Vision-Based Self-Mapping for Autonomous Indoor Drones

Objectives: Without GPS signals the problem of self-autonomous navigation of indoor drones is greatly complicated and consequently there has been a recent surge of interest in Visual SLAM. Such systems enable the localization of a drone in an environment and the building of a map of it using only camera data. However indoor environments generally pose severe problems for these systems such as low illumination, motion blur due to rapid motion and featureless surfaces that do not offer the reliable features required for tracking. Severity of such -problems impairs, in extreme cases, the spatial orientation and the information of the place, which decreases localization quality and the robustness. The goal of this thesis is to develop an adaptive hybrid image enhancement technique to satisfy the real-time and localization accuracy criteria of visual SLAM based indoor drone navigation. Material and Methods: The system combines the simple preprocessing methods, Gamma Correction and Contrast Limited Adaptive Histogram Equalization (CLAHE), with a particular enhancement model based on Convolutional Neural Networks (CNNs). When the Decision Tree classifier determines that the coverage of brightness, sharpness and feature density is such that advanced enhancement is required, advanced enhancement procedures are activated. This system is implemented as an external preprocessing component for ORB-SLAM3 and does not require modifications to the preprocessing pipeline of ORB-SLAM3. Evaluation was performed on two orthogonal data sets EuRoC MH01 and TartanAir P001. The evaluation of localization was performed using the two metrics, Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), obtained with help of the evo toolkit. Results: The strategy shown had a drastic reduction in the ATE with EuRoC MH01, with the ATE of 1.431 m down to the 0.978 m representing a 31.7% reduction in ATE. Moreover, the difficult TartanAir P001 test resulted in a great reduction in the mean error, from 6.667 m to 0.510 m, which is a 92.3% reduction. The Decision Tree classifier had a very high accuracy of 99.97% and a recall of 1, and only 0.41% of the images were passed through the CNN enhancement module. Within the adaptive framework enhancement preprocessing was processed at a rate of 696.67 frames per second (FPS) while in the continuous CNN the enhancement preprocessing rate was only at 69.99 FPS which is a 92.06% decrease in processing time. Conclusion: Hence the adaptive framework integrates localization precision and computational efficiency and is therefore suitable for real time indoor drone navigation. Selective enhancement greatly reduces the computational cost and maintains good indoor localization performance in different indoor environments. Future studies may explore implementation of the presented framework on embedded systems and its application to stereo and visual-inertial SLAM systems.