Depth Estimation in Static Monocular Vision with Stereo Vision Assisted Deep Learning Approach

This paper introduces a method for depth estimation that combines the strengths of geometric and deep learning approaches, tailored for use with fisheye cameras. By leveraging the wide field of view (FOV) provided by fisheye lenses and the accuracy of deep learning models, we aim to overcome the limitations inherent in traditional monocular and stereo vision depth estimation techniques. Our approach integrates a geometry-based stereo vision model with unparallel camera axes, specifically designed for fisheye cameras, with state-of-the-art deep learning methods for monocular depth estimation. We utilize existing datasets and pre-trained networks, including vision transformers and Fully Convolutional Networks (FCNs), to address the challenges of data scarcity and computational resource limitations. The methodology involves calibration of fisheye cameras, depth estimation from geometric and learning-based methods, and a novel combination of these methods to achieve higher accuracy and a broader FOV than currently possible with either approach alone. Our experiments, conducted using an RGB-D camera for validation, demonstrate the potential of our method to enhance depth map quality significantly. This research not only fills a gap in the literature by merging geometric and learning-based depth estimation for fisheye cameras but also proposes a practical application within the drone project's obstacle detection function. Future work will explore the refinement of the geometric model, the integration of different learning-based methods, and the development of a custom network to further improve depth estimation accuracy and applicabllity.