Autonomous Driving System Using CNN and PPO

Autonomous driving has emerged as a solution for the problems which a person faces while driving. Autonomous driving technology not only decreases the transportation challenges but also increases safety. This paper presents an autonomous driving algorithm engineered and executed using Proximal Policy Optimization (PPO), a reinforcement learning (RL) technique, within the Car Learning to Act (CARLA) simulation environment. By keeping in mind, the complexities of Indian roads, such as obstacles or animals on road, unstructured traffic and dynamic conditions, this paper proposes an algorithm designed in a way to adapt these complexities easily. Autonomous driving focuses on two things: Computer Vision and Decision-making. The proposed algorithm prepares a vehicle to be able to work with real world images using the proposed Convolutional Neural Network (CNN) model and generate segmented, labelled images which labels each pixel of the image with a specific class. The RL model is trained on these images which shows that the algorithm is safe and more efficient as compared to the existing work. Through extensive simulations the proposed work shows that the proposed algorithm surpasses most of the existing methods in terms of safety and efficiency. Simulation results shown that the proposed PPO-based framework gain 15% more reward and 20% longer episode duration compared to the prior models, indicating its superior adaptability and robustness.