Polana, R. and Nelson, R. 1994. Low level recognition of human motion. In Proc. of the IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, pp. 77--82.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....domain, the targets small size, deformations and nearly continual partial occlusions limit the applicability of feature based approaches. Using features to help initialize a stronger model is a powerful tracking technique that has been used by many researchers, e.g. with weak models for people in [10, 11, 12, 13, 14] and strong models for vehicles in [8, 9] Models permit restricting the search area for likely features, thereby allowing increased sensitivity without significantly increasing the chance for false alarms. However, these systems require both a reasonably large number of pixels on target and model ....

....path voting process to detect regions of similar mean flow. It is assumed, however, that the target object s motion can be described with a constant vector. Again large targets with minimal occlusion are implicitly presumed. There have been many papers on tracking and analyzing human motion, e.g. [10, 11, 12, 13, 14, 20]. Motion parameter analysis has also been used to distinguish targets. For example, 21] uses motion parameters as the primary method to distinguish between human and vehicle. However, it is presumed that targets are not occluded and consist of many hundreds or thousands of pixels. This limits its ....

R. Polana and R. Nelson, "Low level recognition of human motion," in Workshop on Non-rigid Motion, pp. 77-82, Nov. 1994.

....inherently needs to understand the temporal aspects of human motion. Boyd and Little [14] analyzed the moments of the optical flow of walking people to recognize them by their gait. They used maximum entropy spectrum estimation to analyze the short and noisy feature sequences. Polana et al. [15] used nearest neighbor search through spario temporal templates of observed independent motion for gait recognition. Both works emphasize the recognition rather than the modeling of temporal patterns. Bregler [7] used HMMs with states associated with dynamical systems to model human dynamics. On ....

R. Polana and R. Nelson, Low level recognition of human motion (or how to get your man without finding his body parts), in Motion of Non-Rigid and Articulated Objects, Proceedings of lEEE Workshop on, 1994, pp. 77 82.

....10, 11] In [2] the Condensation based Trajectory Recognition (CTR) is proposed which can be seen as a generalization of HMM s. Alternative methods based on template techniques convert an image sequence into a static shape pattern. The most commonly used features for this technique are 2D meshes [8]. Furthermore, these approaches also comprise motion energy images (MEI) and motion history images (MHI) which are presented in [4] A detailed review of human motion analysis can be found in [1] The outline of this paper is given as follows. In the next section we introduce the tracking ....

R. Polana and R. Nelson. Low Level Recognition of Human Motion (Or How to Get Your Man Without Finding his Body Parts). IEEE Computer Society Workshop on Motion of Non-Rigid and Articulated Objects, pp. 77-82, 1994.

....segment the image into blobs. Some of these blobs represent the hands and feet of the subject yielding some sort of limb representation. If tracking prepares data for recognition the task is usually to represent data in an appropriate manner. An example of this is published by Polana and Nelson [119] where flow information and down sampling are used to represent image information in a compact manner which are processed by a classifier to recognise six different classes, e.g. walking and running. Independently on the context of tracking three common aspects can be identified. Nearly every ....

....between images from a sequence must originate from the movements of the subject. Two subclasses may be introduced: subtraction and flow. Subtraction is widely used by simply subtracting the current image from the previous image in a pixel by pixel fashion, using either the intensity values [119] or the gradients [4] An improved version is to use three consecutive images instead of two [77] The result reflects movements (and noise) between the images unless the subject has the same intensity colour as the background. The use of background subtraction is very popular. If the scene is ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low Level Recognition of Human Motion. In Workshop on Motion of Non-Rigid and Articulated Objects, Austin, Tx, USA, October 1994.

....ffl 2 D approaches without explicit shape models ffl 2 D approaches with explicit shape models ffl 3 D approaches The first approach bypasses a pose recovery step altogether and describes human movement in terms of simple low level, 2 D features from a region of interest. Polana and Nelson [22] refer to getting your man without finding his body parts . Models for human action are described in statistical terms based on low level features. Foreground regions are typically obtained by skin color detection or background subtraction from which features based on shape [2] 6] 12] texture ....

....finding his body parts . Models for human action are described in statistical terms based on low level features. Foreground regions are typically obtained by skin color detection or background subtraction from which features based on shape [2] 6] 12] texture [7] 8] 21] or motion [4] 9] [22] are extracted. In some cases [12] 21] the requirement of a separate foreground segmentation is relaxed by the employment of window search procedures. The second approach uses explicit a priori knowledge of how the human body (or hand) appears in 2 D, taking essentially a model and view based ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Motion of Non-Rigid and Articulated Objects, pages 77-- 82, Austin, 1994.

