AI model for local environment mapping and real-time images The system performs analysis, obstacle detection, identification, and classification to obtain environmental perception results. S600, based on pose estimation Local environment map Based on the environmental perception results, the edge computing unit runs a local path planning algorithm in real time. This algorithm, combined with the robot's kinematic constraints, obstacle avoidance requirements, and task objectives, generates an optimal local motion path. .

[0028] In this embodiment of the invention, the lightweight vision algorithm is deployed in a dedicated AI acceleration chip. The lightweight vision algorithm includes a pruned and quantized YOLOvX series algorithm and a lightweight semantic segmentation network. The YOLOvX series algorithm is used for object detection, and the lightweight semantic segmentation network is used to extract key visual features from the image. Key visual features This includes feature points, line segments, planar information, identified object bounding boxes, and semantic regions in the environment.

[0029] In this embodiment of the invention, the step of running the visual-inertial odometry (VIO) algorithm on the edge computing unit specifically includes: fusing the relative pose obtained by visual feature matching with the high-frequency pose change obtained by IMU pre-integration using Kalman filtering or factor graph optimization methods, wherein the IMU pre-integration is expressed as: in , , They represent arrive Position, velocity, and rotational pre-integral quantities within the time period, and These are the acceleration and angular velocity measured by the IMU. and It's the bias of the accelerometer and gyroscope. and It's noise. It is the gravity vector. It is an antisymmetric matrix. By fusing visual and IMU data, VIO can overcome the limitations of a single sensor and achieve robust and high-precision pose estimation in challenging scenarios such as insufficient visual information (e.g., sparse texture, lighting variations) or rapid robot movement and severe shaking. This technology is particularly suitable for complex and dynamic environments. This multi-sensor fusion technology is one of the key aspects of achieving high-precision and robust navigation at the edge in this invention embodiment. Through complementary advantages, it significantly improves the system's adaptability and reliability in various complex environments.

[0030] In this embodiment of the invention, the process of the edge computing unit running the simultaneous localization and mapping algorithm (visual SLAM algorithm) in real time includes front-end visual odometry, which is responsible for estimating the relative motion between adjacent frames; and back-end optimization, which uses graph optimization methods to adjust the pose and map for global consistency, and builds and updates the local environment map. For edge devices, lightweight or sparse SLAM algorithms, such as ORB-SLAM2 and VINS-Mono, are typically used. These algorithms maintain keyframes and local maps and use Bundle Adjustment (BA) or graph optimization methods to minimize reprojection errors, thereby reducing the demand for computing resources while ensuring a certain level of accuracy.

[0031] In this embodiment of the invention, the graph optimization method is expressed as minimizing the following error function: in These are the state variables to be optimized, including camera pose and 3D points. It is a measurement error. It is the covariance matrix. It is a constraint set, a local environment map. It includes information on the environmental structure and obstacle locations within the robot's current perception range, and is represented in the form of point clouds, grid maps, or semantic maps.

[0032] In this embodiment of the invention, the steps of obstacle detection, identification and classification specifically include: distinguishing between static obstacles (walls, furniture) and dynamic obstacles (pedestrians, other robots), and using a lightweight convolutional neural network (CNN) or a Transformer-based detection model. The lightweight convolutional neural network and the detection model need to be processed by optimization techniques such as model pruning, quantization (such as INT8 quantization) and knowledge distillation to adapt to the computing power of edge devices.

[0033] In this embodiment of the invention, model pruning reduces model size and computational cost by removing unimportant connections or neurons from the network; quantization converts floating-point parameters to low-ratio specific-point numbers to accelerate inference; and local semantic segmentation enables the robot to understand semantic information of different areas in the environment (such as traversable areas, dangerous areas, and specific target areas), providing a basis for more intelligent path planning. For example, lightweight backbone networks such as MobileNetV3 or EfficientNet are used for semantic segmentation to ensure real-time performance. The output of semantic segmentation is typically a pixel-level category label map, representing the semantic category to which each pixel belongs.

