Analysis and demonstration of vision driven lane changing system of self-driving car in CARLA simulator

Abstract

Accidents due to driver error persist despite road rules and regulations being updated regularly, mostly because of reckless driving and driving under the influence. Therefore, the inclusion of self-driving cars can greatly reduce these fatal tragedies resulting in safer streets along with an easy mode of transportation for people of all ages and conditions. To implement such self-driving cars, a software simulation model is important. So, in this paper, we focused on the simulation of an automatic lane-changing self-driving system of SAE automation level 4 using the CARLA simulator, to test and validate the avoidance of the slightest intervention of humans, even while changing lanes. The self-driving car system consists of an object detection model based on the Roboflow 3.0 object detection algorithm, which detects objects based on their relative distance to the agent, followed by the decision criteria model which decides the action the self-driving car has to take for the vehicle to maneuver within the CARLA simulated environment. The image datasets being used were uniquely generated by us to help train our machine learning model several times, fine-tuned the decision criteria model and vehicle maneuver techniques for the simulated environment, and provided analysis of the models to give a better understanding of the system. In simulation, the results look promising to be applied in practical applications in controlled environments.