<![CDATA[Autonomous driving systems have become a topic of interest in academia, industry, and the military in recent years. Active sensors such as Light Detection and Ranging (LiDAR) can generally be used to measure the distance to an object, but the costs and computations required are very large. To obtain a relatively lower cost, a monocular camera is the solution. Based on previous research, estimating the depth value in images from monocular cameras using the Deep Neural Networks (DNN) method is proven to work well. The U-Net DNN architecture using Residual Network (ResNet) on the encoder section is used in this research. The process of training, validation, and model testing were carried out on the DIODE: A Dense Indoor and Outdoor Depth Dataset with a total of 8899 data. The training and validation phases are carried out using Adam optimization. This research obtained the best model using a learning rate of 1e-3 and a weight on the loss function of 0.3. This model obtains an evaluation metrics that are able to compete with previous studies with an RMSE value of 0.2272, an REL of 1.3676, an accuracy with a threshold of 1.25 of 56.22%, an accuracy of a threshold of 1.252 of 78.97%, and an accuracy of a threshold of 1.253 of 89.29%. Testing the inference model obtains the number of frames per second (FPS) in the range of 5-12 FPS.]]>
Copyrights © 2022
