Vehicle detection in wide area aerial surveillance using Temporal Context

Moving vehicle detection from wide area aerial surveillance is an important and challenging task, which can be aided by context information. In this paper, we present a Temporal Context(TC) which can capture the road information. In contrast with previous methods to exploit road information, TC does not need to get the location of the road first or to use the Geographical Information System's (GIS) information. We first use background subtraction to generate the candidates, then build TC based on the candidates that have been classified as positive by Histograms of Oriented Gradient(HOG) with Multiple Kernel Learning(MKL). For each positive candidate, a region around the candidate is divided into several subregions based on the direction of the candidate, then each subregion is divided into 12 bins with a fixed length; and finally the TC, a histogram, is built according to the positions of the positive candidates in 8 consecutive frames. In order to benefit from both the appearance and context information, we use MKL to combine TC and HOG. To evaluate the effect of TC, we use the publicly available CLIF 2006 dataset, and label the vehicles in 102 frames which are 2672 × 1200 subregions that contain expressway of the original 2672 × 4008 images. The experiments demonstrate that the proposed TC is useful to remove the false positives that are away from the road, and the combination of TC and HOG with MKL outperforms the use of TC or HOG only.