Multi-frame Depth Estimation for Autonomous Driving Using Motion Masking and Self-supervised Learning

Fisheye cameras are often employed in automatic driving for 360 \(\circ \) near-field sensing around the car. This paper suggests a multi-task visual perception network using rectified fisheye photographs and motion masking to assist the vehicle in detecting its surroundings. However, strong nonlinear distortions constitute a trade-off and necessitate more complex algorithms. Dense-depth supervision in practice has shown promise and may be able to address the problem, even though it can be difficult to offer accurate methods for self-supervised learning. The supervision signal from the neighbouring frames is used to train self-supervised depth estimation networks to forecast scene depth. Additionally, the information between the consecutive frames consists of moving objects which has relative motion between them with respect to the motion of the camera which will produce bias and must be tackled with motion masking annotations before feeding to the network. Also, for further ablation study, the proposed system is trained with the unstructured environments of roads that most developing countries have. Thus, we used AdasInd dataset along with WoodScape and Kitti dataset. Further, to measure the accuracy of the output from the depth pixel, a novel measure relative pixel depth alignment (RPDA) is proposed. Most monocular networks do not make use of this additional signal, ignoring crucial information that can enhance expected depth. Therefore, we propose multi-frame depth estimation from fisheye images using self-supervised learning, a dense adaptive depth estimation method that makes use of the sequential data accessible during test time.