Research On 3D Reconstruction Method Of Indoor Scene Based On RGB-D Camera

With the development of artificial intelligence technology in recent years,in various complex industrial applications,robots have been assisting humans to perform some simple tasks,and more and more have begun to turn to some complex tasks.In order to solve the problem of positioning and mapping of industrial robots in complex unknown scenes,especially in low-light scenes,this paper applies Vision-SLAM technology to 3D reconstruction of indoor scenes by RGB-D camera.An Improved Gamma Correction is proposed to optimize the pose estimation based on feature point method in low-light scenes,and the SLAM algorithm flow is disassembled into pose estimation,back-end optimization and loopback detection to improve and optimize.It is verified that the proposed algorithm performs well in low-light scenes and achieves a good three-dimensional reconstruction effect for indoor scenes.The specific contents are as follows:First of all,according to the research content of the subject,we chose to study the RGB-D camera with higher reconstruction efficiency in the indoor scene,and finally selected Microsoft's Kinect V2 camera as our experimental equipment through parameter comparison.In order to improve the accuracy of camera acquisition of image information,we recalibrate the internal parameters and distortion coefficients and the registration of color and depth cameras based on Zhang Zhengyou's checkerboard calibration method.Then,the reconstruction scheme based on RGB-D SLAM is designed and improved.Firstly,the mainstream feature extraction and matching algorithm is compared and analyzed,and the pose estimation based on feature point method is completed.Then,in order to eliminate the cumulative error,some key points are selected and the pose map is optimized;in addition,the loopback detection methods based on random keyframes and wordbag model are compared and analyzed.Subsequently,an adaptive gamma correction algorithm combining image average gray value and empirical value is proposed,which firstly obtains the first gray average of the image captured by the camera in low-light scene,and adjusts the gamma correction accordingly.After comparison,the algorithm has better performance in different low-light environments compared to several classical image detail enhancement algorithms,both in terms of accuracy of feature matching and applicability of visual reconstruction algorithms.Finally,a three-dimensional reconstruction experiment for typical indoor scenes is designed.The ORB features and FLANN matching methods are selected for preliminary motion estimation,and the camera is loop-detected based on the word bag model.It is verified that the algorithms in this paper ar e effective in practical applications,and finally complete the 3D reconstruction of indoor scenes based on RGB-D camera.