Humanistic Intelligence (HI) is proposed as a new signal processing framework in which the processing apparatus is inextricably intertwined with the natural capabilities of our human body and mind. Rather than trying to emulate human intelligence, HI recognizes that the human brain is perhaps the best neural network of its kind, and that there are many new signal processing applications, within the domain of personal technologies, that can make use of this excellent but often overlooked processor. The emphasis of this paper is on personal imaging applications of HI, as we take a first step toward an intelligent wearable camera system that can allow us to effortlessly capture our day-to-day experiences, help us remember and see better, provide us with personal safety through crime reduction, and facilitate new forms of communication through collective connected humanistic intelligence. The wearable signal processing hardware, which began as a cumbersome backpack-based photographic apparatus of the 1970s, and evolved into a clothing-based apparatus in the early 1980s, currently provides the computational power of a UNIX workstation concealed within ordinary-looking eyeglasses and clothing. Thus it may be worn continuously during all facets of ordinary day-to-day living, so that, through long-term adaptation, it begins to function as a true extension of the mind and body.to signal processing hardware that is both in close physical proximity to the user, and is constant. By constant, what is meant is that the apparatus is both interactionally and operationally constant.
This constancy of user-interface separates this signal processing architecture from other related devices such as pocket calculators and Personal Data Assistants (PDAs).
By operationally constant, what is meant is that although it may have ``sleep" modes, it is never ``dead" (as is typically a calculator which may be worn in a shirt pocket). By interactionally constant, what is meant is that the inputs and outputs of the device are always potentially active. Thus, for example, a pocket calculator, worn in a shirt pocket, and left on all the time is still not interactionally constant, because it cannot be used in this state (e.g. one still has to pull it out of the pocket to see the display or enter numbers). A wrist watch is a borderline case. Although it continues to keep proper time, and it is worn on the body, one must make a conscious effort to orient it within one's field of vision.
Devices embodying humanistic intelligence are not merely intelligent signal processors that a user might wear or carry in close proximity to the body, but instead, are devices that turn the user into part of an intelligent control system where the user becomes an integral part of the feedback loop.Human-Computer Interaction, the goal is not just to construct a device that can model (and learn from) the user, but, more importantly, to construct a device in which the user also must learn from the device. Therefore, in order to facilitate the latter, devices embodying HI should provide a constant user-interface -- one that is not so sophisticated and intelligent that it confuses the user. Although the device may implement very sophisticated signal processing algorithms, the cause and effect relationship of this processing to its input (typically from the environment or the user's actions) should be clearly and continuously visible to the user, even when the user is not directly and intentionally interacting with the apparatus. Accordingly, the most successful examples of HI afford the user a very tight feedback loop of system observability (ability to perceive how the signal processing hardware is responding to the environment), even when the controllability of the device is not engaged (e.g. at times when the device is not being used). A simple example is the viewfinder of a wearable camera system, which provides framing, a photographic point of view, and facilitates to the user a general awareness of the visual effects of the camera's own image processing algorithms, even when pictures are not being taken. Thus the human operator is always in the feedback loop of the imaging process, even though pictures may only be taken occasionally. A more sophisticated example is the biofeedback-controlled wearable camera system, in which the biofeedback process happens continuously, whether or not a picture is actually being taken. In this sense, the user becomes one with the machine, over a long period of time, even if the machine is only ``used'' (e.g. to actually take a picture) occasionally.
HI attempts to both build upon, as well as re-contextualize, concepts in intelligent signal processing, and related concepts such as neural networks, fuzzy logic, and artificial intelligence. HI also suggests a new goal for signal processing hardware, that is, to directly assist, rather than replace or emulate human intelligence. What is needed to facilitate this vision is a simple computational signal processing framework that empowers the human intellect.
Due to advances in low power microelectronics, we are entering a pivotal era in which it will become possible for us to be inextricably intertwined with computational technology that will become part of our everyday lives in a much more immediate and intimate way than in the past.
Physical proximity and constancy were simultaneously realized by the `WearComp' project of the 1970s and early 1980s (Figure 1)
However, what makes WearComp particularly useful in new and interesting ways, and what makes it particularly suitable as a basis for humanistic intelligence, is the collection of other input devices, not all of which are found on a desktop multimedia computer.
