Vision-based road slope estimation methods using road lines or local features from instant images

In this study, the authors study lateral and longitudinal road slope estimation, which are important parameters to control a vehicle laterally in lane-keeping assistance and longitudinally in adaptive cruise control, respectively. They present three methods estimating road slopes from instant road images obtained from a front monocular camera. In the first geometry-based (GB) method, they obtain the slope estimates using 2D road lines derived from 3D road line and pin-hole camera models. In the second local features-based (LFB) method, they use the SIFT local features between two consecutive image pairs. In the first two methods, they derive computationally simple novel analytic expressions for the road slope estimates. As a third solution in covariance based (CB) method, they consider supervised learning where they use 2D road lines as features to train their algorithm. GB and CB methods rely on road lines and provide in advance estimations that enable proactive driving, whereas LFB uses the entire image and performs estimation when the road slope actually changes. They use multi-layer perceptron learning method as a benchmark in their tests. They compare the estimation performance, system requirements, and applicable conditions of the proposed methods under controlled and uncontrolled test environments.