Spacecraft Pose Estimation Using a Monocular Camera

Abstract

Spacecraft pose estimation is an essential input to the Guidance, Navigation and Control process. Pose estimation based on monocular camera images does not require modifications to the target vehicle and has fewer resource requirements than LIDAR systems. We provide a comprehensive investigation and novel contributions in foreground recognition and extraction, image feature generation, and pose estimation. We evaluated $12$ image feature detector and descriptor performances and proposed a new biologically inspired image descriptor. We also assessed the bag of visual words codebook technique for object localisation and evaluated linear, non-linear and non-parametric classifiers. We tested the ResNet and Inception-ResNet convolutional neural networks on target localisation. We developed compact semantic segmentation autoencoders based on AlexNet, U-Net and VGG. We made several new contributions in the image saliency generation. First, we developed a novel principal component analysis based formula for graph manifold ranking Optimal Affinity Matrix inversion which reduces computation time and stabilises the ranking inversion process. We developed a novel weighted gradient orientation histogram feature for monochromatic image superpixel identification and provided three enhanced versions of the graph manifold ranking tested on $32,536$ images. Our technique out-performs the state-of-the-art saliency method in precision and our fastest method is $12{\times}$ faster than the original graph manifold ranking technique. We introduce an innovative false-coloured high-frequency salient feature image to enhance foreground and background pixel histogram distinction. We propose a novel space background classification scheme using pixel statistics to detect Earth passage. We evaluated appearance based pose matching using principal components analysis, SoftPOSIT and $e$P$n$P for pose estimation. We propose a novel homography transform projection method that simplifies the perspective-$n$-point correspondence. We introduce improvements to the SoftPOSIT initiation to reduce the effects of local minimum trapping using centroid matching. We developed region-based pose estimation using level-set segmentation and pixel statistics. Our tests show the region-based method out-performs the appearance-based and point-based methods in speed, precision and stability.