Air-ground vehicle detection with a reduced object category specific visual dictionary

Abstract

Nowadays ground vehicle detection on airborne platforms is becoming very important for intelligent visual surveillance applications. Object detection using cascade structured classifiers is booming fast in recent decade, and very successful in real-time applications. However, most of them apply a sliding window on multi-scaled images which commonly need heavy computational expense, therefore, are only suitable for using simple features. In this paper, a biologically inspired object detection algorithm is proposed, which exploits image patch based feature learning and visual saliency detection. The image patch based local features are learnt by unsupervised learning to generate an object category specific visual dictionary. Visual saliency detection is performed to extract candidate object regions from a whole image using the learnt local features. Instead of a sliding window, a candidate object region is sent to an object classifier only when its features are salient on the whole image. Since the number of candidate object regions decreases dramatically, it allows to utilize much complex features to represent object images so that it can increase the descriptive capability of the learnt features. The experimental results on practical vehicle image datasets indicate that less computational expense and good detection performance can be achieved.