Advanced Driver Assistance Systems (ADAS) aids the driver by providing information or automating the driving related tasks to improve driver comfort, reduce workload and improve safety. The vehicle senses its external environment using sensors, building a representation of the world used by the control systems. In on-road applications, the perception focuses on establishing the location of other road participants such as vehicles and pedestrians and identifying the road trajectory. Perception in the off-road environment is more complex, as the structure found in urban environments is absent. Off-road perception deals with the estimation of surface topography and surface type, which are the factors that will affect vehicle behaviour in unstructured environments. Off-road perception has seldom been explored in automotive context. For autonomous off-road driving, the perception solutions are primarily related to robotics and not directly applicable in the ADAS domain due to the different goals of unmanned autonomous systems, their complexity and the cost of employed sensors. Such applications consider only the impact of the terrain on the vehicle safety and progress but do not account for the driver comfort and assistance. This work addresses the problem of processing vision sensor data to extract the required information about the terrain. The main focus of this work is on the perception task with the constraints of automotive sensors and the requirements of the ADAS systems. By providing a semantic representation of the off-road environment including terrain attributes such as terrain type, description of the terrain topography and surface roughness, the perception system can cater for the requirements of the next generation of off-road ADAS proposed by Land Rover. Firstly, a novel and computationally efficient terrain recognition method was developed. The method facilitates recognition of low friction grass surfaces in real-time with high accuracy, by applying machine learning Support Vector Machine with illumination invariant normalised RGB colour descriptors. The proposed method was analysed and its performance was evaluated experimentally in off-road environments. Terrain recognition performance was evaluated on a variety of different surface types including grass, gravel and tarmac, showing high grass detection performance with accuracy of 97%. Secondly, a terrain geometry identification method was proposed which facilitates semantic representation of the terrain in terms of macro terrain features such as slopes, crest and ditches. The terrain geometry identification method processes 3D information reconstructed from stereo imagery and constructs a compact grid representation of the surface topography. This representation is further processed to extract object representation of slopes, ditches and crests. Thirdly, a novel method for surface roughness identification was proposed. The surface roughness descriptor is then further used to recommend a vehicle velocity, which will maintain passenger comfort. Surface roughness is described by the Power Spectral Density of the surface profile which correlates with the acceleration experienced by the vehicle. The surface roughness descriptor is then mapped onto vehicle speed recommendation so that the speed of the vehicle can be adapted in anticipation of the surface roughness. Terrain geometry and surface roughness identification performance were evaluated on a range of off-road courses with varying topology showing the capability of the system to correctly identify terrain features up to 20 m ahead of the vehicle and analyse surface roughness up to 15 m ahead of the vehicle. The speed was recommended correctly within +/- 5 kph. Further, the impact of the perception system on the speed adaptation was evaluated, showing the improvements in speed adaptation allowing for greater passenger comfort. The developed perception components facilitated the development of new off-road ADAS systems and were successfully applied in prototype vehicles. The proposed off-road ADAS are planned to be introduced in future generations of Land Rover products. The benefits of this research also included new Intellectual Property generated for Jaguar Land Rover. In the wider context, the enhanced off-road perception capability may facilitate further development of off-road automated driving and off-road autonomy within the constraints of the automotive platform.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:731437 |
| Date | January 2017 |
| Creators | Gaszczak, Anna |
| Publisher | University of Warwick |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://wrap.warwick.ac.uk/97618/ |
Page generated in 0.0017 seconds