In typical embodiments of `WearComp' these measurement (input) devices include the following:
Certain applications use only a subset of these devices, but including all of them in the design facilitates rapid prototyping and experimentation with new applications. Most embodiments of WearComp are modular, so that devices can be removed when they are not being used.
A side-effect of this `WearComp' apparatus is that it replaces much of the personal electronics that we carry in our day-to-day living. It enables us to interact with others through its wireless data communications link, and therefore replaces the pager and cellular telephone. It allows us to perform basic computations, and thus replaces the pocket calculator, laptop computer and personal data assistant (PDA). It can record data from its many inputs, and therefore it replaces and subsumes the portable dictating machine, camcorder, and the photographic camera. And it can reproduce (``play back'') audiovisual data, so that it subsumes the portable audio cassette player. It keeps time, as any computer does, and this may be displayed when desired, rendering a wristwatch obsolete. (A calendar program which produces audible, vibrotactile, or other output also renders the alarm clock obsolete.)
However, it goes beyond replacing all of these items, because not only is it currently far smaller and far less obtrusive than the sum of what it replaces, but these functions are interwoven seamlessly, so that they work together in a mutually assistive fashion. Furthermore, entirely new functionalities, and new forms of interaction arise, such as enhanced sensory capabilities, as will be discussed in Sections 3 and 4.
Some more recent related work by others , also involves building circuits into clothing, in which a garment is constructed as a monitoring device to determine the location of a bullet entry. The underwearable differs from this monitoring apparatus in the sense that the underwearable is totally reconfigurable in the field, and also in the sense that it embodies humanistic intelligence (the apparatus reported in  performs a monitoring function but does not facilitate human interaction).
In summary, there were three reasons for the signal processing hardware being `underwearable':
Smart clothing was inspired by the need for comfortable signal processing devices that could be worn for extended periods of time. The inspiration for smart clothing arose out of noticing that some of the early headsets typically used with ``crystal radios'' were far more comfortable than the newer headsets, and could often be worn for many hours (some such early headsets had no head bands but instead was sewn into a cloth cap meant to be worn underneath a helmet). Of particular interest was the cords used in some of the early headsets (Fig 5(a)), early telephones, early patch cords,
The notion that cloth be rendered conductive, through the addition of metallic fibers interwoven into it, is one thing that makes possible clothing that serves as an RF shield (Fig 5(b)), manufactured to address response to the growing fear of the health effects of long-term exposure to radio-frequency exposure . However, it may also be used to shield signal processing circuits from outside interference, or as a ground plane for various forms of conformal antennas sewn into the clothing.
Smart clothing is made using either of the following two approaches:
Conductive materials have been used in certain kinds of drapery for many years for appearance and stiffness, rather than electrical functionality, but these materials can be used to make signal processing circuits, as depicted in Figure 6. Simple circuits like this suggest a future possible direction for research in this area.
Picard also suggests its possible use to estimate human emotion .
The emphasis of this paper will be on visual image processing with the WearComp apparatus. The author's dream of the 1970s, that of an intelligent wearable image processing apparatus, is just beginning to come to fruition.22].
One ultimate goal of the proposed apparatus and methodology is to ``learn" what is visually important to the wearer, and function as a fully automatic camera that takes pictures without the need for conscious thought or effort from the wearer. In this way, it might summarize a day's activities, and then automatically generate a gallery exhibition by transmitting desired images to the World Wide Web, or to specific friends and relatives who might be interested in the highlights of one's travel. The proposed apparatus, a miniature eyeglass-based imaging system, does not encumber the wearer with equipment to carry, or with the need to remember to use it, yet because it is recording all the time into a circular buffer, merely overwriting that which is unimportant, it is always ready. Thus, when the signal processing hardware detects something that might be of interest, recording can begin in a retroactive sense (e.g. a command may be issued to start recording from thirty seconds ago), and the decision can later be confirmed with human input. Of course this apparatus raises some important privacy questions which are beyond the scope of this article, but have been addressed elsewhere in the literature .