[0034] In this embodiment of the invention, commonly used local path planning algorithms include DWA (Dynamic Window Approach), TEB (Timed-Elastic Band), or sampling-based RRT / PRM variants. DWA predicts the robot's trajectory within a short time by sampling multiple sets of linear and angular velocities within the robot's current velocity window, and selects the optimal trajectory based on target proximity, obstacle distance, and velocity evaluation functions. TEB represents the robot's trajectory as a series of time-optimized discrete poses and minimizes the energy function through an optimizer to meet dynamic constraints and obstacle avoidance requirements. These algorithms can perform dynamic obstacle avoidance and path optimization under real-time requirements. Simultaneously, they enable real-time decision-making, such as immediately adjusting the path or executing an emergency stop command upon detecting a sudden obstacle. The generated motion commands are directly sent to the robot's actuators.

[0035] In this embodiment of the invention, to balance the advantages of local processing and cloud services, an efficient cloud-edge collaboration mechanism is adopted. The edge device only interacts with the cloud infrequently and efficiently when necessary (such as for global map updates, complex task scheduling, long-term data storage, or iterative training of algorithm models). The data uploaded to the cloud is typically high-level information processed and aggregated by the edge device (such as compressed local map keyframes, anomaly reports, and model inference result statistics), rather than raw high-bandwidth image streams, thereby minimizing network load. The cloud can then perform global path optimization, multi-robot collaborative scheduling, or retrain and update the edge AI model based on the data uploaded by the edge device, and issue new model versions or global instructions to the edge device.

[0036] Compared to cloud computing or traditional architectures, this system offers significant advantages. Firstly, in terms of navigation accuracy, by processing the latest visual data locally in real-time and fusing multiple sensors, combined with the hardware collaborative processing capabilities of the high-performance AI acceleration chip in the edge computing unit, visual data is efficiently transformed into an executable hardware instruction set. This enables the acquisition of high-precision positioning information and environmental perception, effectively avoiding navigation deviations caused by network latency, especially in dynamic environments. Secondly, regarding response speed, by offloading core decision-making logic to the edge, latency caused by data transmission and cloud computing is significantly reduced, achieving millisecond-level real-time response, which is crucial for dynamic obstacle avoidance and emergency handling. Furthermore, this system significantly reduces network bandwidth requirements; most image data is processed locally, with only a small amount of high-level information or instructions transmitted, effectively alleviating network bandwidth bottlenecks and thus reducing communication costs. In terms of energy consumption, the system reduces the power consumption of wireless transmission of large amounts of raw data, while utilizing the high energy efficiency of the edge AI acceleration chip to reduce overall system energy consumption and extend the robot's runtime. This system also enhances data privacy and security. Sensitive visual data is processed and anonymized locally, reducing the risk of data exposure on public networks and meeting data privacy protection requirements. Finally, this system improves robustness. Even in situations of unstable or interrupted network connections, the robot can still rely on its edge computing capabilities for autonomous navigation and decision-making, enhancing the system's independent operation and reliability.

[0037] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0038] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0039] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0040] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A robot vision navigation optimization system based on edge computing, characterized in that, The system includes a vision sensor module, an edge computing unit, a navigation algorithm module, a communication module, and a robot actuator. The vision sensor module is used to acquire high-definition images or video streams. The edge computing unit is responsible for locally executing the visual navigation algorithm. The navigation algorithm module is deployed on the edge computing unit and is used to perform visual odometry, simultaneous localization and mapping (SLT), environmental perception, obstacle detection, local path planning, and motion control. The communication module is used for model updates, task scheduling, and non-real-time data uploads with the cloud. The communication module supports high-speed local communication with the robot actuator and the vision sensor module. The robot actuator is used for motion control according to navigation commands.

2. The robot vision navigation optimization system based on edge computing according to claim 1, characterized in that, The edge computing unit integrates a high-performance embedded processor, a dedicated AI acceleration chip, memory, and storage.

