An event camera based low complexity pose estimation system and method

Abstract

This invention discloses a low-complexity pose estimation system and method based on an event camera, belonging to the interdisciplinary fields of machine vision and neuromorphic perception. It includes an event camera module, a spatiotemporal joint data selection module, a state space modeling module, an event-triggered state estimation module, and a pose estimation result post-processing module. The method relies on these modules, proceeding through event acquisition, spatiotemporal data filtering, state modeling, ellipsoidal set-value estimation, and result optimization output steps. Combined with observability analysis to ensure convergence, it achieves dynamic recursive pose estimation. The low-complexity pose estimation system and method based on an event camera provided by this invention eliminates spatiotemporal redundancy in data, balances estimation accuracy and computational complexity, and possesses advantages such as low power consumption and high real-time performance. It can stably output high-precision pose results in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of machine vision and neuromorphic perception, and in particular to a low-complexity pose estimation system and method based on an event camera. Background Technology

[0002] Traditional pose estimation systems rely on periodic sampling mechanisms, which suffer from core drawbacks such as high power consumption, poor real-time performance, and limited dynamic range, making it difficult to meet the application requirements of wheeled robots in complex scenarios such as high-speed movement and drastic changes in lighting. With the development of neuromorphic engineering, event cameras that simulate the working mechanism of biological vision systems have emerged, providing a new approach to overcome the bottlenecks of traditional pose estimation techniques.

[0003] However, the event-driven output characteristics of event cameras break the traditional periodic sampling data pattern, making it impossible to directly apply traditional pose estimation algorithms based on frame data, thus bringing new challenges to pose estimation technology: First, event-driven data modeling is difficult. The asynchronous pulse sequence output by event cameras contains information such as pixel coordinates, event timestamps, and polarity of light intensity changes, rather than traditional continuous image frames. The feature extraction and matching relied upon by traditional pose estimation algorithms are based on the synchronous processing of continuous image frames, which cannot directly adapt to the non-periodic and discrete characteristics of pulse events; Second, multi-sensor spatial redundancy processing is complex. Event cameras are usually equipped with a massive number of pixel units to expand the sensing range. First, the measurement data from adjacent cameras are often highly correlated, resulting in spatial redundancy. Traditional methods do not consider the processing of such redundant data, leading to an exponential increase in data transmission volume and computational complexity with the number of sensors, which restricts the real-time deployment of embedded platforms. Second, it is difficult to balance positioning accuracy and computational complexity. Existing pose estimation methods based on event cameras either rely on complex deep learning models, resulting in high computational load, or simplify the algorithm, resulting in insufficient positioning accuracy, which is difficult to meet the actual application needs of wheeled robots. Third, there is a lack of systematic theoretical support. Classical observability theory and optimal estimation theory are difficult to directly apply to systems with event-driven characteristics, and the performance of the algorithms is not guaranteed.

[0004] Therefore, developing a pose estimation method that adapts to the characteristics of event cameras and combines low computational complexity with high positioning accuracy to meet the real-time and low power consumption requirements of wheeled robot embedded systems has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a low-complexity pose estimation system and method based on an event camera, which solves the problems of computational complexity, low accuracy, poor stability, and difficulty in deployment on embedded platforms in the existing technology, and achieves robot pose estimation that balances low complexity, high accuracy and strong robustness.

[0006] To achieve the above objectives, the present invention provides a low-complexity pose estimation system based on an event camera, comprising: an event camera module, a spatiotemporal joint data selection module, a state space modeling module, an event-triggered state estimation module, and a pose estimation result post-processing module. Event camera module: As the system's perception front end, one or more event cameras are deployed on the wheeled robot according to a preset spatial structure. Each event camera independently collects information on changes in ambient light intensity, simulating the working mechanism of a biological retina, and generates event information when the perceived change in light intensity exceeds a threshold. Spatiotemporal joint data selection module: used to perform dual filtering of event data output by the event camera module in both the time and spatial domains to obtain event data that has both temporal validity and spatial independence; State space modeling module: Transforms the pose estimation problem into a dynamic system state estimation problem, defines the pose of the wheeled robot as the system state, establishes the state evolution equation based on the robot's kinematic model, and constructs the measurement equation by combining the event output characteristics of the event camera; Event-triggered state estimation module: Based on ellipsoidal set-valued estimation theory, a low-complexity estimation algorithm is designed. It achieves robot pose estimation through a two-step iterative process of prior state prediction and measurement constraint update, while simultaneously performing... - Observability analysis; Post-processing module for pose estimation results: Optimizes the estimation results, judges the validity of the estimation results based on the characteristics of the ellipsoid shape matrix, and outputs the results when the accuracy meets the requirements; when the accuracy is insufficient, it dynamically adjusts the camera scheduling strategy to increase the amount of effective data and improve the stability and reliability of the pose estimation results.

[0007] Preferably, the camera trigger threshold in the event camera module is adaptively adjusted according to the ambient lighting conditions to balance noise suppression and dynamic information capture rate.

