Monocular Vision Aided Autonomous UAV Navigation in Indoor Environments

Abstract

In recent years, Unmanned Aerial Vehicles (UAV), popularly known as drones, have swiftly gained popularity in various sectors such as disaster hit environments, military and industrial applications, agriculture, etc. Vision-based pose estimation of UAV has recently been very popular among the computer vision community. In this doctoral research, an attempt has been made to navigate UAV autonomously in GPS-denied indoor environments with the feed from a monocular camera only. Understanding the 3D perspective of a scene is imperative in improving the precision of intelligent autonomous systems. However, with only one camera (monocular vision) feed, the difficulty in understanding becomes compounded. This research focuses on estimating 3D primitives using only the visual data captured with a monocular static camera and without any additional sensors. The primitives thus estimated are used for safe navigation of UAV in the presence of static obstacles in indoor environments. A monocular vision assisted optical flow method is proposed to measure the depth of a UAV from an impending frontal obstacle. The approach follows the fundamental principle of perspective vision that the size of an object relative to its field of view (FoV), increases as the center of projection moves closer towards the object. This involves modeling the depth followed by its realization through scale-invariant visual features. Noisy depth measurements arising due to the external wind, or the turbulence in the UAV, are rectified by employing a constant velocity based Kalman filter model. Rigorous experiments with scale-invariant features reveal an overall accuracy of 89.6% with varying obstacles, in both indoor and outdoor environments. However, this method requires frequent command updates to keep the UAV on safe path. A two-stage deep neural network (DNN) model is developed to predict the instantaneous optimal trajectory of the UAV from an input RGB image. In the first stage, the depth map and surface normal of the image are predicted using a fully convolutional deep architecture. The network is trained separately for both the tasks while keeping the architecture same. In the second stage, the predicted depth and surface normal are jointly processed through another trained DNN classifier model to predict the optimal trajectory of the UAV. The suggested architecture, compared to the counterparts, uses fewer training samples and model parameters. Instantaneous optimal trajectories also help to overcome the issue of low frequency command updates, which is a drawback of the previous method.
The previous methods do not work well inside long corridor environments. Hence, an optical flow based vanishing point method is proposed to navigate a UAV safely inside indoor corridor environments. The proposed algorithm makes sure that the UAV avoids the side wall as well as the frontal wall at the end of the corridor. The knowledge of the vanishing point location alongside a formulated mechanism governs the necessary set of commands to safely navigate the UAV avoiding any collision with the side walls. Furthermore, the relative Euclidean distance scale expansion of matched scale-invariant keypoints in a pair of frames is taken into account to estimate the depth of a frontal obstacle; usually a wall at the end of the corridor. Exhaustive experiments in different corridors reveal the efficacy of the proposed scheme. A DNN based monocular vision assisted algorithm is proposed to safely localize a UAV in indoor corridor environments. Always, the aim is to navigate the UAV through a corridor in the forward direction by keeping it at the center with no orientation either to the left or right side. The algorithm makes use of the RGB image, captured from the UAV front camera, and passes it through trained DNN models to predict the position of the UAV with respect to the central bisector line (CBL) of the corridor. If the UAV is disoriented, an appropriate command is generated to rectify the pose. A new corridor dataset, named NitrUAVCorridorV1, that contains images as captured by the UAV front camera when the UAV is at all possible locations of a variety of corridors, is also proposed. As per our knowledge, we are the first ones to propose such a dataset to navigate a UAV inside corridor environments. The performance of all the proposed methods is experimentally validated in real-world indoor environments. Also, the results are compared qualitatively as well as quantitatively with the respective counterparts.