Perception Stack of a Self-Driving Car

Abstract

Over 1.3 million people die in car crashes each year, according to the WHO (World Health Organization), and it appears that most accidents are caused by driver error. Self-driving cars can improve safety by drastically reducing collisions and saving lives. Because making self-driving cars can remove human error, it gives some serious safety benefits to these sophisticated artificial intelligence systems. The aim of our project is to develop algorithms for robotic perception. This includes main tasks of static and dynamic object detection, object tracking, depth estimation, collision avoidance, visual odometry and semantic segmentation for drivable surface area. This project is developed on the extensive knowledge of computer vision, deep learning and robotics which employs different AI algorithms to help the vehicle analyse the environment around it We used and integrated number of state-of-the-art algorithms such as YOLO and VGG that gathers information around the vehicle using monocular and stereo cameras and process it in real time on jetson development kit and enables our vehicle to move efficiently.