3D COMPUTER VISION

SUMMARY

The trend towards asymmetric operations signifies the need to operate within a military environment characterised by complexity and uncertainty. Rich and up-to-date surveillance data needs to be collected from around and within potentially intricate (e.g. urban) structure. Such structure is also potentially of the most uncertain and dangerous kind for military personnel.

Coverage of such complexity implies multiplicity of sensors, which also implies cheapness (and disposability). The difficulties of deployment, and particularly the dangers of human deployment, implies the use of intelligent sensors and intelligent delivery platforms.

The intelligent sensor needs not only to direct its data gathering activity for a primary surveillance role, but it must also take measures to ensure its own survival, to get to the places it needs to be, to position itself to obtain quality data, to adapt to events and changing circumstances and even to take specific actions (e.g. designate for a weapon).

Computer vision is the sine qua non of machine intelligence. In particular, passive vision, attempts to do what human beings do so readily – use the huge quantity of free data that exists most of the time in most places. While computer vision is very difficult and remains an infant technology, there have been significant advances in both 3D generic vision for structure analysis, and scene interpretation in terms of 3D model-based processes, which are promising real capabilities in autonomy. These algorithmic advances are generally such that they require little beyond current sensor capability and are capable of profitably using the smallest and cheapest imaging devices.

The proposed work seeks to demonstrate proof of principle for the basic functionality required of autonomous machines in these environments.

MILITARY BENEFITS

Intelligent platforms can act as a force multiplier, providing data or performing actions that military personnel might otherwise have to carry out. While such devices will be necessarily much less capable than human beings, they can add to effectiveness either by weight of numbers, by going where human beings cannot or do not wish to go or by playing sacrificial or dispensable roles. The essential requirements, if such force multiplication is anticipated, are cheapness and intelligence.

CRP programmes have already indicated a significant military interest in the use of unmanned vehicles and it is understood that further focus is now being directed at micro-UAVs. Autonomy is the obvious logical extension to the UAV concept and the inevitable requirement of the micro-UAV. Additionally, as UAVs are increasingly used, there is an increasing likelihood of such machines straying into civil environments and there is thus a need to provide greater intelligence simply to reduce the burden on military personnel of ensuring such events have minimal impact.

RESEARCH OBJECTIVE

The research intends to demonstrate that small-scale sensor platforms both for ground and air operation, are capable of operating in directed autonomous mode within complex environments, both containing natural (e.g. vegetative) entities and human artefact (e.g. buildings). We aim to show that such platforms can journey through such environments, surviving, navigating, mapping and planning, largely on the basis of passive computer vision.

RESEARCH OUTLINE

The overall concept of sensor platforms that can survive, navigate and plan for action or information gathering within complex environments, is ambitious and not all such functionality is achievable within a programme of this size. Nevertheless, it is believed that the essential survival and navigation functionality will be demonstrable in real-time on a real platform, with sufficient human supervision to ensure that infelicities of performance have no major operational consequences.

Because the Consortium brings to this project significant expertise in both generic and model-based 3D vision, it is both feasible and sensible to introduce a demonstration platform early in the programme. It is believed that at the end of year 1, a live demonstration can be given of structure-from-motion processes interpreting the visual scene on a simple mobile platform and providing navigation and driving commands to it. This demonstration platform is envisaged as the development machine for implementing more powerful vision, navigation, mapping and path planning algorithms during years 2-3.

Structure-from-motion/stereo processes only provide a partial description of the world at present. Being feature-based, it is rich where features are well distributed, sometimes difficult to interpret where features are dense (e.g. within a tree), and impoverished where features are sparse. In the latter case human beings can fill in data by other means (e.g. shape-from-shading), which algorithmic processes have yet to imitate. Until computer vision algorithms advance to deal with some of these more difficult aspects of scene understanding, there is some promise of assistance from active processes. 3D extraction through active laser devices is explored as a sub-theme within the EO theme and it is anticipated that the passive 3D robotic vision programme will be able to incorporate such inputs in the 2nd and 3rd years of the research. It is anticipated, therefore, that a machine vision concept can be developed, in which intelligently directed active devices can be used to resolve ambiguities or fill in the gaps within the 3D representations from the passive vision processes.

CO-ORDINATION WITH EXISTING / PREVIOUS RESEARCH

This research will develop themes of 3D computer vision which have been the object of significant previous internally funded programmes and which represent the state-of-the-art in machine intelligence. The programme will attempt to align itself with ongoing MOD concept developments in the area of UAVs and, in particular micro-UAVs. While the principal vision process to be investigated is generic (directed at extracting structure in an unknown world), it is anticipated that it may develop so as to bring in model-based vision processes. This draws on internally funded research programmes (including for applications in the civil domain, such as ball and player tracking in sports broadcasting and tool tracking in surgical navigation) and aligns to current MOD work in precision guidance.

Leave a Reply

Your email address will not be published. Required fields are marked *