Bayesian Multi-Lane Road Geometry Estimation
Examensarbete för masterexamen
Systems, control and mechatronics (MPSYS), MSc
Abstract Advanced Driver Assistance Systems (ADAS) are automotive technologies that improve driver comfort and road safety. Many such features, like adaptive cruise control and automatic lane keeping rely on knowledge about the geometry of the road around the ego vehicle. This purpose of this work is to design and evaluate a model based algorithm to estimate the center curves of highway lanes as well as on ramps and off ramps based only on internal vehicle sensors and a front camera system detecting lane markings. In particular, we design a novel road model with support for describing any number of lanes of the ego road, on ramps and off ramps within some range of the ego vehicle. To evaluate whether this road model is suitable for tracking with a Bayesian filter, a multiple hypothesis coupled multi object tracking filter is designed specifically for the chosen road model. Multiple hypothesis are used to deal with data association for detections of lane markers and coupling lanes is useful for modeling parallel lanes. The filter additionally deals with object spawning and death in a way inspired by multi-Bernoulli mixture filters by giving each lane within each hypothesis a probability of existence. The filter is evaluated on highway driving scenarios provided by the Zenseact company and is demonstrated to track the ego lane and parallel lanes well, but sometimes struggles to distinguish main road lanes from on or off ramps since the filter only knows about lane markings and not e.g. road edges or guard rails. The current filter implementation is also computationally intensive and does not reliably run in real time on consumer hardware, but demonstrates the potential of the road model used to track multiple lanes and branching roads.