[0008] Preferably, the dual filtering in the spatiotemporal joint data selection module specifically involves: retaining valid events in the time domain where the light intensity change reaches a threshold, and discarding invalid and redundant data; and filtering independent and valid data in the spatial domain based on the pixel coordinate distribution of the event data, and removing adjacent pixels or camera height-related information.

[0009] Preferably, the prior state prediction in the event-triggered state estimation module is to predict the current pose range using the estimation result of the previous moment and the state equation; the measurement constraint update is to combine the filtered valid event information output by the spatiotemporal joint data selection module, simplify the constraint update through ellipsoidal operation, and recursively obtain the pose estimation result.

[0010] Preferably, the post-processing module for pose estimation results optimizes the estimation by using a sliding window averaging method to smooth the estimated values ​​at consecutive time intervals and suppress fluctuations caused by random disturbances.

[0011] This invention also provides a low-complexity pose estimation method based on an event camera, comprising the following steps: S1. The event camera module collects information on changes in ambient light intensity, generates and outputs event data according to an adaptively adjusted trigger threshold; S2, the spatiotemporal joint data selection module processes the event data: in the time domain, valid events are filtered through interruption, a cache of historical valid events is maintained, signal-related information of the last valid event is stored, and invalid data is discarded; in the spatial domain, redundant information is removed based on measurement vector similarity to obtain a set of valid independent events. S3, the state space modeling module defines the pose of the wheeled robot as the system state, establishes the state evolution equation and measurement equation, and forms a mathematical model; S4. The event-triggered state estimation module performs prior state prediction and measurement constraint update, and obtains the robot pose estimation value through iterative calculation. S5. The pose estimation result post-processing module performs sliding window smoothing on the posterior ellipsoid center, performs validity verification based on the trace of the ellipsoid shape matrix, adjusts the sensor scheduling strategy according to the verification result, and outputs stable pose estimation results.

[0012] Preferably, the formula for spatial domain redundancy removal in step S2 is: in, For the first Time of the first Spatial scheduling variables for each sensor; It is an infinite norm distance metric; For the first The selected time A set of measurement vectors from a sensor; For the first Time of the first Measurement vectors of each sensor; This is the spatial similarity threshold; This is a time index, representing the current discrete moment; The sensor index represents the current sensor number to be determined. One sensor.

[0013] Preferably, the state evolution equation in step S3 is as follows: in, for System status at all times; for System status at all times; This is the state transition matrix; For process perturbations, satisfy bounded constraints , Indicates that the origin is the center. The ellipsoidal set is the shape matrix; the process perturbation considers the uncertainty of robot motion, and is achieved through reasonable settings. The parameters ensure the rationality of the disturbance modeling; The measurement equation, the formula is: in, The sensor's measured value; The measurement matrix is ​​determined by the spatial position of the sensors; To measure noise, bounded constraints must be satisfied. , Indicates that the origin is the center. For the ellipsoid set of shape matrices, The noise boundary parameter is defined; different measurement matrices correspond to sensors at different locations, clarifying the mapping relationship between event data and robot pose, and truly reflecting the measurement characteristics of the camera.

[0014] Preferably, the prior state prediction in step S4 is based on the state estimation ellipsoid of the previous time step, and the prediction is carried out using the state equation and ellipsoid affine transformation, Minkowski sum operation, etc., to complete the prediction of the prior state ellipsoid of the current time step. Specifically, it includes the following steps: S411. Determine the prediction basis and core computational methods: based on the state estimation ellipsoid of the previous time step, including the state estimation center of the previous time step. Shape matrix of the previous time step By utilizing the robot's kinematic state equations and combining ellipsoidal affine transformation, ellipsoidal Minkowski sum operation, the prediction calculation of the prior state ellipsoid at the current moment is carried out. S412. The prior state center is obtained by operating on the state transition matrix and the state estimation center from the previous time step, as shown in the formula: in, It is the a priori state center; The state estimation center of the previous time step; This is the state transition matrix; S413. Calculate the prior state shape matrix: The prior state shape matrix is ​​obtained by operating on the state transition matrix, the previous time step shape matrix, and the process perturbation constraint ellipsoid. The formula is as follows: in, The shape matrix represents the prior state. The shape matrix of the state at the previous time step; For the shape matrix operations of the Minkowski sum of the ellipsoid; State transition matrix transpose; The shape matrix of the process perturbation constraint ellipsoid; S414. Perform the Minkowski sum operation on the ellipsoid: The Minkowski sum operation on the ellipsoid involved in the formula follows these rules: For ellipsoids... Centered on The ellipsoid set of shape matrices and Centered on The ellipsoid set of shape matrices Its Minkowski and external approximations are based on Centered on The ellipsoid set of shape matrices ,in To optimize parameters; S415. Output prediction results: Through steps S411-S414, obtain the complete prior state ellipsoid at the current moment, including the prior state center. and the shape matrix of prior states This serves as the input basis for subsequent measurement constraint update steps; The measurement constraint update utilizes valid event measurement information output by the spatiotemporal joint data selection module to update the constraints of the prior ellipsoid, thereby obtaining the posterior state ellipsoid. The specific steps are as follows: S421. Constructing the set of effective measurements: For each effective sensor, the measured value is known, and the ellipsoid corresponding to the measurement constraint is... ; S422, Ellipsoidal Intersection Outside Approximation: Constructing an Auxiliary Matrix Auxiliary vectors By simplifying the calculation using the matrix inverse lemma, intermediate quantities are obtained. , The formula is: in, To integrate the intermediate matrix of prior state uncertainty and effective measurement constraints, To integrate the intermediate vector between the prior state center and the effective measurement value, The shape matrix of the prior states The inverse matrix, For measurement matrix transpose, For an effective sensor set; S423. Calculate the center and shape matrix of the posterior ellipsoid: The center of the posterior ellipsoid is calculated recursively using an external approximation through ellipsoid intersection operations. and shape matrix The formula is: in, For the intermediate matrix The inverse matrix; Let be the ellipsoidal scale parameter, satisfying ; S424, Optimize scale parameters By minimizing The upper bound yields the optimal parameters, ensuring the compactness of the estimated ellipsoid. The formula is: in, and To optimize parameters, As an auxiliary vector, For auxiliary matrix; S45, Verification - Observability: Calculate the observability matrix using the following formula: in, For a finite amount of time, To represent the time from the initial time 0 to the finite time... A set; when ,in, For the state dimension; the system satisfies -Observability; The pose estimation error will converge to ,in, , For matrix The Cholesky decomposition is calculated using the following formula: in, This represents the optimal observable information matrix.