The system might use the inputs from the biosensors on the body, as a multidimensional feature vector with which to classify content as important or unimportant. For example, it might automatically record a baby's first steps, as the parent's eyeglasses and clothing-based intelligent signal processor make an inference based on the thrill of the experience. It is often moments like these that we fail to capture on film: by the time we find the camera and load it with film, the moment has passed us by.
In this kind of situation, it is desirable that the apparatus would function autonomously, without conscious effort from the wearer, even though the wearer might be aware of the signal processing activities of the measuring (sensory) apparatus he or she is wearing.
As a simplified example of how the processing might be done, we know that the wearer's heart rate, averaged over a sufficient time window, would likely increase dramatically with no corresponding increase in footstep rate (in fact footsteps would probably slow at the request of the gunman). The computer would then make an inference from the data, and predict a high visual saliency. (If we simply take heart rate divided by footstep rate, we can get a first-order approximation of the visual saliency index.) A high visual saliency would trigger recording from the wearer's camera at maximal frame rate, and also send these images together with appropriate messages to friends and relatives who would look at the images to determine whether it was a false alarm or real danger.
Such a system is, in effect, using the wearer's brain as part of its processing pipeline, because it is the wearer who sees the shotgun, and not the WearComp apparatus (e.g. a much harder problem would have been to build an intelligent machine vision system to process the video from the camera and determine that a crime was being committed). Thus humanistic intelligence (intelligent signal processing arising, in part, because of the very existence of the human user) has solved a problem that would not be possible using machine-only intelligence.
Furthermore, this example introduces the concept of `collective connected humanistic intelligence', because the signal processing systems also rely on those friends and relatives to look at the imagery that is wirelessly sent from the eyeglass-mounted video camera and make a decision as to whether it is a false alarm or real attack. Thus the concept of HI has become blurred across geographical boundaries, and between more than one human and more than one computer.
An important aspect of HI is that the conscious will of the user may be inserted into or removed from the feedback loop of the entire process at any time. A very simple example, taken from everyday experience, rather than another new invention, is now presented.
One of the simplest examples of HI is that which happens with some of the early autofocus Single Lens Reflex (SLR) cameras in which autofocus was a retrofit feature. The autofocus motor would typically turn the lens barrel, but the operator could also grab onto the lens barrel while the autofocus mechanism was making it turn. Typically the operator could ``fight'' with the motor, and easily overpower it, since the motor was of sufficiently low torque. This kind of interaction is particularly useful, for example, when shooting through a glass window at a distant object, where there are two or three local minima of the autofocus error function (e.g. focus on particles of dust on the glass itself, focus on a reflection in the glass, and focus on the distant object). Thus when the operator wishes to focus on the distant object and the camera system is caught in one of the other local minima (for example, focused on the glass), the user merely grasps the lens barrel, swings it around to the approximate desired location (as though focusing crudely by hand, on the desired object of interest), and lets go, so that the camera will then take over and bring the desired object into sharp focus.
This very simple example illustrates a sort of humanistic intelligent signal processing in which the intelligent autofocus electronics of the camera work in close synergy with the intellectual capabilities of the camera operator.
It is this aspect of HI, that allows the human to step into and out of the loop at any time, that makes it a powerful paradigm for intelligent signal processing.
Since the emphasis of this paper is on personal imaging, the treatment here will focus on the wearable camera (discussed here in Section 3) and the wearable radar (discussed in Section 4). The other measurement instruments are important, but their role is primarily to facilitate exploiting the human intellect for purposes of processing data from the imaging apparatus.
The theoretical framework for processing video is based on regarding the camera as an array of light measuring instruments capable of measuring how the scene or objects in view of the camera respond to light . This framework has two important special cases, the first of which is based on photometric self-calibration to build a lightspace map from images which differ only in overall exposure, and the second of which is based on algebraic projective geometry as a means of combining information from images related to one-another by a projective coordinate transformation.
These two special cases of the theory are now presented in Sections 3.2.1 and 3.2.2 respectively, followed by bringing both together in Section 3.2.3. The theory is applicable to standard photographic or video cameras, as well as to the wearable camera and personal imaging system.Section 3.2.2, is found to be useful in the context of the personal imaging system, as will be described in Section 3.2.3.
