Lane Keeping and Navigation of a Self-driving RC Car Based on Image Semantic Segmentation and GPS Fusion

Abstract

This work aims to develop a lane-keeping and GPS navigation system for a self-driving vehicle on a real road. For the lane-keeping task, we trained and deployed a light-weight bilateral semantic segmentation network to segment out drivable area on the road from the camera along with other objects like cars and humans and derived the steering formula for the vehicle. As for navigation, we use a GPS receiver and an IMU module to estimate the position, heading, and orientation of the vehicle with respect to a designated GPS track. To improve the result of state estimation, we modified the Extended Kalman Filter to effectively estimate the state of the vehicle by fusing GPS and IMU sensors. In the end, we discuss a fusion strategy to navigate the vehicle in different road scenarios. The system runs on an Nvidia Jetson TX2 on a 1/10 RC car model, the experiment and measurements were conducted on the internal road of HCMCUTE campus.