A thresholding scheme of eliminating false detections on vehicles in wide-area aerial imagery

Post-processings are usually necessary to reduce false detections on vehicles in wide-area aerial imagery. In order to improve the performance of vehicle detection, we propose a two-stage scheme, which consists of a thresholding method by constructing a pixel-weight based thresholding policy to classify pixels in the greyscale feature map of an automatic detection algorithm followed by morphological filtering. We use two aerial videos for performance evaluation, and compare the automatic detection results with the ground-truth objects. We compute average F-score and percentage of wrong classifications towards six detection algorithms before and after applying the proposed scheme. We measure the variation of overlap ratios from detections to objects, and establish sensitivity analysis to evaluate the performance of proposed scheme by combining it on each of two representative algorithms. Simulation results verify both validity and efficiency of the proposed thresholding scheme, also display the difference of detection performance between datasets and among algorithms.