Most everyday scenes have a far greater dynamic range than can be recorded on a photographic film or electronic imaging apparatus (whether it be a digital still camera, consumer video camera, or eyeglass-based personal imaging apparatus as described in this paper). However, a set of pictures, that are identical except for their exposure, collectively show us much more dynamic range than any single picture from that set, and also allow the camera's response function to be estimated, to within a single constant scalar unknown.
A set of functions,
where are scalar constants, is known as a Wyckoff set, and describes a set of images, , when is the spatial coordinate of a piece of film or the continuous spatial coordinates of the focal plane of an electronic imaging array, q is the quantity of light falling on the sensor array, and f is the unknown nonlinearity of the camera's response function (assumed to be invariant to .
Because of the effects of noise (quantization noise, sensor noise, etc.), in practical imaging situations, the dark (``underexposed'') pictures show us highlight details of the scene that would have been overcome by noise (e.g. washed out) had the picture been ``properly exposed''. Similarly, the light pictures show us some shadow detail that would not have appear above the noise threshold had the picture been ``properly exposed''.
A means of simultaneously estimating f and , given a Wyckoff set , has been proposed. A brief outline of this method follows. For simplicity of illustration (without loss of generality), suppose that the Wyckoff set contains two pictures, and , differing only in exposure (e.g. where the second image received k times as much light as the first). Photographic film is traditionally characterized by the so-called ``D logE'' (Density versus log Exposure) characteristic curve. Similarly, in the case of electronic imaging, we may also use logarithmic exposure units, , so that one image will be K = log(k) units darker than the other:
(f^-1(I_1)) = Q = (f^-1(I_2)) - K
The existence of an inverse for f follows from the semimonotonicity assumption . (We expect any reasonable camera to provide a semimonotonic relation between quantity of light received, q, and the pixel value reported.) Since the logarithm function is also monotonic, the problem comes down to estimating the semimonotonic function and the scalar constant K, given two pictures and :
The unknowns (F and K) may be solved by regression (e.g. in a typical imaging situation with 480 by 640 by 256 grey values, this amounts to solving 307200 equations in 257 unknowns: 256 for F and one for K). An intuitive way to solve this problem, which also provides valuable insight into how to combine the differently exposed images into a single image of extended dynamic range, is as follows: recognize that
provides a recipe for ``registering'' (appropriately lightening or darkening) the second image with the first. This registration procedure differs from image registration procedure commonly used in image resolution enhancement (to be described in Section 3.2.2) because it operates on the range (tonal range) of the image as opposed to its domain (spatial coordinates) . (In Section 3.2.3, registration in both domain and range will be addressed).
Now if we construct a cross histogram of the two images, we will have a matrix (typically of dimension 256 by 256 assuming 8-bit-deep images) that completely captures all of the information about the relationship between the two pictures. This representation discards all spatial information in the images (which is not relevant to estimating f). Thus the regression problem (that of solving (3) can be done on the cross histogram instead of the original pair of images. This approach has the added advantage of breaking the problem down into two separate simpler steps:
The above method allows us to estimate, to within a constant scale factor, the photometric response function of the camera without making any assumptions on the form of f, other than semimonotonicity. However, if we use a parametric model (e.g. to fit a smooth parameterized curve through the cross histogram), then the results can be somewhat more noise-immune.
A suitable parameterization is motivated by the fact that
the ``D log E'' curve of most typical photographic emulsions is linear
over a relatively wide region, which suggests the commonly used empirical
law for the response function of film :
This formulation has been found to apply well to the eyeglass-based
camera system designed and built by the author.
The constant that
characterizes the density of unexposed film,
applies equally well to the electronic imaging array
in the eyeglass-based camera. The quantity
may be subtracted off, either
through design and adjustment of the circuits connected to the
sensor array, or through the capture of
one picture (or several pictures signal-averaged)
taken with the lens covered,
to be subtracted from each of the incoming pictures, or
it may be estimated (e.g. treated
as an additional unknown parameter).
The range-range plot
then takes the form
where k is the ratio of exposures relating the two pictures. Thus to find the value of the linear constant, , in we simply apply linear regression to points in the joint histogram. From we can obviously find the camera's contrast parameter, .
