|
This paper details and evaluates a system that aims to provide continuous robust localisation ('tracking') of vehicles
throughout the scenes of aerial video footage captured by Unmanned Aerial Vehicles (UAVs). The scientific field of
UAV object tracking is well studied in the field of computer vision, with a variety of solutions offered. However,
rigorous evaluation is infrequent, and further novelty lies here in our exploration of the benefits of combined modality
processing, in conjunction with a proposed adaptive feature weighting technique. Building on our previously reported
framework for object-tracking in multi-spectral video1, moving vehicles are initially located by exploiting their intrascene
displacement within a camera-motion compensated video-image domain. For each detected vehicle, a
spatiogram2-based representation is then extracted, which is a representative form that aims to bridge the gap between
the 'coarseness' of histograms and the 'rigidity' of pixel templates. Spatiogram-based region matching then ensues for
each vehicle, towards determining their new locations throughout the subsequent frames of the video sequence. The
framework is flexible in that, in addition to the exploitation of traditional visible spectrum features, it can accommodate
the inclusion of additional feature sources, demonstrated here via the attachment of an infrared channel. Furthermore, the
system provides the option of enabling an adaptive feature weighting mechanism, whereby the transient ability of certain
features to occasionally outperform others is exploited in an adaptive manner, to the envisaged benefit of increased
tracking robustness. The system was developed and tested using the DARPA VIVID2 video dataset3, which is a suite of
multi-spectral (visible and thermal infrared) video files captured from an airborne platform flying at various altitudes.
Evaluation of the system is quantitative, which differentiates it from a large portion of the existing literature, whilst the
results observed serve to further reveal the challenging nature of this problem.
|
