A recurrent deep architecture for enhancing indoor camera localization using motion blur elimination

Abstract

Rapid growth and technological improvements in computer vision have enabled indoor camera localization. The accurate camera localization of an indoor environment is challenging because it has many complex problems, and motion blur is one of them. Motion blur introduces significant errors, degrades the image quality, and affects feature matching, making it challenging to determine camera pose accurately. Improving the camera localization accuracy for some robotic applications is still necessary. In this study, we propose a recurrent neural network (RNN) approach to solve the indoor camera localization problem using motion blur reduction. Motion blur in an image is detected by analyzing its frequency spectrum. A low-frequency component indicates motion blur, and by investigating the direction of these low-frequency components, the location and amount of blur are estimated. Then, Wiener filtering deconvolution removes the blur and obtains a clear copy of the original image. The performance of the proposed approach is evaluated by comparing the original and blurred images using the peak signal-to-noise ratio (PSNR) and structural similarity index(SSIM). After that, the camera pose is estimated using recurrent neural architecture from deblurred images or videos. The average camera pose error obtained through our approach is (0.16m, 5.61◦). In two recent research, Deep Attention and CGAPoseNet, the average pose error is (19m, 6.25◦) and (0.27m, 9.39◦), respectively. The results obtained through the proposed approach improve the current research results. As a result, some applications of indoor camera localization, such as mobile robots and guide robots, will work more accurately.