Once f is determined, each picture becomes a different estimate of the same
q_n = 1k_nf^-1(I_n)
true quantity of light falling on each pixel of the image sensor. Thus one may regard each of these measurements (pixels) as a light meter (sensor element) that has some nonlinearity followed by a quantization to a measurement having typically 8-bit precision.
It should be emphasized that most image processing algorithms incorrectly assume that the camera response function is linear (e.g. almost all current image processing, such as blurring, sharpening, unsharp masking, etc., operates linearly on the image) while in fact it is seldom linear. Even Stockham's homomorphic filtering, which advocates taking the log, applying linear filtering, and then taking the antilog, fails to capture the correct nonlinearity, as it ignores the true nonlinearity of the sensor array. It has recently been shown that, in the absence of any knowledge of the camera's nonlinearity, simply selecting a value of two or three, and using (5) to linearize the image (e.g. square or cube all pixel values in the image), followed by filtering, followed by the inverse operation (e.g. extracting the square root or cube root of each pixel in the image) provides much better results than the approach advocated by Stockham. Of course, finding the true response function of the camera allows us to do even better, as we may then apply our linear signal processing methodology to the original light falling on the image sensor.
obtaining a first-order estimate of the yaw, pitch, and roll of the wearer's head, or making an important first step in the more difficult problem of estimating depth and structure from a scene . Thus, in this section, the assumption is that most of the image motion arises from that of generating an environment map, zero-parallax is assumed.
The problem of assembling multiple pictures of the same scene into a single image commonly arises in mapmaking (with the use of aerial photography) and photogrammetry, where zero-parallax is also generally assumed. Many of these methods require human interaction (e.g. selection of features), and it is desired to have a fully automated system that can assemble images from the eyeglass-based camera. Fully automatic featureless methods of combining multiple pictures have been previously proposed (see also here), but with an emphasis on subpixel image shifts; the underlying assumptions and models (affine, and pure translation, respectively) were not capable of accurately describing more macroscopic image motion. A characteristic of video captured from a head-mounted camera is that it tends to have a great deal more macroscopic image motion, and a great deal more perspective `cross-chirping' between adjacent frames of video, while the assumptions of static scene content and minimal parallax are still somewhat valid. This assumption arises for the following reasons:
Accordingly, two featureless methods of estimating the parameters of a projective group of coordinate transformations were first proposed in , and in more detail in , one direct and one based on optimization (minimization of an objective function). Although both of these methods are multiscale (e.g. use a coarse to fine pyramid scheme), and both repeat the parameter estimation at each level (to compute the residual errors), and thus one might be tempted to call both iterative, it is preferable to refer to the direct method as repetitive to emphasize that does not require a nonlinear optimization procedure such as Levenberg-Marquardt, or the like. Instead, it uses repetition with the correct law of composition on the projective group, going from one pyramid level to the next by application of the group's law of composition. A method similar to the optimization-method was later proposed. The direct method has also been subsequently described in more detail.
The direct featureless method for estimating the 8 scalar parameters , of an exact projective (homographic) coordinate transformation is now described. In the context of personal imaging, this result is used to multiple images to seamlessly combine images of the same scene or object, resulting in a single image (or new image sequence) of greater resolution or spatial extent.
Many papers have been published on the problems of motion estimation and frame alignment (review and comparison). In this Section the emphasis is on the importance of using the ``exact'' 8-parameter projective coordinate transformation, particularly in the context of the head-worn miniature camera.
The most common assumption (especially in motion estimation for coding, and optical flow for computer vision) is that the coordinate transformation between frames is translation. Tekalp, Ozkan, and Sezan have applied this assumption to high-resolution image reconstruction. Although translation is less simpler to implement than other coordinate transformations, it is poor at handling large changes due to camera zoom, rotation, pan and tilt.
Zheng and Chellappa considered the image registration problem using a subset of the affine model -- translation, rotation and scale. Other researchers (also these) have assumed affine motion (six parameters) between frames.
The only model that properly captures the ``keystoning'' and ``chirping'' effects of projective geometry is the projective coordinate transformation. However, because the parameters of the projective coordinate transformation had traditionally been thought to be mathematically and computationally too difficult to solve, most researchers have used the simpler affine model or other approximations to the projective model.
