Tracking Keypoints from Consecutive Video Frames Using CNN Features for Space Applications

Hard time constraints in space missions bring in the problem of fast video processing for numerous autonomous tasks. Video processing involves the separation of distinct image frames, fetching image descriptors, applying different machine learning algorithms for object detection, obstacle avoidance, and many more tasks involved in the automatic maneuvering of a spacecraft. These tasks require the most informative descriptions of an image within the time constraints. Tracking these informative points from consecutive image frames is needed in flow estimation applications. Classical algorithms like SIFT and SURF are the milestones in the feature description development. But computational complexity and high time requirements force the critical missions to avoid these techniques to get adopted in real-time processing. Hence a time conservative and less complex pre-trained Convolutional Neural Network (CNN) model is chosen in this paper as a feature descriptor. 7-layer CNN model is designed and implemented with pre-trained VGG model parameters and then these CNN features are used to match the points of interests from consecutive image frames of a lunar descent video. The performance of the system is evaluated based on visual and empirical keypoints matching. The scores of matches between two consecutive images from the video using CNN features are then compared with state-of-the-art algorithms like SIFT and SURF. The results show that CNN features are more reliable and robust in case of time-critical video processing tasks for keypoint tracking applications of space missions.