Research On Visual Odometry Based On Fusion Of ORB Keypoints And BEBLID Descriptor

Along with the continuous development and popularity of artificial intelligence technology,various intelligent unmanned operation devices represented by mobile robots have gradually entered into thousands of households,bringing great convenience to people’s daily lives.For mobile robots,one of the basic functions is to be able to sense their own location and the content of the surrounding environment in a complex environment.Simultaneous Localization and Mapping(SLAM)is an important tool to achieve this function,which refers to the technology that locates the robot’s own position and builds a map of its surroundings at any time during its movement.Visual odometry(VO),as the front-end of SLAM system,undertakes important tasks such as feature point matching and pose estimation,and is the core part to realize localization and map building.Therefore,the study of visual odometry system is of crucial importance for SLAM.Under dim conditions,the scene content becomes blurred,which results in the degradation of image texture clarity and affects the feature point extraction work,which may eventually lead to the failure of robot pose estimation.Meanwhile,the quality of feature points is also important for visual odometry,and an excellent feature point should have the advantages of high differentiation,good repeatability and high detection efficiency.Therefore,it is of great importance to reduce the influence of low-light scenes and improve the performance of feature points.To address the problem of low-light image enhancement,this study applies Retinex and FGS(Fast Global Image Smoothing)theories to improve the effect of dim scenes on visual odometry.Specifically,Retinex theory divides the image observed by human eye into ambient light component and the reflection component of the object to the ambient light,and the purpose of low-light image enhancement is to recover the reflection component,which is also the most original appearance of the object.The ambient illumination component is first obtained from the image observed by the human eye using the FGS filtering scheme.Subsequently,the reflection component is recovered from the illumination component using Retinex theory.Finally,the reflection component is denoised because the noise is also enhanced during the image enhancement process.Also,the contrast of the reflection component is improved to enhance the clarity of texture detail information.To address the problem of low quality of feature points,this study uses a feature point detection scheme with the fusion of ORB key-points and BEBLID descriptors to improve the performance of image feature points.ORB key-points have the advantages of fast operation speed and simple principle,which are ideal for scenarios with high real-time requirements and limited equipment performance.By using multi-resolution image construction and grayscale prime center method,we solve the problems such as key-points sensitive to scale and direction,and improve the key-point robustness.The use of quadtree method avoids keypoints to overly gather inside the region,reduces the redundancy of key-points,and helps to improve the quality of key-points.As a newly emerged descriptor calculation method,BEBLID uses the Ada Boost algorithm to select a combination of strong classifiers and calculate the average gray value difference between them,and finally generates a string of binary descriptors.BEBLID can both maintain the matching rate of feature points and reduce the computational time consumption.Finally,a dataset of real scenes is used for testing.The experimental results show that the low-light image enhancement method we use can take into account both global and local texture information of the image,and can better recover the image content of dim scenes.the feature point detection scheme with the fusion of ORB key-points and BEBLID descriptors reduces the time loss by about 85% while maintaining the matching accuracy.The experimental data all verify the effectiveness of the low-light image enhancement and feature point detection schemes used in this study,which can significantly improve the localization and map building capabilities of mobile robots in low-light environments,laying an important foundation for the further development and popularization of robotics.