....based on active sensing. 3 2 D approaches without explicit shape models One general approach to the analysis of human movement has been to bypass a pose recovery step altogether and to describe human movement in terms of simple low level, 2 D features from a region of interest. Polana and Nelson [65] refer to getting your man without finding his body parts . Models for human action are then described in statistical terms derived from these lowlevel features, or by simple heuristics. The approach without explicit shape models has been especially popular for applications of hand pose ....

....features is to superimpose a grid on the interest region, after a possible normalization of its extent. In each tile of the grid a simple feature is computed, and these features are combined 7 to form a K Theta K feature vector to describe the state of movement at time t. Polana and Nelson [65] use the sum of the normal flow (see Figure 1) Yamamoto et al. 86] use the number of foreground pixels and Takahashi et al. 78] define an average edge vector for each tile. Both Darell and Pentland [19] and Kjeldsen and Kender [44] use the image pixels directly as input. The work by Darell and ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. of IEEE Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, Austin, 1994.

....subsequently. It is unclear whether these applications will materialize; the 2 D head tracking application provides modest compression gains and is specific to scenes with human faces; the 3 D head (or FIG. 1. Detection of periodic activity using low level motion features (from Polana and Nelson [65], c # 1994 IEEE) body) tracking application has not been solved satisfactorily yet. See Aizawa and Huang [2] for a good overview. In all the applications discussed above, a nonintrusive sensory method based on vision is preferable over a (in some cases a not even feasible) method that relies ....

....on active sensing. 3. 2 D APPROACHES WITHOUT EXPLICIT SHAPE MODELS One general approach to the analysis of human movement has been to bypass a pose recovery step altogether and to describe human movement in terms of simple low level, 2 D features from a region of interest. Polana and Nelson [65] refered to getting your man without finding his body parts. Models for human action are then described in statistical terms derived from these low level features or by simple heuristics. The approach without explicit shape models has been especially popular for applications of hand pose ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson, Low level recognition of human motion, in Proc. of IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, 1994, pp. 77--82.

....at maximum, the power of higher dimensional descriptions. It is then essential to count on experimental results that support their efficiency and demonstrate their power. The importance of motion recognition problems is evidenced by the increasing attention they have received in recent years [24, 25, 1, 16, 20, 19]. One of the main areas of research is the analysis of humans in motion. There are numerous domains that motivate the research in this area: video surveillance, human computer interaction, athletics, dance, robot motion, among others. 1 Motion of objects is much more semantically rich than ....

....knowledge model, for example [5, 11, 21, 22] The most relevant group of approaches related to ours, as in [6] are those that try to achieve recognition of the motion directly from the sequence of images. No explicit use of the static images is done besides what is needed to represent the motion [20, 2, 6]. We give a brief overview of their fundamentals. According to [20] recognition of repetitive motion can be achieved on the basis of bottom up processing, without identifying specific parts or classification of the object. Their approach needs to estimate local velocity to undo the translation ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

....articulated models comprised of 2D or 3D solid primitive, sometimes accounting for self occlusion by having an explicit object model. Shape models were used by [2] for human tracking. The most relevant motion recognition approaches related to our work are those that employ view based models [3, 9, 19]. In particular, 9] uses motion history images (MHI) and motion energy images (MEI) temporal templates that are matched using a nearest neighbor approach against examples of given motions already learned. The main problem of this method is the requirement of having stationary objects, and the ....

....learned. The main problem of this method is the requirement of having stationary objects, and the insufficiency of the representation to discriminate among similar motions. Motion analysis techniques have had the problem that registration of useful, filtered information is a hard labor by itself [9, 4, 19]. Our system provides a functional front end that supports such tasks. 3 Basic Approach A diagram of our approach is illustrated in Fig. 1. The first stage of the algorithm is based on the background subtraction methods of [30] The system is initialized by acquiring statistical measurements of ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

....articulated models comprised of 2D or 3D solid primitive, sometimes accounting for self occlusion by having an explicit object model. Shape models were used by [2] for human tracking. The most relevant motion recognition approaches related to our work are those that employ view based models [3, 9, 19]. In particular, 9] uses motion history images (MHI) and motion energy images (MEI) temporal templates that are matched using a nearest neighbor approach against examples of given motions already learned. The main problem of this method is the requirement of having stationary objects, and the ....