The 8-parameter pseudo-perspective model does, in fact, capture both the converging lines and the chirping of a projective coordinate transformation, but not the true essence of projective geometry.
Of course, the desired ``exact'' eight parameters come from the projective group of coordinate transformations, but they have been perceived as being notoriously difficult to estimate. The parameters for this model have been solved by Tsai and Huang, but their solution assumed that features had been identified in the two frames, along with their correspondences. The main contribution of the result summarized in this Section is a simple featureless means of automatically solving for these 8 parameters.
A group is a set upon which there is defined an associative law of composition (closure, associativity), which contains at least one element (identity) who's composition with another element leaves it unchanged, and for which every element of the set has an inverse.
In the context of this paper, coordinate transformations are the operators (group), and images are the operands (set). When the coordinate transformations form a group, then two such coordinate transformations, and , acting in succession, on an image (e.g. acting on the image by doing a coordinate transformation, followed by a further coordinate transformation corresponding to , acting on that result) can be replaced by a single coordinate transformation. That single coordinate transformation is given by the law of composition in the group.
The orbit of a particular element of the set under the group operation is the new set formed by applying to it all possible operators from the group.
Thus the orbit is a collection of pictures formed from one picture through applying all possible projective coordinate transformations to that picture. This set is referred to as the `video orbit' of the picture in question. Equivalently, we may imagine a static scene, in which the wearer of the personal imaging system is standing at a single fixed location. He or she generates a family of images in the same orbit of the projective group by looking around (rotation of the head).
The projective group of coordinate transformations,
is represented by matrices of the form:
Where, in practical engineering problems, in which d is never zero, the eight scalar parameters are denoted by , , , and .
The `video orbit' of a given 2-D frame is defined to be the set of all images that can be produced by applying operators from the 2-D projective group of coordinate transformations (8) to the given image. Hence, the problem may be restated: Given a set of images that lie in the same orbit of the group, find for each image pair, that operator in the group which takes one image to the other image.
If two frames of the video image sequence, say, and , are in the same orbit, then there is an group operation such that the mean-squared error (MSE) between and is zero. In practice, however, the element of the group that takes one image ``nearest'' the other is found (e.g. there will be a certain amount of error due to violations in the assumptions, due to noise such as parallax, interpolation error, edge effects, changes in lighting, depth of focus, etc).
The brightness constancy constraint equation which gives the flow velocity components, is:
As is well-known, the optical flow field in 2-D is under-constrained. The model of pure translation at every point has two parameters, but there is only one equation (10) to solve, thus it is common practice to compute the optical flow over some neighborhood, which must be at least two pixels, but is generally taken over a small block, , , or sometimes larger (e.g. the entire image, as in the Video Orbits algorithm described here).
However, rather than estimating the 2 parameter translational flow, the task here is to estimate the eight parameter projective flow (8) by minimizing:
_flow=(u_m^TE_x + E_t)^2
Although a sophisticated nonlinear optimization procedure, such as Levenberg-Marquardt, may be applied to solve (11), it has been found that solving a slightly different but much easier problem, allows us to estimate the parameters more directly and accurately for a given amount of computation:
(This amounts to weighting the sum differently.)
Differentiating (eq:nonlinearopt) with respect to the free parameters , and setting the result to zero gives a linear solution:
( ^T ) [a_11,a_12,b_1,a_21,a_22,b_2,c_1,c_2]^T
In practice, this process has been improved significantly by using an `Abelian pre-processing step' based on generalizations of Fourier transform cross spectra such as the cross-Fourier-Mellin transform and subspaces of the cross-chirplet transform , as well as non-commutative subgroup pre-processing based on the full cross-chirplet transform which is a generalization of the multi-resolution Fourier transform (see also).
) is also allowed to change of its own accord (e.g. arbitrarily).