[0015] Preferably, the ellipsoid center smoothing in step S5 specifically includes the following steps: S511. Determine the smoothing method: Use the sliding window averaging method to smooth the estimated ellipsoid center values ​​at consecutive time points; S512, Adaptive adjustment of window length: The length of the sliding window is adaptively adjusted according to the robot's movement speed; S513. Handling window initialization issues: The smoothing result in the initial stage directly uses the current estimated value, and the complete window averaging operation is performed after sufficient window data is available. S514. Store the estimated values ​​from the most recent multiple time periods using a circular buffer. Each time an update occurs, remove the oldest data, add the new data, and calculate the average. The specific formula is as follows: in, These are the smoothed state estimates. The length of the sliding window. For the first State estimate at time 1; Validation and parameter adjustment specifically include the following steps: S521. Define the validity judgment criterion: based on the trace of the ellipsoid shape matrix. Quantify the estimation accuracy and preset the accuracy threshold. The judgment criteria are: in, This indicates that the estimation result is valid. This indicates that the estimation result is invalid and the sensor scheduling strategy needs to be adjusted. S522. Perform the operation based on the verification result: when In the effective scenario, the smoothed pose estimation result is directly output. ; when In the case of an invalid scenario, the current measurement information is deemed insufficient, triggering an adjustment to the camera scheduling strategy, reducing the spatial similarity threshold, increasing the number of valid events, and improving the richness of measurement information. S523, Dynamic Recovery Threshold: When multiple consecutive time intervals... When the estimation accuracy meets the requirements, the spatial similarity threshold is restored to the initial value to balance the pose estimation accuracy and computational complexity. S524. Standardized output results: Pose estimation results are output through a standardized communication interface.

[0016] Therefore, the present invention employs the above-mentioned low-complexity pose estimation system and method based on an event camera, which has the following beneficial effects: (1) Pose recursion is realized based on ellipsoid set-value estimation, which has low computational cost and low complexity, and is suitable for real-time operation on embedded platforms; (2) Redundant information is filtered out by spatiotemporal joint data, which greatly reduces the amount of computation while ensuring accuracy; (3) Introduce smoothing, validity verification and adaptive scheduling mechanisms to improve the stability and robustness of estimation; (4) Based on - Observability ensures error convergence and reliable positioning accuracy, making it more suitable for real-world scenarios of wheeled robots.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a diagram showing the relationship between the modules in an embodiment of the present invention; Figure 2 This is a block diagram of a pose estimation system according to an embodiment of the present invention; Figure 3 This is a flowchart of pose estimation according to an embodiment of the present invention. Detailed Implementation

[0019] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] This implementation focuses on the construction and implementation of a low-complexity pose estimation system based on an event camera. The system is centered around an event camera module, a spatiotemporal joint data selection module, a state space modeling module, an event-triggered state estimation module, and a pose estimation result post-processing module. The interrelationships between these modules are shown in Figure 1, and the overall architecture is shown in Figure 2. The method relies on the completed system and strictly follows the execution flow shown in Figure 3: signal acquisition, signal preprocessing, feature extraction, pose estimation algorithm processing, and output of the robot pose result. It is based on ellipsoidal set-valued estimation theory, combined with... - Observability analysis enables low-power, high-precision pose estimation for wheeled robots. The following sections elaborate on the system setup and methodology.