....learned. The main problem of this method is the requirement of having stationary objects, and the insufficiency of the representation to discriminate among similar motions. Motion analysis techniques have had the problem that registration of useful, filtered information is a hard labor by itself [9, 4, 19]. Our system provides a functional front end that supports such tasks. 3 Basic Approach A diagram of our approach is illustrated in Fig. 1. The first stage of the algorithm is based on the background subtraction methods of [30] The system is initialized by acquiring statistical measurements of ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

....repetitive motion is a cue strong enough to enable the recognition of the motion with only highlevel analysis. This is achieved by matching the motion against a set of known spatiotemporal motion templates. Prior to the matching procedure a number of essential preprocessing steps must take place [PN94] These steps are the segmentation of the moving actor, and the normalization of its motion, both in the space and time dimensions, so that it can be compared to the motion templates. The first step makes sure that the system isolates the segments of the image where a moving actor exists, using ....

Ramprasad Polana and Randal Nelson. Low Level Recognition of Human Motion (or How to Get Your Man Without Finding His Body Parts. In IEEE Computer Society Workshop on Motion of Nonrigid and Articulate Objects, Austin, TX, USA, 1994.

....high resolution support from the data is available. As we will discuss, we have found it is often difficult to separate the tracking from the boundary estimation problem. Therefore, some people tracking methods that require accurate object boundary detection based on intensity[1] and optical flow[12] are difficult to apply in our tracking task. Further the difficulty in segmenting colliding children and their rapid changes in appearance and motion prevents the use of differential motion estimators that use smooth or planar motion models [3] and tracking techniques that require reasonably ....

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. Work. Non-Rigid Motion, pages 77--82, Nov. 1994.

.... An increasing interest in the recognition of human motion and action using computer vision has appeared, with much emphasis on real time computability (for example Davis 1997; Bradski 1998; Bradski 1999; Freeman 1998; Ju 1996; Akita 1984; Darrell 1994; Cuttler 1998; Hogg 1983; Little 1995; Polana 1994; Rohr 1994; Haritaoglu 1998; Pinhanez 1999; Yamato 1992; Bregler 1997; Wren 1995; Cham 1998 and surveys in Cedres 1995; Shah 1997; Moeslund 1999a; Moeslund 1999b; Gavrila 1999) In particular, tracking surveillance systems, human computer interfaces, and entertainment domains have a heightened ....

Polana, R. and R. Nelson. Low level recognition of human motion. In IEEE Wkshp. On Nonrigid and Articulated Motion, 1994.

....joint angles and translational changes of the center of coordinates (see [8, 10, 170] Such representations pose many problems for the inverse problem, that is, to recognize and classify observed movement. Rohr [145] used 3D cylindrical models to recognize walking movements. Polana and Nelson [139] used the distribution of motion in the area of the action to Chapter 2 Representing Action 35 recognize among seven types of activities such as running and jogging. Gavrila and Davis [53] used 3D models to track upper body movements. Recent work by Bregler and Malik [29] proposes exponential ....

R. Polana and R. Nelson. "Low Level Recognition of Human Motion", Proc. of IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, Texas, pp. 77-82. November. 1994.

....or feasible to try to recover 3 D body pose from 2 D image sequences for the purpose of recognizing human movement. An alternative approach is to work directly with 2 D features derived from the images, using some form of 2 D model [Goddard, 1994] Guo et al. 1994] Leung and Yang, 1995] or not [Polana and Nelson, 1994] [Darrell and Pentland, 1993] Recognition systems using 2 D model free features have been able to claim early successes in matching human movement patterns. For constrained types of human movement (such as walking parallel to the image plane, involving periodic motion) many of these features ....

....using 2 D model free features have been able to claim early successes in matching human movement patterns. For constrained types of human movement (such as walking parallel to the image plane, involving periodic motion) many of these features have been successfully used for classification, as in [Polana and Nelson, 1994]. This may indeed be the easiest and best solution for several applications. But we find it unlikely that reliable recognition of more unconstrained and complex human movement 1 The tracking results described in this paper are also available as video clips from our home pages. e.g. humans ....

R. Polana and R. Nelson, "Low Level Recognition of Human Motion," IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, TX, 1994.

