Autonomous Driving Control Based on the Technique of Semantic Segmentation

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Current Advanced Driver Assistance Systems (ADAS) are limited to simple driving scenarios.
• Driver intervention is required in emergencies, posing safety risks.
• The development of fully autonomous vehicles is a key research goal.

Purpose of the Study:

• To develop and evaluate autonomous driving control using Deep Reinforcement Learning (DRL).
• To compare the performance of different DRL models, specifically Deep Deterministic Policy Gradient (DDPG) and Recurrent Deterministic Policy Gradient (RDPG).
• To investigate the impact of sensor input, namely Red-Green-Blue (RGB) cameras and semantic segmentation cameras, on autonomous driving performance.

Main Methods:

• Implementation of autonomous driving control within the CARLA simulator.
• Utilizing RGB and semantic segmentation camera feeds as input for DRL models.
• Training and evaluating DDPG and RDPG algorithms with a custom reward mechanism.
• Testing model performance on driving scenarios not encountered during training.

Main Results:

• Recurrent Deterministic Policy Gradient (RDPG) strategies successfully completed driving missions in unseen scenarios.
• The use of semantic segmentation cameras significantly improved the efficiency of the RDPG control strategy.
• Deep Deterministic Policy Gradient (DDPG) models showed limitations in handling novel driving situations.

Conclusions:

• RDPG demonstrates superior capability for robust autonomous driving control compared to DDPG.
• Semantic segmentation enhances the effectiveness of DRL-based autonomous driving systems.
• This research contributes to the advancement of AI-driven autonomous vehicle technology.