Moving vehicle detection in wide area motion imagery is a challenging task due to the large motion of the camera and the small number of pixels on the target. At the same time, this task is very important for surveillance applications, and the result can be used for urban traffic management, accident and emergency responder routing. Also, the effectiveness of the context in object detection task can be further explored to increase target tracking accuracy. In this paper, we propose to use Spatial Context(SC) to improve the performance of the vehicle detection task. We first model the background of 8 consecutive frames with median filter, and get candidates by using background subtraction. The SC is built based on the candidates that have been classified as positive by Histograms of Oriented Gradient(HOG) with Multiple Kernel Learning(MKL). The region around each positive candidate is divided into m subregions with a fixed length l, then, the SC, a histogram, is built based on the number of positive candidates in each region. We use the publicly available CLIF 2006 dataset to evaluate the effect of SC. The experiments demonstrate that SC is useful to remove false positives, around which there are few positive candidates, and the combination of SC and HOG with multiple kernel learning outperforms the use of SC or HOG only.

Date Published: 28 May 2013
PDF: 9 pages
Proc. SPIE 8751, Machine Intelligence and Bio-inspired Computation: Theory and Applications VII, 875105 (28 May 2013); doi: 10.1117/12.2015967

Published in SPIE Proceedings Vol. 8751:
Machine Intelligence and Bio-inspired Computation: Theory and Applications VII
Misty Blowers; Olga Mendoza-Schrock, Editor(s)