....may be recognized by motion information alone. Work with very low resolution American Sign Language images by Sperling et al. 13] further supports the notion that in many domains a full geometric reconstruction of the moving object is unnecessary for recognition. For example, Polana and Nelson [10] use low level features of motion to recognize periodic motions such as walking. A number of researchers have developed novel representations for motion trajectories that are useful in gesture recognition. Gould and Shah [5] show the analysis of motion trajectories to identify event boundaries. ....

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. of the Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, Austin, Texas, Nov. 1994.

....missing in our representation system is a mechanism to specify the intensity of an action. For computer theater purposes, the difference between talking and shouting is crucial. 5. 2 Recognition Research in visual action recognition has been restricted to recognizing human body movements ([13, 31, 29, 10, 4]) 18, 37, 12] are among the few works which actually examined some of the issues related to understanding actions and their effects in the world. Bruce Blumberg s dog mentioned above uses the recognition capabilities of ALIVE ( 22] to react to commands like pointing , sitting , and catch the ....

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. of IEEE Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, Austin, Texas, November 1994.

....actions, and that the expectation about the next action plays a fundamental role in the recognition and segmentation processes ( 13, 22] Thus, a non timed version of the script would, theoretically, give most of the information needed. Some research has been done in recognizing human movements [24, 18, 1, 16, 8] and in action recognition [10] though most methods were developed for situations much more constrained than those found in normal TV studios. However, we believe that the use of approximate models can significantly facilitate the provision of the contextual information which is essential for ....

R. Polana and R. Nelson, "Low Level Recognition of Human Motion," Proc. of IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, Texas, November 11--12, pp. 77--82, 1994.

....is the requirement that there be individual features or properties that can be extracted from each frame of the image sequence. These approaches accomplish motion understanding by recognizing a sequence of static configurations. Alternatively, there is the work on direct motion recognition [12, 15, 16, 2, 8]. These approaches attempt Frame 0 20 40 Figure 2: Example of someone sitting. Top row contains key frames; bottom row is cumulative motion images starting from Frame 0. to characterize the motion itself without any reference to the underlying static images. Of these techniques, the work of ....

Polana, R. and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Non-rigid and Articulated Motion, 1994.

....a considerable task. Part of the difficulties comes from the problem of recognizing actions and body movements. Some research has been done in recognizing movements as, for instance, the works of Tsotsos et al. 23] and Rohr ( 19] and on theoretical grounds by Allen ( 1] Polana and Nelson ([17]) and Israel et al. 8] Research in action recognition has been more rare (see the work of Kuniyoshi and Inoue [10] though we believe the use of approximate models can significantly facilitate the provision of the contextual information which is essential for action recognition. Another ....

R. Polana and R. Nelson, "Low Level Recognition of Human Motion," Proc. of IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, Texas, November 11--12, pp. 77--82, 1994.

....model we will develop attempts to span as wide an angular range as possible using a single, low order representation of the appearance of the motion. However, when not possible, our model can also accommodate discrete regions or aspects. Finally there is the work on direct motion recognition [18, 21, 23, 3]. These approaches attempt to characterize the motion itself without any reference to the underlying static images. Of these techniques, the work of Black and Yacoob [3] is the most relevant to the results presented here. The goal of their research is to recognize human facial expressions as a ....

Polana, R. and R. Nelson, "Low Level Recognition of Human Motion," IEEE Workshop on Non-rigid and Articulated Motion, Austin, 1994.

.... illustrated the use of simple color histograms for retrieving images from a diverse database, and Syeda Mahmood has shown how a combination of color and texture features can speed up selection of items of interest in photos [8] Texture has also been shown to be powerful for recognition of motions [9]. 1.1 Texture: beyond the traditional definition There is much more texture in the world than most people realize. Texture is ubiquitous; it is felt on the tiny surface of a shriveled pea, can be heard in the interwoven melodies of a fugue, can be seen in the rocking motion of a boat, and even ....

R. Polana and R. Nelson, "Low level recognition of human motion," in IEEE Workshop on Motion of Non-rigid and Articulated Objects, (Austin, TX), 1994.

....In related work, Darrell and Pentland [3] use dynamic time warping and normalized correlation to match the interpolated responses of several learned image templates. Murase and Nayar [8] parameterize multiple eigenspaces over pose and illumination angle for object recognition. Polana and Nelson [10] match low level templates of spatiotemporal motion to recognize periodic human motions in image sequences. Cui and Weng [2] use learned decision boundaries to recognize sequences of vector quantized images of hands. Starner and Pentland [15] extract the position and dominant orientation of both ....

