tailieunhanh - Machine Learning and Robot Perception - Bruno Apolloni et al (Eds) Part 9

Tham khảo tài liệu 'machine learning and robot perception - bruno apolloni et al (eds) part 9', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | 5 Efficient Incorporation of Optical Flow 195 Fig. . Processing time averaged over 7-window-frames vs frames for the original sequence left for the sequence subsampled by 6 in time right The polygonal tracker with its ability to utilize various region-based descriptors could be used for tracking textured objects on textured backgrounds. A specific choice based on an information-theoretic measure 41 196 G. Unal et al. Fig. . A flatworm in a textured sea terrain 15 frames are shown left-right topbottom . Polygonal tracker successfully tracks the flatworm whose approximation uses high order moments of the data distributions leads the image based integrand f in Eq. 17 to take the form f i uj -vj G J I - uj G J I - vj with functions G chosen j A 2 2 -r 2 2 for instance as G1 e and G2 S e . When the correction step of our method involves the descriptor f Just given with a adaptive number of vertices a flatworm swimming in the bottom of the sea could be captured through the highly textured sequence by the polygonal tracker in Fig. . The speed plots in Fig. depict the speeds for the tracker with and without prediction. The figure on the right is for the original sequence whose plot is given on the left which is temporally subsampled by two. Varying the number of vertices to account for shape variations of the worm slows down the tracking in general. However the tracker with prediction still performs faster than the tracker without prediction as expected. The difference in speeds becomes more pronounced in the subsampled sequence on the left. Similarly a clownfish on a host anemone shown in Figure could be tracked in a highly textured scene. The continuous trackers we have introduced in this study do not provide a continuous tracking in either of these examples and they split leak to background regions and lose track of the target completely. 5 Efficient Incorporation of Optical Flow 197 Fig. . Processing time averaged over 7-window-frames vs .