[0021] I. Construction of a Low-Complexity Pose Estimation System Based on Event Cameras This system is a complete hardware and software system adapted to the embedded platform of wheeled robots. It is built hierarchically according to the following levels: perception front-end, data preprocessing, mathematical modeling, core calculation, and result optimization output. Each module is functionally independent yet works collaboratively. The hardware relies on the embedded main control board and event camera sensors of the wheeled robot. The software implements the algorithm logic through embedded programming. Modules interact through standardized communication interfaces, and closed-loop feedback signals enable dynamic parameter adjustment, achieving a complete match. Figure 1 Module association and Figure 2 The system architecture requirements.

[0022] (I) Hardware infrastructure setup Perception front-end hardware: Select one or more event cameras and deploy them on the wheeled robot according to a preset spatial structure. The cameras are equipped with standardized communication interfaces for event output. The camera trigger threshold can be adaptively adjusted according to ambient lighting conditions to achieve low power consumption and high dynamic range visual signal acquisition.

[0023] Core processing hardware: The wheeled robot uses an embedded microcontroller and main control board as the hardware carrier for the spatiotemporal joint data selection module, state space modeling module, event-triggered state estimation module, and pose estimation result post-processing module. It maintains a historical valid event buffer for each sensor, sets up an event buffer queue to temporarily store valid event data, and uses an optimized linear algebra library to implement matrix operations, supporting core mathematical operations such as matrix multiplication, inversion, trace calculation, and square root operation.

[0024] Interactive output hardware: A standardized communication interface is reserved on the embedded main control board for outputting pose estimation results. It also supports dynamic adjustment of camera scheduling strategies and various threshold parameters to realize data interaction with the main controller of the wheeled robot.

[0025] (II) Construction of Five Core Functional Modules Combination Figure 1 The module's input / output and coordination logic are implemented to ensure hardware and software compatibility and functionality of each module. This ensures that the output of one module becomes the core input of the next. The pose estimation result post-processing module can feed back parameter adjustment signals to the spatiotemporal joint data selection module. The specific setup is as follows: Event Camera Module: As the system's perception front end, the hardware consists of one or more event cameras deployed on the wheeled robot according to a preset spatial structure, and the software implements an adaptive adjustment strategy for the camera trigger threshold. The core function of the module is to simulate the working mechanism of the biological retina, generating event information only when the perceived light intensity change exceeds the threshold, achieving low power consumption and high dynamic range visual signal acquisition, and providing the system with raw event data.

[0026] Spatiotemporal Joint Data Selection Module: The hardware carrier is an embedded microcontroller of the wheeled robot, and the software is implemented in collaboration with the main program through the microcontroller's interrupt service routine. The core function of the module is to design a joint filtering mechanism for the massive event data output by the event camera. In the time domain, only valid events with light intensity changes reaching a threshold are retained, and invalid and redundant data are discarded. In the spatial domain, independent and valid data are filtered based on the pixel coordinate distribution of the event data, and adjacent pixels or camera height-related information are removed. Only event data that has both temporal validity and spatial independence is retained, which significantly reduces the complexity of subsequent data transmission and processing.

[0027] State space modeling module: The hardware relies on the embedded main control board computing unit of the wheeled robot, and the software is implemented through matrix operation functions combined with an optimized linear algebra library. The core function of the module is to transform the pose estimation problem into a dynamic system state estimation problem. The pose (position and attitude) of the wheeled robot is defined as the system state. The state evolution equation is established based on the robot's kinematic model. The measurement equation is constructed by combining the event output characteristics of the event camera. The mathematical relationship between event data and robot pose is clarified, providing rigorous mathematical model support for subsequent state estimation.

[0028] Event-triggered state estimation module: As the core solution module of the system, the hardware relies on the embedded main control board computing unit of the wheeled robot, and the software is designed and implemented based on ellipsoidal set-valued estimation theory. The core function of the module is to achieve robot pose estimation through two iterative steps of prior state prediction and measurement constraint update. First, the estimation result of the previous moment and the state equation are used to predict the current pose range. Then, the constraint is updated by simplifying the ellipsoidal operation by combining the filtered valid event information, and the pose estimation result is obtained recursively. At the same time, based on... - Observability analysis ensures algorithm convergence, guarantees pose estimation accuracy, and outputs the posterior ellipsoid of pose estimation (including center and shape matrices).

[0029] Post-processing module for pose estimation results: The hardware relies on the embedded main control board of the wheeled robot, and the software implements the sliding window averaging method and the validity verification logic of the estimation results based on the characteristics of the ellipsoidal shape matrix. The core function of the module is to optimize the estimation results output by the event-triggered state estimation module. It uses the sliding window averaging method to smooth the estimated values ​​at consecutive time points and suppress fluctuations caused by random disturbances. It judges the validity of the estimation results based on the characteristics of the ellipsoidal shape matrix. When the accuracy meets the requirements, the results are output. When the accuracy is insufficient, the camera scheduling strategy is dynamically adjusted and fed back to the spatiotemporal joint data selection module to increase the amount of effective data and improve the stability and reliability of the pose estimation results.

[0030] After each module is built, the overall integration is completed to ensure smooth data flow between modules. Closed-loop feedback signals can be accurately transmitted and parameters can be dynamically adjusted. The overall system architecture is shown in Figure 2. It can adapt to the resource-constrained characteristics of the wheeled robot embedded platform and meet the application requirements of low power consumption and high real-time performance.

