Research On SLAM Based On Low-cost Visual/Inertial Sensors

Precise localization in space is a basic function required by technologies such as mobile robots,micro aerial vehicles,virtual reality,and augmented reality.Visual Simultaneous Localization and Mapping(SLAM)and visual-inertial SLAM using cameras and inertial measurement units(IMUs)can perform localization based on the carrier's sensors,and therefore have been widely studied and applied.In the consumer-level field,the cost is a very important consideration.The most widely used low-cost camera is the rolling shutter camera.The rolling shutter camera will bring the rolling shutter effect when the camera moves in the environment,and the low-cost image sensors and inertial sensors are more prone to high image noise and IMU noise.The rolling shutter effect,image noise and IMU noise will all decrease the accuracy and robustness of the SLAM system,but the current research on these aspects is still insufficient.Therefore,this paper will conduct a more detailed study on rolling shutter effect,image noise and IMU noise.Therefore,this paper first aims at the problem that the current visual SLAM dataset cannot evaluate the rolling shutter effect of different levels,and an open-source dataset containing global shutter RGB-D images and rolling shutter RGB-D images is introduced.The dataset contains two trajectories,and each trajectory contains image sequences at three different speeds: slow,medium,and fast.Based on the ground truth,the validity of the dataset is verified through experiments,and the problem that the current visual SLAM dataset cannot evaluate the rolling shutter effect of different levels is solved.Secondly,because of the problem that there is only a single level of IMU noise and a single frequency of IMU measurements in the current visual-inertial SLAM dataset,an IMU data generation method based on the camera pose is introduced.And the IMU data with different noise levels and frame rates were obtained by modeling of the IMU noise.Based on the ground truth,the validity of the dataset is verified through experiments,and the problem of the noise and single frame rate of the IMU in the current visual-inertial SLAM evaluation dataset is solved.Thirdly,because of the problem of ignoring image noise in the current featurebased visual SLAM,a dataset specially used to evaluate the noise in visual SLAM was created based on the previous visual dataset.Then FFDNet was used to denoise the image sequence,and the matching effect of commonly used features in denoising and non-denoising conditions is quantitatively studied.According to the results,a method of improving ORB-SLAM2 by denoising is proposed.Experiments on simulation and real datasets show that this method improves the accuracy and robustness of ORB-SLAM2.Finally,since most of the current visual-inertial SLAM ignores the rolling shutter effect and the monocular visual-inertial SLAM initialization requires specific conditions,VINS-RSD is proposed.The system uses the RGB-D image combined with the IMU measurements to initialize the system,and the problem that the monocular visual-inertial SLAM initialization requires acceleration excitation is solved.The rolling shutter effect is corrected,and a composite loss function is used to improve the visual-inertial SLAM accuracy degradation in the case of high-noise IMU.The results of experiments on the dataset show that the RMSE of VINS-RSD is decreased by an average of 30.76% compared with VINS-Mono.The dataset can be used to evaluate visual/visual-inertial SLAM algorithms using rolling shutter cameras or global shutter cameras.The methods for processing the rolling shutter effect,image noise,and IMU noise explored in this paper have reference value and application prospects for the SLAM fields using low-cost cameras and IMUs.