Building and Climbing based Visual Navigation Framework for Self-Driving Cars

Abstract

This paper proposes a visual navigation strategy for self-driving car running on a constant-width road. The task is to process road image with multiple elements information for planing path and providing the basis for acceleration and deceleration. Common road elements are straight, bend, ramp and crossroad. We propose a novel navigation framework (BCVN) that explicitly decomposes the visual navigation task into navigation line extraction, deviation calculation and curvature calculation. The core idea of navigation line extraction is Building-Climbing. Building is to build foundations with a small number of consecutive points. Climbing is to climb points on the basis of the foundations. Building and Climbing are both used in search of bilateral edges. Deviation calculation use the method of dynamic weighting for self-driving car to control steering. Curvature calculation is to obtain a suitable value for self-driving car to achieve acceleration and deceleration control. We use least squares algorithm to assist in bilateral edges search and curvature calculation. We describe our real-time implementation of the BCVN framework, the method of dynamic weight and Building-Climbing. We test the strategy on the self-driving car platform, which shows strong adaptability and high efficiency.