Academia.eduAcademia.edu

Figure 23 – uploaded by Dan Schonfeld

See full PDFdownloadDownload figure
DAN SCHOMNFELD was born in Westchester, Pennsylvania, on June 11, 1964. He received the B.S. in electrice engineering and computer science from the University of California, Berkeley, California, in 1986 and the M.S and Ph.D. in electrical and computer engineering from the Johns Hopkins University, Baltimore, Maryland, i 1988 and 1990, respectively. Since September 1990 he has been with the Department of Electrical Engineerin and Computer Science at the University of Illinois, Chicago, Illinois, where he is currently an associate professo: He has authored over 30 technical papers in various journals and conferences. He has served as an associate edito of the IEEE Transactions on Image Processing on Nonlinear Filtering. He currently serves as an associate edito of the IEEE Transactions on Signal Processing on Multidimensional Signal Processing and Multimedia Signa Processing. His current research interests are in one- and multidimensional signal processing, speech processin¢ image processing, video processing, computer vision, pattern recognition, medical imaging, and multimedi communication networks.

Figure 23 DAN SCHOMNFELD was born in Westchester, Pennsylvania, on June 11, 1964. He received the B.S. in electrice engineering and computer science from the University of California, Berkeley, California, in 1986 and the M.S and Ph.D. in electrical and computer engineering from the Johns Hopkins University, Baltimore, Maryland, i 1988 and 1990, respectively. Since September 1990 he has been with the Department of Electrical Engineerin and Computer Science at the University of Illinois, Chicago, Illinois, where he is currently an associate professo: He has authored over 30 technical papers in various journals and conferences. He has served as an associate edito of the IEEE Transactions on Image Processing on Nonlinear Filtering. He currently serves as an associate edito of the IEEE Transactions on Signal Processing on Multidimensional Signal Processing and Multimedia Signa Processing. His current research interests are in one- and multidimensional signal processing, speech processin¢ image processing, video processing, computer vision, pattern recognition, medical imaging, and multimedi communication networks.

Related Figures (25)

