Autonomous Aerial Vehicle Navigation without Reliance on GPS

点击翻译按钮获取中文摘要

Small autonomous rotorcraft have seen a surge of activity over the last decade, and have demonstrated some fascinating indoor capabilities such as autonomous flight through windows, synchronized flying, and even playing catch. However, progress seems stalled at this point, as systems are currently required to operate within an indoor positioning system (IPS), or in the open where GPS satellite visibility can be ensured. Small rotorcraft are ideally suited to inspection tasks that require precise positioning in close proximity to objects, but their reliance on IPS or GPS for accurate localization and their inability to perceive their environment and navigate through it autonomously drastically reduces their current utility. As a result, today’s small UAVs are essentially tethered to within line-of-sight of operators and away from any objects in their surroundings. It is time to cut the tether. By augmenting quadrotor vehicles with multi-camera clusters, I plan to develop vision-based localization and mapping algorithms that can complement GPS information and detect obstacles in the environment. Vision-based localization can act as an estimation refinement for GPS and inertial measurements or independently if necessary. Multiple cameras can be easily integrated into small rotorcraft and will provide robust motion estimation and reasonable 3D scene reconstruction, as long as sufficient computation power is available. The approach to full quadrotor autonomy outlined in this proposal involves four main components. The first component is to develop new low-power, light weight vision and computing modules that exploit the latest advances from smartphone chips to offload much of the computation effort from the central unit. The high framerates that can be achieved with this architecture will be the first defense against poor feature correspondence that plagues current approaches. The next component involves developing robust methods for vehicle localization in static environments. Given known sensor models and accurate motion predictions, it will be possible to apply integrity monitoring techniques such as parity space thresholding to not only reject erroneous measurements, but do so knowing the probability of failure and false positives when designing the system. The challenge lies in developing the necessary fundamental changes to the description of the visual sensor data, and the expansion of parity space integrity monitoring techniques to nonlinear systems with large numbers of measurements and multiple outliers. The third component involves the development of a dense reconstruction of the scene for motion planning purposes. Relying on occupancy grid map representations that permit rapid assimilation of new measurements, an overapproximation to the environment will be constructed based on all tracked features in the environment. Finally, the above methods rely on a static environment assumption, but most applications require autonomous operation near moving objects such as people and vehicles. We will develop methods to detect and track moving objects, eliminating their current negative effect on motion estimation, and enabling vision based tracking and collision avoidance. With these components integrated onto a single quadrotor vehicle, it will finally be possible to operate autonomously in close proximity to obstacles with and without reliable GPS measurements. Vast inspection markets in manufacturing, energy generation and distribution, infrastructure maintenance and videography await once rotorcraft are able to assess dangers and fly with guaranteed reliability regardless of their surroundings.