Comparative Analysis of Deep Learning-Based Depth Estimation Using KITTI Dataset for Mobile Robotics

In autonomous mobile robotics, depth estimation is an important vision perception task. It is mainly used for safe path planning with obstacle avoidance in real time environment. Traditional method involves LIDAR, RADAR, camera and GPS module to estimate the depth information. These sensors although give reliable measurement but less suitable such as low cost and light weight deployment. The advance development of deep learning predicts the depth estimation without any external hardware using KITTI dataset. In this paper U-Net, ResNet, and UniDepth models were compared and its performance was evaluated. About 40000 images were used for evaluation and estimation metrics were evaluated.it has been observed that UniDepth model outperforms the other models  with highest accuracy of 90% as it involves transformer architecture and suitable for complex environment.