Video tracking architectures for small low-power embedded systems are severely constrained by their limited processing
capacity and must therefore be highly optimized to meet modern performance requirements. Consequently the various
design trade-offs have a direct and significant impact on the overall system performance.
The evaluation is based on a test framework and a set of metrics for defining tracking performance. Well-known metrics
appropriate to multi-target video-tracking applications have been selected to provide a generalized and meaningful
characterisation of the system and also to allow easier comparison with other video tracking algorithms. The selected set
is extended further with additional architecture-specific metrics to extract a finer level of granularity in the analysis and
support embedded system issues. The tracking system is evaluated within the test framework using a broad spectrum of
real and synthetic video imagery across multiple scenarios.
In each case the embedded systems are required to robustly track multiple targets and accurately control the sensor
platform in real-time. The key focus (hence the requirement for a rapid design and test cycle) is on evaluating the
different system behaviours through testing and then analysing the results to identify how the various design
methodologies affect the overall performance. We briefly compare some analysis of the tracking performance between
two different types of internal track processes we frequently use.
An adaptive image pre-processor has been developed for a next-generation video tracking system. In a previous paper
we presented a wavelet-based enhancement pre-processor (AWEP) which showed good segmentation capability. Here
we discuss the impact of structural and implementation constraints placed on the algorithm during targeting on a low
power FPGA device. We discuss the underlying issues and outline our approach to compensating the effect and
regaining stability. Output results are given illustrating the segmentation performance after applied optimization of the
decomposition filter kernels. A set of results from the tracking system are presented to demonstrate the effectiveness of
the AWEP implementation on the tracking performance applied to real video.
An adaptive image pre-processor has been developed as a high-performance front-end for a next-generation multi-target
tracking (MTT) system. The tracking system is designed to track targets across potentially multiple and distributed
electro-optic video sensors. Typically a pre-processor operates to enhance targets and assist the tracking. However, they
frequently rely on expert knowledge to configure the algorithm for the particular application and hence do not cope
adequately given unexpected variations or generic application. The pre-processor developed for our MTT system
achieves a significantly improved and robust performance by using an adaptive approach based on wavelet
decomposition and a "supporting-classifier" method. It is capable of detecting and dynamically maintaining a target-definition
optimized for tracking, whilst maximally suppressing non-related clutter. This paper presents an overview of
the architecture and demonstrates its performance on real video scenes.
A significant capability of unmanned airborne vehicles (UAV's) is that they can operate tirelessly and at maximum
efficiency in comparison to their human pilot counterparts. However a major limiting factor preventing ultra-long
endurance missions is that they require landing to refuel. Development effort has been directed to allow UAV's to
automatically refuel in the air using current refueling systems and procedures. The 'hose & drogue' refueling system
was targeted as it is considered the more difficult case. Recent flight trials resulted in the first-ever fully autonomous
airborne refueling operation.
Development has gone into precision GPS-based navigation sensors to maneuver the aircraft into the station-keeping
position and onwards to dock with the refueling drogue. However in the terminal phases of docking, the accuracy of the
GPS is operating at its performance limit and also disturbance factors on the flexible hose and basket are not predictable
using an open-loop model. Hence there is significant uncertainty on the position of the refueling drogue relative to the
aircraft, and is insufficient in practical operation to achieve a successful and safe docking.
A solution is to augment the GPS based system with a vision-based sensor component through the terminal phase to
visually acquire and track the drogue in 3D space. The higher bandwidth and resolution of camera sensors gives
significantly better estimates on the state of the drogue position. Disturbances in the actual drogue position caused by
subtle aircraft maneuvers and wind gusting can be visually tracked and compensated for, providing an accurate
This paper discusses the issues involved in visually detecting a refueling drogue, selecting an optimum camera
viewpoint, and acquiring and tracking the drogue throughout a widely varying operating range and conditions.
The ability to automatically detect and track moving targets whilst stabilizing and enhancing the incoming video would be highly beneficial in a range of aerial reconnaissance scenarios. We have implemented a number of image-processing algorithms on our ADEPT hardware to perform these and other useful tasks in real-time. Much of this functionality is currently being migrated onto a smaller PC104 form-factor implementation that would be ideal for UAV applications. In this paper, we show results from both software and hardware implementations of our current suite of algorithms using synthetic and real airborne video. We then investigate an image processing architecture that integrates mosaic formation, stabilisation and enhancement functionality using micro-mosaics, an architecture which yields benefits for all the processes.
Many airborne platforms have high performance electro optical sensor suites mounted on them. Such sensor systems can provide vital, real time reconnaissance information to users on the platform or on the ground. However such sensor systems require control and output large amounts of data of which the user may require only a relatively small amount for his decision processes. This paper describes a payload management system, designed to automatically control an airborne sensor suite to improve the 'quality' of the data provided to the user and other systems on the airborne platform. The system uses real time image-processing algorithms to provide low-level functions e.g. closed loop target tracking, image stabilization, automatic focus control and super-resolution. The system combines such real time outputs and incorporates contextual data inputs to provide higher-level surveillance functions such as recognition and ranging of navigational waypoints for geo-location; registration of image patches for large area terrain imaging. The paper outlines the physical and processing architecture of the system and also gives an overview of the algorithms and capabilities of the system. The issues surrounding the integration into existing airborne platforms are discussed.
During a pre-programmed course to a particular destination, an autonomous vehicle may potentially encounter environments that are unknown at the time of operation. Some regions may contain objects or vehicles that were not anticipated during the mission-planning phase. Often user-intervention is not possible or desirable under these circumstances. Thus it is required for the onboard navigation system to automatically make short-term adjustments to the flight plan and to apply the necessary course corrections. A suitable path is visually navigated through the environment to reliably avoid obstacles without significant deviations from the original course.
This paper describes a general low-cost stereo-vision sensor framework, for passively estimating the range-map between a forward-looking autonomous vehicle and its environment. Typical vehicles may be either unmanned ground or airborne vehicles. The range-map image describes a relative distance from the vehicle to the observed environment and contains information that could be used to compute a navigable flight plan, and also visual and geometric detail about the environment for other onboard processes or future missions.
Aspects relating to information flow through the framework are discussed, along with issues such as robustness, implementation and other advantages and disadvantages of the framework. An outline of the physical structure of the system is presented and an overview of the algorithms and applications of the framework are given.