چكيده به لاتين
Deep learning is a branch of machine learning and artificial intelligence and a set of algorithms that attempt to model high-level abstract concepts using learning at different levels and layers. Due to their high accuracy and efficiency, they have been used in self-driving cars. Many car accidents are caused by human faults. One of the reasons for paying attention to self-driving cars is to prevent these accidents by eliminating or reducing human intervention, and the result will be reducing human and financial losses. In this project, the decision-making part of self-driving cars has been studied. The decision-making or control section for most approaches in self-driving cars falls into three general categories: mediated perception-based approach, direct-perception approach, and behavioral response approach. The advantage of behavioral response approaches is performing all tasks required for driving, in an End-to-End manner. We have developed a behavioral response system based on the perception and control of self-driving vehicles and tested the proposed model in the simulator. The approach presented here consists of three parts: perception, memory, and control. In the perception section, the dimension of the images is reduced by an autoencoder. These smaller images, known as latent representation, then enter the memory section and predict the next frame latent representation. Finally, the various inputs, which in the complete model are the raw input images, the current latent representation, the next frame latent representation, semantic segmentation, and non-visual inputs containing the current control commands, and the current conditions enter the control section, and different outputs are obtained. The approach presented here is a vision-based algorithm due to the decision algorithm's use of image data. The obtained results are first evaluated by the criteria presented in each of the three sections, and then the performance of the overall system is evaluated in the simulator. Here we have been able to achieve 100% accuracy in the Straight task from the Carla simulator benchmark in different weather and lighting conditions in both existing cities, which has performed better than similar works. Also, the proposed system prediction time is 40 milliseconds for each frame of input images, which is a good time considering the acceptable time for self-driving cars, which is equal to 100 milliseconds.
