Research And Implementation Of SLAM Localization System In Indoor Low-light Condition

SLAM(Simultaneous Localization And Mapping)is considered as an essential basis for autonomous robots and driverless vehicles in detecting the surrounding environment and reconstructing the structure of an unknown environment when estimating the pose(including both location and orientation).Detection and construction using visual sensors is another hot topic in robot navigation.Visual SLAM applies visual sensors to collect information and estimate the motion of the sensor in an unknown environment.Therefore,it relies heavily on the environment’s characteristics.The quality of the images has a great effect on the efficiency.In a low light intensity environment,the performance of motion estimation declines seriously,even with lost tracking.To solve the problem,this thesis mainly studies the low illumination enhancement algorithms as an optimization module in visual SLAM system.The main work and innovation of this thesis are as follows:1.This thesis presents a tone mapping algorithm of FGA(Fast Global Adaption)to compensate the light condition and applies this solution to motion estimation in ORB-SLAM as an optimization module.Key of the solution is to obtain global adaptive brightness gain for image intensifiers.In this thesis,we optimize the luminance mean function and global adaptive function.The performance of the image intensifier is also evaluated in real datasets as well as the proof of higher estimation accuracy.The result of the optimized ORB-SLAM shows that there is a significant performance boost especially in low luminance environment.2.Tone mapping comes with inevitable noise amplification problems.To solve this situation,this thesis use Generative Adversarial Networks(GANs)to enhance low-illumination images with the support of Retinex theory.In this scheme,this thesis defines the image quality loss functions and image brightness loss functions using Single Scale Retinex(SSR),adjusts and optimizes the structure and parameters of the adversarial generation network and completes the training and testing of the network under the dark light data set.RetinexNet and Apple Photo Enhancer are used as control experiments to compare the model performance using Peak Signal-to-Noise Ratio(PSNR)and Structural SIMilarity(SSIM)as evaluation metrics,and the experimental results prove that the solution proposed in this thesis has better performance.According to the results of model analysis,an optimization of the localization accuracy of visual SLAM for low-illumination environments is accomplished on real datasets by fusing the proposed network model with ORB-SLAM.After analyzing the experimental results,it is concluded that the scheme can effectively improve the performance of the visual SLAM system in the position estimation and trajectory estimation in the dark light environment.3.With a combination of dark-light enhancement module,this thesis solves the problem of image quality acquired by vision sensors due to indoor low-light environment.On this basis,a design and implementation of a visual SLAM localization system for indoor dark-light environment is completed,which increases localization accuracy and scene applicability.This thesis presents a technical selection,performance testing and error analysis of the system on an indoor dataset.The results show that the system has high indoor positioning accuracy and can effectively cope with the challenges of low-illumination environment.