Design and Development of an Autonomous Car using Object Detection with YOLOv4

Future cars are anticipated to be driverless; point-to-point transportation services capable of avoiding fatalities. To achieve this goal, auto-manufacturers have been investing to realize the potential autonomous driving. In this regard, we present a self-driving model car capable of autonomous driving using object-detection as a primary means of steering, on a track made of colored cones. This paper goes through the process of fabricating a model vehicle, from its embedded hardware platform, to the end-to-end ML pipeline necessary for automated data acquisition and model-training, thereby allowing a deep learning model to derive input from the hardware platform to control the car’s movements. This guides the car autonomously and adapts well to real-time tracks without manual feature-extraction. This paper presents a computer vision model that learns from video data and involves image processing, augmentation, behavioral cloning and a convolutional neural network model. The darknet architecture is used to detect objects through a video segment and convert it into a 3D navigable path. Finally, the paper touches upon the conclusion, results and scope of future improvement in the technique used.