GEOMETRIC COMPARISON 0F KEY FRAME SELECTION TECHNIQUES IN 3D RECONSTRUCTION BASED ON AERIAL VIDEO IMAGERY

The large number of frames in aerial video sequences, makes it to use all of them for any further application. An abstracting process like key frame selection can be useful. Since the selected frames play and important role on the quality of generated products, especially in geometrics parameters, in this paper a comparison between different key frame selection methods is presented. The capability of each method is investigated based on geometric accuracy for 3D reconstruction of scenes. Five common key frame extraction techniques presented by Gibson, Seo, Pollefeys, Knorr and Ahmed are selected and implemented on a dataset. Before implementation, a radiometric processing is done on all 2706 frames of video and the 30 blur frames are removed using the BluM algorithm. After that the five mention methods are implemented on the remaining frames of video. The key frames from all the five methods are compared based on number and overlapping of frames. Also key frames are used for 3D reconstruction of scene and accuracy of that is used as metric for geometrical evaluation. Results show that the method of Seo selects the min number of key frames, 163 and the key frames by method of Gibson have more mutual area with each other, %78.02 in mean. The results of 3D reconstruction show that accuracy of geometric alignment is almost the same in all methods, so a method that chooses fewer frames will have a higher processing speed and therefore a better fit.