Autonomous Vehicle Based on Deep Q-Learning and YOLOv3 with Data Augmentation

This paper presents a model which imitates human in all ways of driving without human intervention be it recognizing the car states. It incorporates the car speed, position to track, and the surrounding environments based on a combination of Deep Q-Learning and YOLOv3 on the Raspberry Pi computer. Environmental learning is done using Deep Q-Learning. Data can be collected by using Raspberry pi and a front-facing camera. A DNN model with Deep Q-Learning is trained to get high performance in recognition and control tasks. Deep Q-Learning approximates the values by using Deep Neural Networks (DNN). The neural network takes the initial state as input returns the Q-value of all possible actions as an output. Deep Q-Learning addresses correlations between samples and non-stationary by experience replay and fixed Q-targets. YOLOv3 with data augmentation is used to detect and recognize stationary and movable objects, traffic lights, and road signs. YOLOv3 has improved bounding box prediction, more accurate class predictions, and improved abilities at different scales and is much faster. It is easy to optimize since it works based on algorithms that use only one neural network to execute all the task components. YOLOv3, along with data augmentation, significantly increases the data diversity without actually collecting the new training data.