....an approach that can exploit multiple models simultaneously, where the type of models might be quite distinct. Model types useful for characterizing images in an image sequence might include eigenvector decomposition of sets of images [16] orientation histograms [4] peak temporal frequencies [10], tracked position of objects in the frame, and optic flow field summaries. 2.2 State based descriptions In previous work [1] we defined gesture to be a sequence of states in a configuration space. States were defined on some input space (say the joint angles returned by a DataGlove) and were ....

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. of the Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, Austin, Texas, Nov. 1994.

....articulated models comprised of 2D or 3D solid primitive, sometimes accounting for self occlusion by having an explicit object model. Shape models were used by [2] for human tracking. The most relevant motion recognition approaches related to our work are those that employ view based models [3, 9, 20]. In particular, 9] uses motion history images (MHI) and motion energy images (MEI) temporal templates that are matched using a nearest neighbor approach against examples of given motions already learned. The main problem of this method is the requirement of having stationary objects, and the ....

....learned. The main problem of this method is the requirement of having stationary objects, and the insufficiency of the representation to discriminate among similar motions. Motion analysis techniques have had the problem that registration of useful, filtered information is a hard labor by itself [9, 4, 18, 20, 32]. Our system provides a functional front end that supports such tasks. 3 Basic Approach A diagram of our approach is illustrated in Fig. 1. The first stage of the algorithm is based on the background subtraction methods of [31] The system is initialized by acquiring statistical measurements of ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

....high resolution support from the data is available. As we will discuss, we have found it is often difficult to separate the tracking from the boundary estimation problem. Therefore, some people tracking methods that require accurate object boundary detection based on intensity[2] and optical flow[12] are difficult to apply in our tracking task. Further the difficulty in segmenting colliding children and their rapid changes in appearance and motion prevents the use of differential motion estimators that use smooth or planar motion models [4] and tracking techniques that require reasonably ....

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. Work. Non-Rigid Motion, pages 77--82, Nov. 1994.

....to the same spatial reference, while tracking using a single camera does not have this requirement. 3.1 Single Camera Tracking Most methods for tracking moving humans use image sequences taken from a single camera. Features used for tracking are usually points and motion blobs. Polana and Nelson [37] observed that the movements of arms and legs converge to that of the torso. In their work [37] each walking subject image was bounded by a rectangular box, and the centroid of the bounding box was used as the feature to track. Positions of the center point in the previous frames were used to ....

....3.1 Single Camera Tracking Most methods for tracking moving humans use image sequences taken from a single camera. Features used for tracking are usually points and motion blobs. Polana and Nelson [37] observed that the movements of arms and legs converge to that of the torso. In their work [37], each walking subject image was bounded by a rectangular box, and the centroid of the bounding box was used as the feature to track. Positions of the center point in the previous frames were used to estimate the current position. Therefore, correct tracking was resolved even when the two ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion (or how to get your man without finding his body parts). In Proc. of IEEE Computer Society Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, Austin, TX, 1994.

....motion of the parts of an object. Unfortunately, only synthetic images were used in their experiment. In practice, it is difficult to track points through an image sequence, because of occlusion and noise. Spatio temporal curves can also be used to characterize periodic motions. Nelson and Polana [35] track objects using the centroid of the moving pixels in the frame. The size and the position of the object is normalized. The bounding box of the object is divided into square cells and the motion magnitude in each cell is summed. The idea is to use the motion magnitudes as features. If the ....

Ramprasad Polana and Randal Nelson. Low level recognition of human motion. In IEEE Workshop on Motion of Non-rigid and Articulated Objects, pages 77--82, Austin, TX, 1994.

....N spatial grid is superimposed on the motion region, after a possible normalization of its extent. In each of the K Theta N tiles a simple feature is computed, and these are combined to form a K Theta N feature vector to describe the state of movement at time t. This is the approach taken by [19] [23] and [4] Another possibility is to use 2 D model based features, where the assumption is that as result of 2D segmentation and tracking a sequence of 2 D stick figure poses is available. For example, Goddard [7] uses the 2 D angular velocities and orientations of the links as features. Guo ....

....model free features have been able to claim early successes in matching human movement patterns. Indeed, for constrained types of human movement (such as walking parallel to the image plane, involving periodic motion) many of these features have been successfully used for classification, as in [19]. But we find it unlikely that reliable recognition of more unconstrained and complex human movement (e.g. humans wandering around, performing different gestures while walking and turning) can be achieved using these features exclusively. With respect to using 2 D model based features, we note ....

