End-to-End Xception Model Implementation on Carla Self Driving Car in Moderate Dense Environment

Recently, with hardware limitation, many autonomous car developers using a simulator to test their network model or solve some self-driving car issues. With that in mind, Carla simulator provides an open platform with many different varieties of maps and real environment parameter, which indicate multiple challenges to be accomplished. There are many approaches to solve these problems, ranging from a complex model such as imitation learning followed by inverse reinforcement learning to a simple model adopting spatial or time-based network with performance-based oriented putting computational time aside. Pertaining this matter, we look into a light-weight model for spatial classification which can reduce computational time with a slight trade back. Mapping cross-channel correlations and spatial correlations in the feature maps separately in extreme Inception, or Xception has outperformed inception v3 slightly on the Imagenet dataset. Moreover, it has the same number of model parameters as inception, which implies a greater computational efficiency. While, on the recent work, Nvidia model or Pilotnet, a CNN based model, has successfully tested their design to map images into control parameter value on the autonomous car system. Therefore, this development motivates us to use the Xception model in the self-driving car context using Carla simulator. In the test simulation, Xception model can work well, reaching the designated destination. It displays a better steering score in comparison to Nvidia model in the best form.