[0031] II. Implementation of a Low-Complexity Pose Estimation Method Based on Event Cameras This method relies on the system built above. The overall execution flow is shown in Figure 3. It sequentially completes the entire process of signal acquisition, signal preprocessing, feature extraction, pose estimation algorithm processing, and pose result output. It makes full use of the event-driven characteristics of the event camera and combines a low-complexity ellipsoidal set-valued estimation algorithm to achieve a balance between pose estimation accuracy and computational efficiency. The specific implementation steps are as follows: S1, Event Camera Signal Acquisition The system's event camera module completes the acquisition and event-based output of environmental visual signals. The specific steps are as follows: Initialize the operating parameters of the event camera module, configure the communication protocol of the event output interface, and ensure normal data transmission; Activate the camera's adaptive trigger threshold adjustment function, periodically calculate the event occurrence rate per unit time, and dynamically adjust the trigger threshold based on the deviation of the occurrence rate from the preset range. Balancing noise suppression with dynamic information capture rate; The camera senses changes in ambient light intensity in real time and uses a mathematical model of the event triggering mechanism to determine whether to generate a pulse event. Event data is generated and output only when the change in light intensity exceeds the current threshold. No data is output when the triggering condition is not met, thus achieving low-power visual signal acquisition.

[0032] S2, Spatiotemporal Joint Data Preprocessing The system's spatiotemporal joint data selection module filters the raw event data, removing noise and interference and enhancing the effective signal. This process involves two steps: time-domain filtering and spatial-domain scheduling. Time-domain event triggering and filtering: The microcontroller receives event data from each event camera via interrupt, and parses it to obtain key information such as sensor number, event polarity, and timestamp; it maintains a historical valid event buffer for each sensor, storing signal-related information of the last valid event, and judges the validity of the current event according to preset trigger conditions. Events that meet the conditions are marked as valid and the historical buffer is updated, and they are added to the event buffer queue. Events that do not meet the conditions are discarded directly. Spatial domain sensor scheduling: Periodically perform spatial redundancy removal on valid events filtered in the time domain. Construct measurement vectors from historical event data of the sensors within the most recent time window, initialize the set of selected sensors and the set of measurement vectors, and add the first valid event sensor to the set. For subsequent sensors, calculate the infinity norm distance similarity between their measurement vectors and all vectors in the set. Those with similarity higher than a preset threshold are retained and added to the set, while those with similarity lower than the threshold are marked as spatial redundancy and discarded. Finally, output event measurement data that combines temporal validity and spatial independence.

[0033] The formula for eliminating redundancy in the spatial domain is: in, For the first Time of the first Spatial scheduling variables for each sensor; It is an infinite norm distance metric; For the first The selected time A set of measurement vectors from a sensor; For the first Time of the first Measurement vectors of each sensor; This is the spatial similarity threshold; This is a time index, representing the current discrete moment; The sensor index represents the current sensor number to be determined. One sensor.

[0034] S3, State-space modeling The pose estimation problem is standardized through the system's state-space modeling module, a mathematical mapping relationship between event data and robot pose is constructed, and the extraction and quantification of key features of visual signals are completed. The specific steps are as follows: Establish the state evolution equation: Define the pose (position and orientation) of the wheeled robot as the system state. State equations are established based on the robot's kinematic model: in, for System status at all times; for System status at all times; This is the state transition matrix; For process perturbations, satisfy bounded constraints , Indicates that the origin is the center. The ellipsoidal set is the shape matrix; the process perturbation considers the uncertainty of robot motion, and is achieved through reasonable settings. The parameters ensure the rationality of the disturbance modeling; Establish measurement equations: Construct measurement equations by combining the event output characteristics of the event camera: in, The sensor's measured value; The measurement matrix is ​​determined by the spatial position of the sensors; To measure noise, bounded constraints must be satisfied. , Indicates that the origin is the center. For the ellipsoid set of shape matrices, The noise boundary parameter is defined; different measurement matrices correspond to sensors at different locations, clarifying the mapping relationship between event data and robot pose, and truly reflecting the measurement characteristics of the camera.

[0035] Matrix operations are performed using an optimized linear algebra library to ensure the real-time nature of the modeling process and to output standardized state and measurement equations, providing a mathematical model for subsequent pose core calculations.

[0036] S4, Event-triggered pose core estimation The system's event-triggered state estimation module performs dynamic recursive estimation of the robot's pose based on ellipsoidal set-valued estimation theory, combined with... - Observability analysis ensures the convergence of the algorithm and is the core step of the method, specifically divided into two steps: prior state prediction and measurement constraint update. Prior state prediction: Based on the state estimation ellipsoid of the previous time step, including the state estimation center of the previous time step. Shape matrix of the previous time step Using the equation of state, affine transformations of the ellipsoid, and Minkowski sum operations, the prior state center is calculated using the following formula: in, It is the a priori state center; The state estimation center of the previous time step; This is the state transition matrix; The formula for calculating the prior state shape matrix is: in, The shape matrix represents the prior state. The shape matrix of the state at the previous time step; For the shape matrix operations of the Minkowski sum of the ellipsoid; State transition matrix transpose; The shape matrix of the process perturbation constraint ellipsoid; Performing the Minkowski sum operation on the ellipsoid: The Minkowski sum operation on the ellipsoid involved in the formula follows these rules: For ellipsoids... Centered on The ellipsoid set of shape matrices and Centered on The ellipsoid set of shape matrices Its Minkowski and external approximations are based on Centered on The ellipsoid set of shape matrices ,in To optimize parameters; To predict the robot's pose range at the current moment; Measurement constraint update: Using the valid event measurement information output by the spatiotemporal joint data selection module, the prior ellipsoid is updated with constraints to obtain the posterior state ellipsoid. The specific steps are as follows: 1. Constructing a set of effective measurements: For each effective sensor, the measured value is known, and the ellipsoid corresponding to the measurement constraint is... ; 2. Ellipsoidal intersection outside approximation: Constructing an auxiliary matrix Auxiliary vectors By simplifying the calculation using the matrix inverse lemma, intermediate quantities are obtained. , The formula is: in, To integrate the intermediate matrix of prior state uncertainty and effective measurement constraints, To integrate the intermediate vector between the prior state center and the effective measurement value, The shape matrix of the prior states The inverse matrix, For measurement matrix transpose, For an effective sensor set; 3. Calculate the center and shape matrix of the posterior ellipsoid: Through external approximation simplification using ellipsoid intersection operations, recursively calculate the center of the posterior ellipsoid. and shape matrix The formula is: in, For the intermediate matrix The inverse matrix; Let be the ellipsoidal scale parameter, satisfying ; 4. Optimize scale parameters By minimizing The upper bound yields the optimal parameters, ensuring the compactness of the estimated ellipsoid. The formula is: in, and To optimize parameters, As an auxiliary vector, For auxiliary matrix; 5. Verification - Observability: Calculate the observability matrix using the following formula: in, For a finite amount of time, To represent the time from the initial time 0 to the finite time... A set; when ,in, For the state dimension; the system satisfies -Observability; The pose estimation error will converge to ,in, , For matrix The Cholesky decomposition is calculated using the following formula: in, This represents the optimal observable information matrix.

[0037] S5. Post-processing and output of pose estimation results The system's pose estimation result post-processing module optimizes the core estimation results, performs validity verification, and outputs the final pose result, simultaneously displaying the pose and confidence level. The specific steps are as follows: Ellipsoid center smoothing: The posterior ellipsoid center estimates at consecutive time steps are smoothed using a sliding window averaging method. The sliding window length is adaptively adjusted according to the robot's motion speed. The smoothed state estimate is calculated using the following formula: in, These are the smoothed state estimates. The length of the sliding window. For the first State estimate at time 1; Suppress estimation fluctuations caused by random disturbances; when data is insufficient during the window initialization phase, directly use the current estimate to ensure the continuity of smoothing processing; Validation of estimation results: The validity of the estimation results is judged based on the trace operation of the ellipsoid shape matrix, with a preset accuracy threshold. ,like If the label estimation result is valid, then the label estimation result is valid. If so, the label estimation result is invalid; Output and Parameter Feedback: When the estimation result is valid, the smoothed pose estimation result and confidence level are output through the standardized communication interface, and the pose and confidence level are displayed on the terminal, realizing the display of pose and confidence level in Figure 3; when the estimation result is invalid, the spatial similarity threshold of the spatiotemporal joint data selection module is dynamically adjusted to reduce the threshold to increase the number of valid events and improve the richness of measurement information; if the estimation accuracy meets the requirements for multiple consecutive time points, the spatial similarity threshold is restored to the initial value to balance the pose estimation accuracy and computational complexity.

[0038] This method, through the steps described above, achieves dynamic recursive estimation of the pose of a wheeled robot based on the constructed system. The algorithm complexity is linearly related to the amount of event data. It fully utilizes the low power consumption, high dynamic range, and low latency characteristics of the event camera, significantly reducing hardware resource consumption while ensuring pose estimation accuracy. It is fully adaptable to the deployment requirements of resource-constrained embedded systems for wheeled robots and can maintain stable pose estimation performance in complex scenarios such as high-speed movement and drastic changes in lighting.

[0039] Therefore, this invention employs the aforementioned low-complexity pose estimation system and method based on an event camera, relying on the collaborative work of five core modules, and combining ellipsoidal set-valued estimation theory with... - Observability analysis fully leverages the advantages of event camera technology to achieve spatiotemporal redundancy removal and low-complexity pose recursive estimation of event data, balancing estimation accuracy and computational efficiency. It is compatible with wheeled robot embedded systems and can stably, in real time, and with high accuracy complete pose estimation in complex scenarios, providing precise state support for robot navigation, localization, and other tasks, and solving many bottleneck problems of traditional pose estimation technology.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A low-complexity pose estimation system based on an event camera, characterized in that, include: Event camera module, spatiotemporal joint data selection module, state space modeling module, event-triggered state estimation module, pose estimation result post-processing module; Event camera module: As the system's perception front end, one or more event cameras are deployed on the wheeled robot according to a preset spatial structure. Each event camera independently collects information on changes in ambient light intensity, simulating the working mechanism of a biological retina, and generates event information when the perceived change in light intensity exceeds a threshold. Spatiotemporal joint data selection module: used to perform dual filtering of event data output by the event camera module in both the time and spatial domains to obtain event data that has both temporal validity and spatial independence; State space modeling module: Transforms the pose estimation problem into a dynamic system state estimation problem, defines the pose of the wheeled robot as the system state, establishes the state evolution equation based on the robot's kinematic model, and constructs the measurement equation by combining the event output characteristics of the event camera; Event-triggered state estimation module: Based on ellipsoidal set-valued estimation theory, a low-complexity estimation algorithm is designed. It achieves robot pose estimation through a two-step iterative process of prior state prediction and measurement constraint update, while simultaneously performing... - Observability analysis; Post-processing module for pose estimation results: Optimizes the estimation results, judges the validity of the estimation results based on the characteristics of the ellipsoid shape matrix, and outputs the results when the accuracy meets the requirements; when the accuracy is insufficient, it dynamically adjusts the camera scheduling strategy to increase the amount of effective data and improve the stability and reliability of the pose estimation results.

2. The low-complexity pose estimation system based on an event camera according to claim 1, characterized in that, The camera trigger threshold in the event camera module is adaptively adjusted according to ambient lighting conditions to balance noise suppression and dynamic information capture rate.

3. The low-complexity pose estimation system based on an event camera according to claim 1, characterized in that, The dual filtering in the spatiotemporal joint data selection module is as follows: in the time domain, valid events with light intensity changes reaching a threshold are retained, while invalid and redundant data are discarded; in the spatial domain, independent and valid data are filtered based on the pixel coordinate distribution of the event data, while adjacent pixels or camera height-related information are removed.

4. The low-complexity pose estimation system based on an event camera according to claim 1, characterized in that, The prior state prediction in the event-triggered state estimation module uses the estimation result of the previous moment and the state equation to predict the current pose range; the measurement constraint update combines the filtered valid event information output by the spatiotemporal joint data selection module, and simplifies the constraint update through ellipsoidal operation to recursively obtain the pose estimation result.

5. The low-complexity pose estimation system based on an event camera according to claim 1, characterized in that, The post-processing module for pose estimation results optimizes the estimation by using a sliding window averaging method to smooth the estimated values ​​at consecutive time points and suppress fluctuations caused by random disturbances.

6. A method for applying to a low-complexity pose estimation system based on an event camera according to any one of claims 1-5, characterized in that, Includes the following steps: S1. The event camera module collects information on changes in ambient light intensity, generates and outputs event data according to an adaptively adjusted trigger threshold; S2, the spatiotemporal joint data selection module processes event data: the time domain filters valid events through interruption, maintains a cache of historical valid events, stores signal-related information of the last valid event, and discards invalid data; The spatial domain removes redundant information based on the similarity of measurement vectors to obtain an effective set of independent events; S3, the state space modeling module defines the pose of the wheeled robot as the system state, establishes the state evolution equation and measurement equation, and forms a mathematical model; S4. The event-triggered state estimation module performs prior state prediction and measurement constraint update, and obtains the robot pose estimation value through iterative calculation. S5. The pose estimation result post-processing module performs sliding window smoothing on the posterior ellipsoid center, performs validity verification based on the trace of the ellipsoid shape matrix, adjusts the sensor scheduling strategy according to the verification result, and outputs stable pose estimation results.

7. The low-complexity pose estimation method based on an event camera according to claim 6, characterized in that, The formula for spatial domain redundancy removal in step S2 is: in, For the first Time of the first Spatial scheduling variables for each sensor; It is an infinite norm distance metric; For the first The selected time A set of measurement vectors from a sensor; For the first Time of the first Measurement vectors of each sensor; This is the spatial similarity threshold; This is a time index, representing the current discrete moment; The sensor index represents the current sensor number to be determined. One sensor.

8. The low-complexity pose estimation method based on an event camera according to claim 7, characterized in that, The state evolution equation in step S3 is as follows: in, for System status at all times; for System status at all times; This is the state transition matrix; For process perturbations, satisfy bounded constraints , Indicates that the origin is the center. The ellipsoid set is a shape matrix; The measurement equation, the formula is: in, The sensor's measured value; The measurement matrix is ​​determined by the spatial position of the sensors; To measure noise, bounded constraints must be satisfied. , Indicates that the origin is the center. For the ellipsoid set of shape matrices, This represents the noise boundary parameter.