R. Polana and R. Nelson, "Low Level Recognition of Human Motion," IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, 1994.

....are not smooth or rigid, and where multiple objects are interacting. Bregler uses many levels of representation based on mixture models, EM, and recursive Kalman and Markov estimation to learn and recognize human dynamics [4] Deformable trackers that track small images of people are described in [6]. 3 Background Scene Modeling and Foreground Region Detection Frame differencing in W 4 is based on a model of background variation obtained while the scene contains no people. The background scene is modeled by representing each pixel by three values; its minimum and Figure 1: Motion ....

R. Polona, R. Nelson "Low Level Recognition of Human Motion", In Proc. Non Rigid Motion Workshop, November 1994.

....and segmentation can assist in many applications requiring object and activity recognition and representation. The main body of work on periodic motion is modelbased (e.g. 1] 2] More recently there is work on motion recognition directly using low level features of motion information (e.g. [3][4] 5] However, to date, there has not been a method which uses low level motion features to detect and segment periodic motion simultaneously. In this work, we attempt to tackle this problem by using a Fourier spectral based approach. The periodicity templates generated in the process ....

....which is a weakness of the method. The periodicity measure for an entire sequence is the maximum of pf averaged among pixels whose highest power spectrum values appear on the same frequency. The final periodicity measure is used to distinguish periodic and non periodic motion by thresholding. In [3], flow based algorithms are used to transform the Frame 20 Frame 40 Frame 60 Frame 80 Figure 1: Frames 20, 40, 60, and 80 of the 97 frame sequence Walker. Frame size is 320 by 240. image sequence so that the object in consideration stays at the center of the image frame. Then flow magnitudes in ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Motion of Nonrigid and Articulated Objects, Austin, TX, 1994.

....based on trajectory patterns, but does not consider the non rigid motion or shape changes of the moving objects. Such approaches cannot distinguish between totally different actions if they have similar trajectories (e.g. biking vs. walking) On the other hand, appearance change modeling methods [6, 13, 14, 38] look at the image motion of the object itself instead of static configurations or centroids. In particular, Davis and Bobick [14] used motion history images (MHI) and motion energy images (MEI) MHIs and MEIs are temporal image templates that are matched using a nearest neighbor approach against ....

....Reasoning Motion Recognition Figure 1: System Diagram. In general, motion analysis techniques have had the problem that registration of useful, filtered information is a hard labor by itself. Methods for addressing this issue have been proposed previously in the seminal work of [3] and also [8, 14, 17, 35, 38, 53]. Our system provides this information by estimating the 3D trajectories of multiple moving objects, and modeling occlusion. This allows for normalization of incoming appearance with respect to gross changes in scale and 3D orientation and heading. Furthermore, the availability of object 3D ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

....at maximum, the power of higher dimensional descriptions. It is then essential to count on experimental results that support their efficiency and demonstrate their power. The importance of motion recognition problems is evidenced by the increasing attention they have received in recent years [24, 25, 1, 16, 20, 19]. One of the main areas of research is the analysis of humans in motion. There are numerous domains that motivate the research in this area: video surveillance, human computer interaction, athletics, dance, robot motion, among others. Motion of objects is much more semantically rich than static ....

....knowledge model, for example [5, 11, 21, 22] The most relevant group of approaches related to ours, as in [6] are those that try to achieve recognition of the motion directly from the sequence of images. No explicit use of the static images is done besides what is needed to represent the motion [20, 2, 6]. We give a brief overview of their fundamentals. According to [20] recognition of repetitive motion can be achieved on the basis of bottom up processing, without identifying specific parts or classification of the object. Their approach needs to estimate local velocity to undo the translation ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

....able to image the tracked subject in a broad area over a long period of time. In pursuit of this goal, our work has evolved from studying human walking using a fixed camera [1, 2] to tracking non background objects in a single moving camera [3] The studies in tracking using a fixed single camera [4, 2, 5] are limited to a very narrow area due to the restricted viewing angle of the system. A moving camera with a substantial degree of rotational freedom [3] increases the viewing angle to certain degree, however, it complicates the implementation by adding the motion estimation of both the viewing ....