3. The robot vision navigation optimization system based on edge computing according to claim 1, characterized in that, The system operates as follows: The robot continuously acquires images or video streams of the environment through its vision sensor module to obtain raw images. ,in The timestamp is used to transmit the original image to the edge computing unit; The original image is processed by the edge computing unit. Local preprocessing is performed to obtain key visual features using lightweight vision algorithms. ; By combining data from the IMU, a visual-inertial odometry algorithm is run at the edge computing unit to obtain pose estimation. ; Based on key visual features pose estimation The edge computing unit runs synchronous localization and mapping algorithms in real time to build and update local environment maps. ; Edge AI models for local environment mapping and real-time images The system performs analysis, obstacle detection, identification, and classification to obtain environmental perception results. Based on pose estimation Local environment map Based on the environmental perception results, the edge computing unit runs a local path planning algorithm in real time. This algorithm, combined with the robot's kinematic constraints, obstacle avoidance requirements, and task objectives, generates an optimal local motion path. .

4. The robot vision navigation optimization system based on edge computing according to claim 3, characterized in that, The lightweight vision algorithm includes pruned and quantized YOLOvX series algorithms and a lightweight semantic segmentation network. The YOLOvX series algorithms are used for object detection, and the lightweight semantic segmentation network is used to extract key visual features from the image. Key visual features This includes feature points, line segments, planar information, identified object bounding boxes, and semantic regions in the environment.

5. The robot vision navigation optimization system based on edge computing according to claim 3, characterized in that, The step of running the visual-inertial odometry calculation method in the edge computing unit specifically includes: fusing the relative pose obtained by visual feature matching with the high-frequency pose change obtained by IMU pre-integration through Kalman filtering or factor graph optimization methods, wherein the IMU pre-integration is expressed as: in , , They represent arrive Position, velocity, and rotational pre-integral quantities within the time period, and These are the acceleration and angular velocity measured by the IMU. and It's the bias of the accelerometer and gyroscope. and It's noise. It is the gravity vector. It is an antisymmetric matrix.

6. The robot vision navigation optimization system based on edge computing according to claim 3, characterized in that, The edge computing unit runs a synchronous localization and mapping algorithm in real time, including a front-end visual odometry, which is responsible for estimating the relative motion between adjacent frames; and a back-end optimization, which uses graph optimization methods to adjust the pose and map for global consistency.

7. The robot vision navigation optimization system based on edge computing according to claim 6, characterized in that, The graph optimization method is expressed as minimizing the following error function: in These are the state variables to be optimized, including camera pose and 3D points. It is a measurement error. It is the covariance matrix. It is a constraint set, a local environment map. It includes information on the environmental structure and obstacle locations within the robot's current perception range, and is represented in the form of point cloud, grid map, or semantic map.

8. The robot vision navigation optimization system based on edge computing according to claim 3, characterized in that, The steps for obstacle detection, identification, and classification specifically include: distinguishing between static and dynamic obstacles, and using a lightweight convolutional neural network or a Transformer-based detection model. The lightweight convolutional neural network and detection model need to undergo model pruning, quantization, and knowledge distillation to adapt to the computing power of edge devices.

9. The robot vision navigation optimization system based on edge computing according to claim 8, characterized in that, The model pruning reduces model size and computation by removing unimportant connections or neurons from the network; quantization converts floating-point parameters to low-ratio specific-point numbers to accelerate inference; and local semantic segmentation enables the robot to understand semantic information in different regions of the environment.

10. The robot vision navigation optimization system based on edge computing according to claim 3, characterized in that, The local path planning algorithm is DWA, TEB, or a variant of sampling-based RRT / PRM. DWA predicts the robot's trajectory in a short time by sampling multiple sets of linear and angular velocities within the robot's current velocity window, and selects the optimal trajectory based on target proximity, obstacle distance, and velocity evaluation functions. TEB, on the other hand, represents the robot's trajectory as a series of time-optimized discrete poses and minimizes the energy function through an optimizer to meet dynamic constraints and obstacle avoidance requirements.

Images