Lane Detection and Estimation from Surround View Camera Sensing Systems

Autonomous driving poses unique challenges for vehicle environment perception systems. It is highly desirable that we utilize existing vehicle-equipped driver-assistant sensors, without hardware change, to achieve driverless performance. Current product level vehicle surround view camera module (denoted concisely as SVS) is served as a panoramic view visual aid tool for low-automation applications. With proper statistical analysis, the multiple mono-camera information can be very useful for higher vehicle intelligence without significant hardware change. In this study, we focus on lane detection and estimation from a SVS only system. The major difficulty lies in the fact that mono-cameras of the SVS are non-cooperative and essentially of protractor nature; this would lead to large uncertainty on object depth information and incomplete lane observations. We process the highly distorted data in a multi-stage manner. We first utilize a neural network classifier to yield labeled lane-relevant objects. The lane marks/edges point clouds are processed by a truncated Gaussian random field model for the spatial filtering and a fading memory model for the temporal filtering. Then we present polynomial fitting scheme and a statistical analysis of the fitting errors reveals good lane and ego-vehicle orientation cues. In a parking lot real world study, we show promising lane detection and estimation performance of significant practical implications for lane keeping capability in high-automation applications.