DOI: 10.1006/jvci.1999.0432 https: ://www. researchgate. net/publication/222646180
DOI: 10.1006/jvci.1999.0432 https: ://www. researchgate. net/publication/222646180
FIG. 1. VORTEX ina distributed environment.
FIG. 1. VORTEX ina distributed environment.
Following a quantization stage, a variable length (VL) entropy-coding— Huffman coding— is performed at the next stage in order to reduce the symbol redundancy. The decoder performs the inverse operations— VL decoding, inverse DCT, motion decompensation— in order to obtain a lossy reconstruction of the original video sequence.
Following a quantization stage, a variable length (VL) entropy-coding— Huffman coding— is performed at the next stage in order to reduce the symbol redundancy. The decoder performs the inverse operations— VL decoding, inverse DCT, motion decompensation— in order to obtain a lossy reconstruction of the original video sequence.
FIG. 4. Object search and tracking.
FIG. 4. Object search and tracking.
FIG.5. Exhaustive and VORTEX’s approach to tracking for video retrieval.
FIG.5. Exhaustive and VORTEX’s approach to tracking for video retrieval.
FIG. 7. Hits inside the decision area. x, Valid hits; O, center of ADA.
FIG. 7. Hits inside the decision area. x, Valid hits; O, center of ADA.
FIG. 8. Object outside slices. (a) Reference picture; (b) current picture. According to MPEG prediction methodology [31], nonslice macroblocks are not encoded. The decoder simply replicates the corresponding macroblocks from the reference picture. Figure 8 illustrates an example where the determined position of the object of interest in the
FIG. 8. Object outside slices. (a) Reference picture; (b) current picture. According to MPEG prediction methodology [31], nonslice macroblocks are not encoded. The decoder simply replicates the corresponding macroblocks from the reference picture. Figure 8 illustrates an example where the determined position of the object of interest in the
FIG. 10. Occlusion detection and feedback. x, Valid hits; O, origins of motion vectors in current picture; —~, original ADA; —, Recomputed ADA.
FIG. 10. Occlusion detection and feedback. x, Valid hits; O, origins of motion vectors in current picture; —~, original ADA; —, Recomputed ADA.
FIG.11. Motionvector filtering. O, Origins of unfiltered motion vectors; A, Origins of filtered motion vectors.
FIG.11. Motionvector filtering. O, Origins of unfiltered motion vectors; A, Origins of filtered motion vectors.
ios eee A mak PeGatet te SS etre: SECTS SNS: | SNL eRe es cE ieee eee oate Tee ee oe Sn oe ‘If the current picture is isa 1P picture, let us assume that the object i is present i in nthe refer. ence picture at some position that we shall denote by prep. This position might have been
ios eee A mak PeGatet te SS etre: SECTS SNS: | SNL eRe es cE ieee eee oate Tee ee oe Sn oe ‘If the current picture is isa 1P picture, let us assume that the object i is present i in nthe refer. ence picture at some position that we shall denote by prep. This position might have been
FIG. 13. Tracking prediction in I pictures. O, Computed object position in current picture.
FIG. 13. Tracking prediction in I pictures. O, Computed object position in current picture.
FIG. 14. Hit distribution inside the decision area. O, ADA center; x, Valid hits. Figure 15 illustrates the computation and use of the ADA for acquiring the valid hits. Let us assign to the current picture an index k in the picture sequence and to its reference picture index k — 1. For simplicity, k is used here to index only pictures that may be used as reference by the motion compensation process; i.e., I or P pictures. It is possible that there are additional B pictures placed between any two consecutive I or P pictures. Let us
FIG. 14. Hit distribution inside the decision area. O, ADA center; x, Valid hits. Figure 15 illustrates the computation and use of the ADA for acquiring the valid hits. Let us assign to the current picture an index k in the picture sequence and to its reference picture index k — 1. For simplicity, k is used here to index only pictures that may be used as reference by the motion compensation process; i.e., I or P pictures. It is possible that there are additional B pictures placed between any two consecutive I or P pictures. Let us
FIG.15. Adaptive decision area. (a) Computation; (b) valid hit acquisition. x, Valid hits; O, origins of motion vectors. assume that the ADA in picture k — 1 is known and it is us used in computing the new object position in the current ed to collect hits that are to be picture k. All the valid hits in picture k — 1 have corresponding motion vectors whose origins are in the current picture k (see Fig. 15a). At a later time, the current picture k will the following P picture having index k + 1 (see Fig. 15b). T become a reference picture for hus, we shall compute an ADA in the current picture k that will be used to collect the valid hits used for updating the object attributes (position and size) in picture k + 1. This A DA must take into account the information provided by the distribution of the origins of tl corresponding to the valid hits in picture k — 1. he motion vectors in picture k,
FIG.15. Adaptive decision area. (a) Computation; (b) valid hit acquisition. x, Valid hits; O, origins of motion vectors. assume that the ADA in picture k — 1 is known and it is us used in computing the new object position in the current ed to collect hits that are to be picture k. All the valid hits in picture k — 1 have corresponding motion vectors whose origins are in the current picture k (see Fig. 15a). At a later time, the current picture k will the following P picture having index k + 1 (see Fig. 15b). T become a reference picture for hus, we shall compute an ADA in the current picture k that will be used to collect the valid hits used for updating the object attributes (position and size) in picture k + 1. This A DA must take into account the information provided by the distribution of the origins of tl corresponding to the valid hits in picture k — 1. he motion vectors in picture k,
where Vo?(k) is given by
where Vo?(k) is given by
FIG. 17. Limiting function h(.). The limiting function h(.) is depicted in Fig. 17. The threshold ¢ is used in the process of computing the ADA to eliminate very small variations that are due to noise and numerical s is the sign of the ADA variation, and K apa is a scaling constant that controls the rate of variation of the ADA. The signs is determined based on the difference between the variances of the origins of valid motion vectors in the current picture and the reference picture; i.e., s = sgn[o?(k) — o?(k — 1)]. Figure 16 illustrates the computation of the average one-step temporal gradient of the standard deviations of motion vector origins Vo?(k), by using a moving average across M + 1 reference pictures (P).
