We present a novel method to quickly detect and track objects of low resolution within an image frame by comparing
dense, oriented gradient features at multiple scales within an object chip. The proposed method uses vector correlation
between sets of oriented Haar filter responses from within a local window and an object library to create similarity
measures, where peaks indicate high object probability. Interest points are chosen based on object shape and size so that
each point represents both a distinct spatial location and the shape segment of the object. Each interest point is then
independently searched in subsequent frames, where multiple similarity maps are fused to create a single object
probability map. This method executes in real time by reducing feature calculations and approximations using box
filters and integral images. We achieve invariance to rotation and illumination, because we calculate interest point
orientation and normalize the feature vector scale. The method creates a feature set from a small and localized area,
allowing for accurate detections in low resolution scenarios. This approach can also be extended to include the detection
of partially occluded objects through calculating individual interest point feature vector correlations and clustering points
together. We have tested the method on a subset of the Columbus Large Image Format (CLIF) 2007 dataset, which
provides various low-pixel-on-object moving and stationary vehicles with varying operating conditions. This method
provides accurate results with minimal parameter tuning for robust implementation on aerial, low pixel-on-object data
sets for automated classification applications.
Optical flow-based tracking methods offer the promise of precise, accurate, and reliable analysis of motion, but they
suffer from several challenges such as elimination of background movement, estimation of flow velocity, and optimal
feature selection. Wavelet approximations can offer similar benefits and retain spatial information at coarser scales,
while optical flow estimation increases with the reduction of finer details of moving objects. Optical flow methods often
suffer from significant computational overload. In this study, we have investigated the necessary processing steps to
increase detection and estimation accuracy, while effectively reducing computation time through the reduction of the
image frame size. We have implemented an object tracking algorithm using the optical flow calculated from a phase
change between representative coarse wavelet coefficients in subsequent image frames. We have also compared phasebased
optical flow with two versions of intensity-based optical flow to determine which method produces superior
results under specific operational conditions. The investigation demonstrates the feasibility of using phase-based optical
flow with wavelet approximations for object detection and tracking of low resolution aerial vehicles. We also
demonstrate that this method can work in tandem with feature-based tracking methods to increase tracking accuracy.
We present a novel implementation of multi-scale graph-theoretic image segmentation using wavelet decomposition.
This bottom-up segmentation through a weighted agglomeration approach utilizes the specific statistical characteristics
of vehicles to quickly detect regions of interest in image frames. The method incorporates pixel intensity, texture, and
boundary values to detect salient segments at multiple scales. Wavelet decomposition creates gradient and image
approximations at multiple scales for fast edge weighting between nodes in the graph. Nodes with strong edge weights
merge to form a single node at a higher level, where new internal statistics are calculated and edges are created with
nodes at the new scale. Top-down saliency energy values are then calculated for each pixel on every scale, with the pixel
labeled as a member of the node (segment) at the scale of highest energy. Salient node information is then used for
binary classification as a potential object or non-object passes to classification and tracking algorithms. The method
provides multi-scale segmentations by agglomerating nodes that consist of finer node agglomerations (lower scales).
Criteria for weights between nodes include multi-level features, such as average intensity, variance, and boundary
completion values. This method has been successfully tested on an electro-optical (EO) data set with multiple varying
operating conditions (OCs). It has been shown to successfully segment both fully and partially occluded objects with
minimal false alarms and false negatives. This method can easily be extended to produce more accurate segmentations
through the sensor fusion of registered data types.
Automated classification and tracking approaches suffer from the high-dimensionality of the data and information space,
which frequently rely upon both discriminative feature selection and efficient, accurate supervised classification
strategies. Feature selection strategies have the benefit of representing the data in a modified reduced space to improve
the efficacy of data mining, machine learning, and computer vision approaches. We have developed feature-selection
methods involving feature ranking and assimilation to discover reduced feature sets that produce accurate results in
classification for automated classifiers with significant specificity and sensitivity. We have tested a wide range of spatial,
texture, and wavelet-based feature sets for electro-optical (EO) aerial imagery and infrared (IR) land-based image
sequences on several machine-learning algorithms for classification for performance evaluation and comparison. A
detailed experimental evaluation is provided for the classification efficacy of the features and classifiers on the particular
data sets, and is accompanied by a discussion of the particular success or failure. In the second section, we detail our
novel feature set that combines moment and edge descriptors and produces high, robust accuracy when evaluated for
classification. Our method leverages information previously calculated in the detection stage, which includes wavelet
decomposition and texture statistics. We demonstrate the results of our feature set implementation and discuss methods
for creating classifier decision rules to choose a particular classification algorithm dependent on certain operating
conditions or data types adaptively.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.