....flow methods [6] which are widely used for featureless motion tracking, demand small and smooth motion between frames,a restriction that also does not hold in our case. In this paper, we propose to track a moving human in different camera views based on low level recognition of human motion [5]. A simpler form of a 2D human model [7] is applied to detect moving human subjects. Tracking between consecutive frames is mainly based on the consistency of the position, velocity, and average intensity of feature points formulated by multivariate Gaussian models, considered in the views of ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion (or how to get your man without finding his body parts). In Proc. of IEEE Computer Society Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, 1994.

....multimedia storage and retrieval. In this paper, we describe a framework to detect, segment, and track a sequence of monocular images of humans moving in an indoor environment. The common set of assumptions in tracking human motion includes small image motion[1, 2, 3, 4] a fixed viewing system [5, 6, 1, 2, 3, 4, 7], constant velocity[7] uniform intensity and rigidity of the moving object[5, 6] In our study, these assumptions are systematically relaxed. Whether the viewing system is moving or not, the difference between the image motion of the background and that of the moving subjects is a strong cue that ....

....paper, we describe a framework to detect, segment, and track a sequence of monocular images of humans moving in an indoor environment. The common set of assumptions in tracking human motion includes small image motion[1, 2, 3, 4] a fixed viewing system [5, 6, 1, 2, 3, 4, 7] constant velocity[7], uniform intensity and rigidity of the moving object[5, 6] In our study, these assumptions are systematically relaxed. Whether the viewing system is moving or not, the difference between the image motion of the background and that of the moving subjects is a strong cue that we exploit to ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion (or how to get your man without finding his body parts). In Proc. of IEEE Computer Society Workshop on Motion of Non-Rigid and Articulated Objects, pages 77--82, 1994.

....surface patterns may come to mind first, periodicity often involves both space and time, such as cyclic motion. The main body of work on periodic motion is model based (e.g. 1] 2] More recently there is work on motion recognition directly using low level features of motion information (e.g. [3][4] 5] However, to date, there has not been a method which uses low level features to detect and systematically characterize periodicity in space and time. In this work, we attempt to tackle this problem by using periodicity templates to incorporate the location, strength, and other ....

....which is a weakness of the method. The periodicity measure for an entire sequence is the maximum of pf averaged among pixels whose highest power spectrum values appear on the same frequency. The final periodicity measure is used to distinguish periodic and non periodic motion by thresholding. In [3], flow based algorithms are used to transform the image sequence so that the object in consideration is stabilized at the center of the image frame. Then flow magnitudes in tessellated frame areas of periodic motion were used as feature vectors for motion classification. In this paper, we show ....

[Article contains additional citation context not shown here]

R. Polana and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Motion of Nonrigid and Articulated Objects, pages 77--82, Austin, TX, Nov. 11-12 1994.

....sitting. Such capabilities argue for recognizing action from the motion itself, as opposed to first reconstructing a 3dimensional model of a person, and then recognizing the action of the model. The prior work in this area has addressed either periodic or gross motion detection and recognition [17, 21, 24] or the understanding of facial expressions [23, 1, 10] In [2, 5] we propose a representation and recognition theory that decomposes motion based recognition into first describing where there is motion (the spatial pattern) and then describing how the motion is moving. The basic idea is that we ....

Polana, R. and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Non-rigid and Articulated Motion, 1994.

....is the requirement that there be individual features or properties that can be extracted from each frame of the image sequence. These approaches accomplish motion understanding by recognizing a sequence of static configurations. Alternatively, there is the work on direct motion recognition [16, 19, 20, 2, 10, 15, 4]. These approaches attempt to characterize the motion itself without any reference to the underlying static images or a sequence of poses. Of these techniques, the work of Polana and Nelson [16] is the most relevant to the results presented here. The goal of their research is to represent and ....

....configurations. Alternatively, there is the work on direct motion recognition [16, 19, 20, 2, 10, 15, 4] These approaches attempt to characterize the motion itself without any reference to the underlying static images or a sequence of poses. Of these techniques, the work of Polana and Nelson [16] is the most relevant to the results presented here. The goal of their research is to represent and recognize actions as dynamic systems where it is the spatially distributed properties motion (in their case periodicity) that is matched. 3 Temporal templates Our goal is to construct a ....

[Article contains additional citation context not shown here]

Polana, R. and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Nonrigid and Articulated Motion, 1994.

