Air-ground vehicle detection using local feature learning and saliency region detection

Moving vehicle detection is very important for urban traffic surveillance and situational awareness on the battlefield. Algorithms with cascade structure like Adaboost are booming in the recent decade, and successful in realtime application. However, most of them use a sliding window protocol on multi-scale images which involves heavy computing. Therefore, they are only suitable for simple feature. In this paper, a biologically inspired method is proposed. We learn patch-based features for vehicle detection by unsupervised learning, and then employ a visual saliency step after feature extraction. Instead of sliding window, a candidate region is sent to classifier only if its features are "salient" on whole image. As the number of candidate regions decreases dramatically, it allow us to utilize complex feature to increase description ability. Experimental result indicates less computational expense and good performance.