A DQN-Based Autonomous Car-Following Framework Using RGB-D Frames

Modeling car-following behavior has recently garnered much attention due to the wide variety of applications it may be utilized in, such as accident analysis, driver assessment, and support systems. Some of the latest approaches investigate scenario-based autonomous driving algorithms. In this paper, we propose an end-to-end car-following framework that, based on high dimensional RGB-D features only, it ensures autonomous driving by following the actions of a leader car while taking into account other environmental factors (e.g. pedestrians, sidewalk crashing, etc.) To this end, a reinforcement learning (RL) algorithm, precisely an improved Deep Q-Network algorithm, is designed to avoid crashes with the leader car and its detection loss while effectively driving on road. The model is trained and tested using the CARLA simulator in different environments. Our preliminary tests show promising results for enhancing the driving capabilities of autonomous vehicles in many situations such as highways, one-way roads, and no-overtaking roads.