Download Full Text (6.3 MB)
Advances in computer vision have led to development of algorithms that are able to extract semantic information from images/video in order to make high level inferences from data. One of the major steps toward extracting semantic information is to identify useful contextual information present in the scene. In this research, we present a novel technique to extract context information from aerial imagery using concatenated vectors of low level features. The objective of this research is to aid in the identification of threats along the right of way of energy pipelines. The key observation of this research is that aerial imagery consists of various image segments like roads, buildings and trees along with lots of plain ground. All aforementioned segments of the image have definitive properties in terms of low level features. The information content present in plain ground is minimal when compared to other regions in the image. This characteristic was exploited to have a simple thresholding procedure designed on the basis of relative variance and entropy for fast background elimination. Trees are rich in textural content, buildings have higher contrast information and roads have discriminative color features. In this research we have extracted local phase information and local contrast information using the monogenic signal model. These features are used to train a support vector machine (SVM) which is then used for classification. In order to refine the segmentation process, we apply morphological operations on the result of the classifier. We present the results obtained by using the proposed method on various data sets captured using different camera sensors.
Vijayan K. Asari
Primary Advisor's Department
Electrical and Computer Engineering
Stander Symposium poster
"Extracting Context Information from Aerial Imagery for Aiding Threat Detection" (2013). Stander Symposium Posters. 334.