Perception-aware visual odometry

Stephan Visschers
Presentation MSc presentation
Date 2019-01-21
Time 15:30
Location Carré 3446

Visual Odometry (VO) is the practice of motion estimation of a mobile robot from a series of images, which is an appropriate localization tool in a GPS-denied environment. Visual odometry relies on the extraction and matching of distinct features within the recorded images to determine the robot pose. Consequently, when a view lacks the necessary information to estimate the current pose of the
robot, the localization process loses its tracking, which is generally required to control the robot.

A potential solution to prevent the lack of necessary information is to implement a gaze control algorithm that actively actuates the camera towards a more favourable gaze. In this project, the tracking problem has been investigated by preliminary experiments and theoretical analysis, to analyse the conditions which facilitate tracking loss. A control algorithm has been designed in which the setpoints of the
physical system are determined by a virtual mechanical model driven by the feature distribution in camera frame and the state of the system. The influence of the features on the system is continuously adapted to take into account the undesired conditions found in the analysis step.

The resulting change in behaviour of the motion of the camera unit is investigated using a series of simulation experiments, which test and demonstrate the potential benefit of the designed gaze control algorithm and its individual aspects.

Posted on Wednesday, January 2, 2019