FIG. 17. Limiting function h(.). The limiting function h(.) is depicted in Fig. 17. The threshold ¢ is used in the process of computing the ADA to eliminate very small variations that are due to noise and numerical s is the sign of the ADA variation, and K apa is a scaling constant that controls the rate of variation of the ADA. The signs is determined based on the difference between the variances of the origins of valid motion vectors in the current picture and the reference picture; i.e., s = sgn[o?(k) — o?(k — 1)]. Figure 16 illustrates the computation of the average one-step temporal gradient of the standard deviations of motion vector origins Vo?(k), by using a moving average across M + 1 reference pictures (P).
Euclidean Positional Deviation from Exhaustive Matching TABLE 1
Euclidean Positional Deviation from Exhaustive Matching TABLE 1
Average Execution Time per Picture
Average Execution Time per Picture
FIG.18. Snapshots of video tracking for retrieval of multiple templates (cow— black frame and train— white frame).
FIG.18. Snapshots of video tracking for retrieval of multiple templates (cow— black frame and train— white frame).
FIG. 20. VORTEX: A ffine transformations 3. FIG. 19. VORTEX: A ffine transformations 1.
FIG. 20. VORTEX: A ffine transformations 3. FIG. 19. VORTEX: A ffine transformations 1.
FIG.21. Motion vector filtering.
FIG.21. Motion vector filtering.
FIG. 22. VORTEX: Occlusion detection. The third phase of simulations consisted of testing the occlusion detection capability of VO of occ The p of the motion and A the video sequence, as discussed separation of the mo detection depicted in Fig. 22b (bo two rows represent images with tl illustrate the effect that occlusion detec No tion has on the tracking q increasing and decreasing occlusion is determined by the way threshold parameters are set. In general, even though some of its area is partial RTEX. A K-means clustering algorithm with K =2 was used for detection of the tracked and the occluding objects. The decision to accept two motion clusters as indicative usion or severe noise in the area of the object is done by considering the distance between the corresponding magnitudes and phases of the two cluster prototypes. For this purpose, we used a distance threshold in both magnitude and phase for the decision process. hase difference threshold parameter was set to 80° and the magnitude difference threshold parameter to 50%. Thus, if the two motion prototypes differ in motion direction by more than 80° and the relative difference in magnitudes is more than 50%, the algorithm decides in favor of two clusters as opposed to one. This in turn DA data in the reference picture(s) and in Section 5.2. The parame tion prototypes as well as the selection o are adjustable. The two image rows in Fig. 22 contain snaps algorithm without occlusion detection depicted in Fig. 22a (t op ttom row) fora video sequence. 1 he same index in the identica triggers the reassessment thus adjusts the tracking of ters that control the cluster f the correct motion cluster ts of the operation of the row) and with occlusion The video snapshots in the video sequence, in order to uality. The detection of both the cluster- discriminating itis acceptable that the object of interest be tracked y occluded by another obj ect. This situation occurs in the case of increasing partial occlusion. Thus, for some interval of time, both objects are tracked together, with the occluding object covering some of the ADA area. However, detection of decreasing occlusion and separation of the target motion cluster is critical.
FIG. 22. VORTEX: Occlusion detection. The third phase of simulations consisted of testing the occlusion detection capability of VO of occ The p of the motion and A the video sequence, as discussed separation of the mo detection depicted in Fig. 22b (bo two rows represent images with tl illustrate the effect that occlusion detec No tion has on the tracking q increasing and decreasing occlusion is determined by the way threshold parameters are set. In general, even though some of its area is partial RTEX. A K-means clustering algorithm with K =2 was used for detection of the tracked and the occluding objects. The decision to accept two motion clusters as indicative usion or severe noise in the area of the object is done by considering the distance between the corresponding magnitudes and phases of the two cluster prototypes. For this purpose, we used a distance threshold in both magnitude and phase for the decision process. hase difference threshold parameter was set to 80° and the magnitude difference threshold parameter to 50%. Thus, if the two motion prototypes differ in motion direction by more than 80° and the relative difference in magnitudes is more than 50%, the algorithm decides in favor of two clusters as opposed to one. This in turn DA data in the reference picture(s) and in Section 5.2. The parame tion prototypes as well as the selection o are adjustable. The two image rows in Fig. 22 contain snaps algorithm without occlusion detection depicted in Fig. 22a (t op ttom row) fora video sequence. 1 he same index in the identica triggers the reassessment thus adjusts the tracking of ters that control the cluster f the correct motion cluster ts of the operation of the row) and with occlusion The video snapshots in the video sequence, in order to uality. The detection of both the cluster- discriminating itis acceptable that the object of interest be tracked y occluded by another obj ect. This situation occurs in the case of increasing partial occlusion. Thus, for some interval of time, both objects are tracked together, with the occluding object covering some of the ADA area. However, detection of decreasing occlusion and separation of the target motion cluster is critical.
FIG. 23. VORTEX: Occlusion detection (magnified).
FIG. 23. VORTEX: Occlusion detection (magnified).
DAN LELESCU was bom in Lugoj, Romania, on June 3, 1967. He received the B.S./M.S. in electrical engineering from the Technical University “Politehnica” of Timisoara, Romania, in 1991. Currently, he is a Ph.D. candidate in the Department of Electrical Engineering and Computer Science at the University of Illinois, Chicago, Illinois. His research areas of interest include multimedia signal processing and communications, signal detection and estimation, image and video processing, and computer vision.
DAN LELESCU was bom in Lugoj, Romania, on June 3, 1967. He received the B.S./M.S. in electrical engineering from the Technical University “Politehnica” of Timisoara, Romania, in 1991. Currently, he is a Ph.D. candidate in the Department of Electrical Engineering and Computer Science at the University of Illinois, Chicago, Illinois. His research areas of interest include multimedia signal processing and communications, signal detection and estimation, image and video processing, and computer vision.
Related topics:
Data CompressionMotion AnalysisVideo CompressionVisual SearchVideo SearchMultimedia information retrievalImage RetrievalVideo RetrievalTemplate matchingMultiple Object Tracking
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved