Parts-probability-based vehicle detection

Abstract

Detecting vehicles is important in aerial surveillance. Traditional methods used classifiers to detect vehicles, but a single classifier was limited to detecting vehicles of only one intensity and orientation. Therefore, the task required the use of multiple classifiers of different intensities and orientations. To solve this problem, we first used a latent Dirichlet allocation (LDA) model that improved the previous approaches to vehicle detection. Previous text modeling approaches have been generative. They could be used to build probability models of vehicles in different intensities and from various orientations simultaneously using unlabeled data. Using a probability model, we can detect vehicles in a region with high probability. Next, we used a parts-probability model that improves the LDA model. The model effectively encodes spatial structure among visual words by adding spatial relationships among vehicle parts as priors of words. A parts probability model represents a vehicle hierarchically according to parts appearances and a vehicle’s features within the parts to enforce spatial coherency. Then, we used our model to detect vehicles from a collection of images and demonstrate its performs more effectively.