|
Obstacle avoidance from optic flow |
|
For the measurement of image flow, powerful algorithms exist which for a standard CCD chip can be run with 10 or more frames per second on a PC. The detection of egomotion from optic flow has been attempted by several approaches including direct solution of the optic flow equation, separation of translational and rotational components by local vector differences, and matching large vector fields to flow patterns. In the case of independently moving objects, the segmentation of such objects and egomotion detection will have to be done in an iterative scheme. |
|
|
|
|
In translatory movement, the magnitude of image motion detected at the camera target depends on the distance to the imaged objects. A simple strategy for obstacle avoidance therefore measures the average length of image flow vectors on both sides of the vertical mid-line. The agent then turns to the side, where optic flow is less. In addition to this left-right balancing of optic flow, keeping flow magnitude at a constant level generates automatic speed control in which the agent slows down if obstacles comes close and speeds up again if space is free. This principle has been shown operate both in flying insects and in humans steering a car.
|
|
|
|
|
|
While optic flow balancing ensures that the agent will not collide with obstacles, it does not guarantee the approach of a goal. Goal vectors can be provided in various ways, including a visible goal in the image, a represented vector updated by path integration (odometry), or a goal specified by an external user. Attraction to the goal and repellence from the obstacle can be combined in various ways, using either a local 3D map around the agent, or the camera image. |
|
|
|
|
Independently moving obstacles play an important role in automatic vehicles and automotive driver assistance systems. While methods involving object recognition and non-rigid motion detection are relevant for detecting other vehicles or pedestrians, the detection of independently moving objects seems to be the most important strategy for flying robots. This again can be achieved by optic flow calculation by first detecting egomotion and then comparing individual motion vectors with all motion vectors consistent with the known egomotion.
|
|
|
|
|
|
Initially, obstacle avoidance should focus on static objects. The scheme for corridor centering by an earthbound robot will have to be extended to flyers in three-dimensional environments. In addition, speed of the robot can be controlled by the closeness of the obstacle. Indeed, this is implicit to the optic flow approach which gives a graded estimate of obstacle closeness. More elaborate schemes tracking independent motion will have to rely on improved optic flow algorithms. Moreover at this time, the approach of optic flow and of assessment of goal approach has not been extended to 3D environments for flying agents. |
|
|
|
Localisation and navigation |
|
Localization task use several sensory data measurements to produce an estimation of the robot's state in the environment. For that, proprioceptive sensors, like odometry or inertial measurement units, could be used to give an estimation of the state of the robot (position, orientation ...). But as these data are often subject to drift, they need to be corrected or filtered with more information. Exterioceptive data like GPS measurements, laser or ultrasound range finder, video based landmarks are then combined with the objective to build a map, representing in a topological way or geometrical way the sensed environment. The robustness of this sensor fusion depend on the quality of the extracted features. |
|
|
|
|
|
CEA-LIST already has worked in this field with the development of a small autonomous indoor robot. It uses mainly vision, with a monocular camera looking at the ceiling, to achieve localization. After an exploration step, visual natural landmarks on the ceiling are extracted and incrementally added to a 'knowledge visual data set' as the robot explores new surroundings. These extracted features are not only used to build a visual description of the environment for localization, but are also used to perform navigation task. Indeed, trajectories are defined as a graph of key images the robot must follow. This visual path allows to compute in the sensory space control laws for the robot. All the developed algorithms were optimized and integrated in a single chip, FPGA based, computing architecture. This architecture was chosen to implement efficiently a particle filter algorithm used for visual localization. |
|
|
|
|
|
In the case of perception for localization purposes, the difficulties lie in the selection, the extraction and the identification of characteristic features. For terrestrial robots, laser based sensors could lead to accurate and robust environment sensing but these important weight and size sensors are not adapted to be embedded on small flying robots. Vision based systems seem a more promising approach. When applied to flying robots, stereoscopic vision could be used to extract and archive landmarks and then reconstruct 3D environments thanks to Kalman filtering. As stereo vision may not be suitable in every situation when disparity information is difficult to estimate (high altitude, low separation between cameras because of small sized robots), thus monocular based systems may also be used. |
|
|
|
Industrial UAV state of the art |
|
UAVs have so far been mainly limited to the military sector. Several candidates vehicles were designed to meet these military needs from industry and governments developments programs. The US ones are reviewed in the US DoD roadmap. The military UAVs have very specific characteristics: they are medium size and medium weight and reach high speed. Using fixed wing mechanical architecture, more simple than helicopter-like rotary wing, they are usually not able to perform stationary flight. Moreover, they are designed to manoeuvre in free environments and little efforts have been done to make them fully autonomous or reactive to complex environments. This prevents their use in civil and populated area, for safety reasons. |
|
|
|
|
|
The development of civil UAVs and of civil VTOL UAVs in particular, is a growing field of interest. The design of these vehicles involves most of the challenges of analysis of full scale rotorcraft, both in the determination of rotor performance and the study of complex full-airframe interactional aerodynamics issues. |
|
|
|
|
Many different VTOL UAVs have been designed. They can be classified into the following main categories, based on their rotor configuration:
- single main rotor
- co-axial rotors
- tilt rotor
- tandem rotor
- intermeshing rotor
Such rotor configurations use rotary blades to produce the needed propulsion. In addition, there are other configurations such as jet stream, in fan rotors. The most common are single rotors followed by coaxial rotors. |
|
|
|
|
A lot of teleoperated VTOL industrial products exist in Europe, Asia and USA, for instance:
- Hovereye (from the French companies Bertin technologies)
- Survey copter
- Highcam 1 (from Highcam in the Netherlands)
Moreover, many research groups are currently working to build completely autonomous vehicles able to reach a given location without teleoperation (for example in the framework of a university competition entitled 'Miniature UAV system design' organized by DGA abd ONERA in France, or that of the IARC International Aerial Robotics competition).
This results in many prototypes coming from university, aiming at an increased autonomy. However there is still no real prototype able to navigate in urban environment, avoid obstacles and follow a predefined trajectory. |
|
|
|
Mission Planning |
|
In current UAV systems the mission planning is typically performed by a human operator. The human operator must be aviation trained, and the process of planing the mission is time-consuming. In this context aviation trained implies that the operator has been trained to plan missions, which will be conducted under visual flight-rules (VFR) conditions. |
|
|
|
|
|
The planned mission is provided to the UAV, which flies the mission but which cannot autonomously make changes to the mission plan during flight. This approach is both inflexible (the UAV cannot change its own mission plan) and costly (in terms of planning time and operator training). Onboard mission planning is therefore preferred in order to achieve higher flexibility and lower costs. |
|
|
|
|
|
Past approaches have typically operated in two dimensions and have only permitted the UAV to fly withing narrow corridors. Most available mission planning systems address big military UAVs. But few address the micro drone area, as micro-drones are still in development and remain mostly experimental. Therefore an efficient and reliable 3D planning for micro-drones still has to be developed. |
|
|
|
|
|
|
State of the art