....with also account for parallax [17] Recognition of the extracted motion is an important new area of research which merits increased effort. Sometimes low level features can be very successful for recognition of motion textures (water, leaves, etc. and even human activities (see Polana and Nelson [18]) Figure 2: Stroboscopic image made from several frames of a noisy football video. See http: wwwwhite. media.mit.edu steve orbits orbits.html for a closer look. 2.8 Browsing video It is unclear at this stage what will be the best way to browse video. Tonomura et al. have outlined several ....

R. Polana and R. Nelson, "Low level recognition of human motion," in IEEE Workshop on Motion of Nonrigid and Articulated Objects, (Austin, TX), 1994.

....3 D One may question whether it is desirable or feasible to try to recover 3 D body pose from 2 D image sequences for the purpose of recognizing human movement. An alternative approach is to work directly with 2 D features derived from the images, using some form of 2 D model [7] 11] or not [3] [16]. Recognition systems using 2 D model free features have been able to claim early successes in matching human movement patterns. For constrained types of human movement (such as walking parallel to the image plane, involving periodic motion) many of these features have been successfully used for ....

....using 2 D model free features have been able to claim early successes in matching human movement patterns. For constrained types of human movement (such as walking parallel to the image plane, involving periodic motion) many of these features have been successfully used for classification, as in [16]. This may indeed be the easiest and best solution for several applications. But we find it unlikely that reliable recognition of more unconstrained and complex human movement (e.g. humans wandering around, making different gestures while walking and 1 The tracking results described in this ....

R. Polana and R. Nelson, "Low Level Recognition of Human Motion," IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, TX, 1994.

....One advantage of having the recovered model is the ability to estimate and predict the feature locations, for instance edges, in the following frames. Given the past history of the model configurations, prediction is commonly attained using Kalman filtering [29, 28, 16] and velocity constraints [26, 15]. Because of the self occlusions that frequently occur in articulated objects, some employ multiple cameras and restrict the motion to small regions [28, 15] to help with projective model occlusion constraints. A single camera is used in [19, 16, 29] but the actions tracked in these works had ....

....features or properties that can be extracted and tracked from each frame of the image sequence. Hence, motion understanding is really accomplished by recognizing a sequence of static configurations. This understanding generally requires previous recognition and segmentation of the person [26]. We now consider recognition of action within a motion based framework. 2.2 Motion based recognition Directional motion recognition [26, 30, 24, 3, 34, 31, 11] approaches attempt to characterize the motion itself without reference to the underlying static poses of the body. Two main approaches ....

[Article contains additional citation context not shown here]

Polana, R. and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Non-rigid and Articulated Motion, 1994.

No context found.

Polana, R. and Nelson, R. 1994. Low level recognition of human motion. In Proc. of the IEEE Workshop on Motion of Non-Rigid and Articulated Objects, Austin, pp. 77--82.

No context found.

R. Polana and R. Nelson, Low level recognition of human motion (or how to get your man without finding his body parts), in Motion of Non-Rigid and Articulated Objects, Proceedings of IEEE Workshop on, 1994, pp. 77--82.

No context found.

R. Polana and R. Nelson. Low level recognition of human motion. In Proc. of Workshop on Non-rigid Motion, pages 77--82, 1994.

No context found.

R. Polana and R. Nelson, "Low level recognition of human motion," in Proc. IEEE Workshop Nonrigid and Articulated Motion, Austin, TX, 1994, pp. 77--82.

No context found.

R. Polana and R. Nelson. Low level recognition of human motion. Proc. IEEE Workshop on Nonrigid and Articulate Motion, 1994.

No context found.

Polana R, Nelson R. (1994). Low Level Recognition of human motion(or how to get your man without finding his body parts). Proceedings IEEE Computer Society Workshop on Motion of Non-Rigid and Articulated Objects, Austin Texas

No context found.

R. Polana and R. Nelson, "Low level recognition of human motion," in Proc. of IEEE Workshop on Motion of Non-rigid and Articulated Object, 1994, pp. 77--82.

No context found.

R. Polana and R. Nelson. Low level recognition of human motion. In IEEE Workshop on Motion of Non-Rigid and Articulated Objects, pages 77-- 82, Austin, 1994.

No context found.

R. Polona, R. Nelson #Low Level Recognition of Human Motion", In Proc. Non Rigid Motion Workshop,November 1994.

No context found.

Polana, R. & Nelson, R. 1994 "Low level recognition of human motion," IEEE Workshop on Non-rigid and Articulated Motion, Austin, Texas.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC