Research On State Estimation Of Ground Robot Based On Multi-sensor Fusion

In recent years,with the variety of sensors and the declining cost,multi-sensor fusion technology has been extensively researched and developed,and researchers are using multisensor fusion technology to complete the state estimation tasks of various intelligent robots.At present,researchers have utilized cameras,inertial measurement units(IMU),lidar and other sensors to obtain better state estimation accuracy for robot trajectory navigation in some scenarios.But for robots moving on a ground,the scale contained on IMU will drift slowly,guaranteeing accuracy and robustness is still a very critical and challenging task.This paper proposed a tightly-coupled robot state estimation algorithm that combines stereo vision,IMU and wheel odometer sensors,including initialization,measurement preprocessing,pre-integration correction,calibration of internal and external parameters of the camera and wheel odometer,and multi-sensor Joint optimization and other modules.This study uses vision to correct IMU noise and reduce the data noise of low-performance IMUs.For ground robots,the addition of wheel odometer sensors can solve the problem of scale drift and improve the robustness of the system.The core contribution of the entire study can be divided into three parts.The first is to correct the IMU pre-integration value by visual information and reduce the noise caused by random walk of IMU and discretization of continuous integration;the second is to realize the calibration procedure of internal and external parameters of camera and wheel odometer;the last is the research of multi-sensor data fusion method,realized a simple and effective fusion method of wheel odometer.In this paper,the internal and external calibration procedures of the wheel odometer are implemented.The calibration parameters have very good consistency,and the error fluctuation is 2.5%.This calibration method has been open sourced.We collected a representative data set in a large factory environment,The state estimation accuracy is improved by about 50% compared to the best systems of the same type.It is worth mentioning that the data set contains a number of unstable factors,such as dark light,reflective light,linear motion,dynamic objects,etc.The system can still run efficiently,reflecting the better robustness of our system.Multi-sensor fusion state estimation is a research field that attaches great importance to applications.The system implemented in this paper can be widely used in positioning and navigation tasks of restaurant service robots,warehouse logistics robots,sweeping robots,and autonomous vehicles.