9. The low-complexity pose estimation method based on an event camera according to claim 8, characterized in that, The prior state prediction in step S4 is based on the state estimation ellipsoid from the previous time step. It utilizes the state equation and ellipsoidal affine transformation, Minkowski summation, and other operations to perform the prediction, thus completing the prediction of the prior state ellipsoid at the current time step. Specifically, it includes the following steps: S411. Determine the prediction basis and core computational methods: based on the state estimation ellipsoid of the previous time step, including the state estimation center of the previous time step. Shape matrix of the previous time step By utilizing the robot's kinematic state equations and combining ellipsoidal affine transformation, ellipsoidal Minkowski sum operation, the prediction calculation of the prior state ellipsoid at the current moment is carried out. S412. The prior state center is obtained by operating on the state transition matrix and the state estimation center from the previous time step, as shown in the formula: in, It is the a priori state center; The state estimation center of the previous time step; This is the state transition matrix; S413. Calculate the prior state shape matrix: The prior state shape matrix is ​​obtained by operating on the state transition matrix, the previous time step shape matrix, and the process perturbation constraint ellipsoid. The formula is as follows: in, The shape matrix represents the prior state. The shape matrix of the state at the previous time step; For the shape matrix operations of the Minkowski sum of the ellipsoid; State transition matrix transpose; The shape matrix of the process perturbation constraint ellipsoid; S414. Perform the Minkowski sum operation on the ellipsoid: The Minkowski sum operation on the ellipsoid involved in the formula follows these rules: For ellipsoids... Centered on The ellipsoid set of shape matrices and Centered on The ellipsoid set of shape matrices Its Minkowski and external approximations are based on Centered on The ellipsoid set of shape matrices ,in To optimize parameters; S415. Output prediction results: Through steps S411-S414, obtain the complete prior state ellipsoid at the current moment, including the prior state center. and the shape matrix of prior states This serves as the input basis for subsequent measurement constraint update steps; The measurement constraint update utilizes valid event measurement information output by the spatiotemporal joint data selection module to update the constraints of the prior ellipsoid, thereby obtaining the posterior state ellipsoid. The specific steps are as follows: S421. Constructing the set of effective measurements: For each effective sensor, the measured value is known, and the ellipsoid corresponding to the measurement constraint is... ; S422, Ellipsoidal Intersection Outside Approximation: Constructing an Auxiliary Matrix Auxiliary vectors By simplifying the calculation using the matrix inverse lemma, intermediate quantities are obtained. , The formula is: in, To integrate the intermediate matrix of prior state uncertainty and effective measurement constraints, To integrate the intermediate vector between the prior state center and the effective measurement value, The shape matrix of the prior states The inverse matrix, For measurement matrix transpose, For an effective sensor set; S423. Calculate the center and shape matrix of the posterior ellipsoid: The center of the posterior ellipsoid is calculated recursively using an external approximation through ellipsoid intersection operations. and shape matrix The formula is: in, For the intermediate matrix The inverse matrix; Let be the ellipsoidal scale parameter, satisfying ; S424, Optimize scale parameters By minimizing The upper bound yields the optimal parameters, as shown in the formula: in, and To optimize parameters, As an auxiliary vector, For auxiliary matrix; S425, Verification - Observability: Calculate the observability matrix using the following formula: in, For a finite amount of time, To represent the time from the initial time 0 to the finite time... A set; when ,in, For the state dimension; the system satisfies -Observability; The pose estimation error will converge to ,in, , For matrix The Cholesky decomposition is calculated using the following formula: in, This represents the optimal observable information matrix.

10. A low-complexity pose estimation method based on an event camera according to claim 9, characterized in that, The ellipsoid center smoothing in step S5 specifically includes the following steps: S511. Determine the smoothing method: Use the sliding window averaging method to smooth the estimated ellipsoid center values ​​at consecutive time points; S512, Adaptive adjustment of window length: The length of the sliding window is adaptively adjusted according to the robot's movement speed; S513. Handling window initialization issues: The smoothing result in the initial stage directly uses the current estimated value, and the complete window averaging operation is performed after sufficient window data is available. S514. Store the estimated values ​​from the most recent multiple time periods using a circular buffer. Each time an update occurs, remove the oldest data, add the new data, and calculate the average. The specific formula is as follows: in, These are the smoothed state estimates. The length of the sliding window. For the first State estimate at time 1; Validation and parameter adjustment specifically include the following steps: S521. Define the validity judgment criterion: based on the trace of the ellipsoid shape matrix. Quantify the estimation accuracy and preset the accuracy threshold. The judgment criteria are: in, This indicates that the estimation result is valid. This indicates that the estimation result is invalid and the sensor scheduling strategy needs to be adjusted. S522. Perform the operation based on the verification result: when In the effective scenario, the smoothed pose estimation result is directly output. ; when In the case of an invalid scenario, the current measurement information is deemed insufficient, triggering an adjustment to the camera scheduling strategy, reducing the spatial similarity threshold, increasing the number of valid events, and improving the richness of measurement information. S523, Dynamic Recovery Threshold: When multiple consecutive time intervals... When the estimation accuracy meets the requirements, the spatial similarity threshold is restored to the initial value to balance the pose estimation accuracy and computational complexity. S524. Standardized output results: Pose estimation results are output through a standardized communication interface.

Images