This paper examines the problem of estimating vehicle position and direction, i.e., pose, from a single vehicle-mounted camera. A drawback of pose estimation using vision only is that it fails when image information is poor. Consequently, other information sources, e. g., motion models and sensors, may be used to complement vision to improve the estimates. We propose to combine standard in-vehicle sensor data and vehicle motion models with the accuracy of local visual bundle adjustment. This means that pose estimates are optimized with regard not only to observed image features but also to a single-track vehicle model and standard in-vehicle sensors. The described method has been experimentally tested on challenging data sets at both low and high vehicle speeds as well as a data set with moving objects. The vehicle motion model in combination with in-vehicle sensors exhibit good accuracy in estimating planar vehicle motion. Results show that this property is preserved, when combining these information sources with vision. Furthermore, the accuracy obtained from vision-only in direction estimation is improved, primarily in situations in which there are few matched visual features.
Read full abstract