Nyquist showed how a signal can be reconstructed from a sampling of finite resolution in the domain (e.g. space or time), but assumed infinite dynamic range (e.g. infinite precision or word length per sample). On the other hand, if we have infinite spatial resolution, but limited dynamic range (even if we have only 1 bit of image depth), Curtis and Oppenheim showed that we can also obtain perfect reconstruction using an appropriate modulation function. In the case of the personal imaging system, we typically begin with images that have very low spatial resolution and very poor dynamic range (video cameras tend to have poor dynamic range, and this poor performance is especially true of the small CCDs that the author uses in constructing unobtrusive lightweight systems). Thus, since we lack both spatial and tonal resolution, we are not at liberty to trade some of one for more of the other. Thus the problem of `spatiotonal' (simultaneous spatial and tonal) resolution enhancement is of particular interest in personal imaging.
In Section 3.2.1, a new method of allowing a camera to self-calibrate was proposed. This methodology allowed the tonal range to be significantly improved. In Section 3.2.2, a new method of resolution enhancement was described. This method allowed the spatial range to be significantly enhanced.
In this Section (3.2.3), a method of enhancing both the tonal range and the spatial domain resolution of images is proposed. It is particularly applicable to processing video from miniature covert eyeglass-mounted cameras, because it allows very noisy low quality video signals to provide not only high-quality images of great spatiotonal definition, but also to provide a rich and accurate photometric measurement space which may be of significant use to intelligent signal processing algorithms. That it provides not only high quality images, but also linearized measurements of the quantity of light arriving at the eyeglasses from each possible direction of gaze, follows from a generalization of the photometric measurement process outlined in Section 3.2.1.
Most notably, this generalization of the method no longer assumes that the camera need be mounted on a tripod, but only that the images fall in the same orbit of a larger group, called the `projectivity+gain' group of transformations.
Thus the apparatus can be easily used without conscious thought or effort, which gives rise to new intelligent signal processing capabilities. The method works as follows: As the wearer of the apparatus looks around, the portion of the field of view that controls the gain (usually the central region of the camera's field of view) will be pointed toward different objects in the scene. Suppose for example, that the wearer is looking at someone so that their face is centered in the frame of the camera, . Now suppose that the wearer tips his or her head upward so that the camera is pointed at a light bulb up on the ceiling, but that the person's face is still visible at the bottom of the frame, . Because the light bulb has moved into the center of the frame, the camera's AGC causes the entire image to darken significantly. Thus these two images, which both contain the face of the person the wearer is talking to, will be very differently exposed. When registered in the spatial sense (e.g. through the appropriate projective coordinate transformation), they will be identical, over the region of overlap, except for exposure, if we assume that the wearer swings his or her head around quickly enough to make any movement in the person he is talking to negligible. While this assumption is not always true, there are certain times that it is true (e.g. when the wearer swings his or her head quickly from left to right and objects in the scene are moving relatively slowly). Because the algorithm can tell when the assumptions are true (by virtue of the error), during the times it is true, it use the multiple estimates of , the quantity of light received, to construct a high definition environment map.
An example of an image sequence captured with a covert eyeglass-based version of the author's WearComp7, and transmitted wirelessly to the Internet, appears in Fig 9.
and proceeds as follows: as in Mann & Picard, consider one dimensional ``images'' for purposes of illustration, with the understanding that the actual operations are performed on 2-D images. The 1-D projective+gain group is defined in terms of the ``group '' of projective coordinate transformations, taken together with the one-parameter group of gain (image darkening/lightening) operations:
where characterizes the gain operation, and admits a group representation:
giving the law of composition: where the first law of composition on the right hand side is the usual one for the projective group (a subgroup of the projective+gain group), and the second one is that of the one-parameter gain (homometric lightening/darkening) subgroup.
Two successive frames of a video sequence are related through a group-action that is near the identity of the group, thus one may think of the Lie algebra of the group as providing the structure locally. As in previous work an approximate model which matches the `exact' model in the neighbourhood of the identity is used.
For the `gain group' (which is a one parameter group isomorphic to addition over the reals, or multiplication over the positive reals), the approximate model may be taken from Eq 5, by noting that:
g(f(q)) = f(kq) = + (kq)^
= - k^+ k^+ k^q^
Thus we see that g(f) is a ``linear equation'' (is affine) in f. This affine relationship suggests that linear regression on the cross histogram between two images would provide an estimate of and , while leaving unknown, which is consistent with the fact that the response curve may only be determined up to a constant scale factor.
we have that the (generalized) brightness change constraint equation is:
Combining this equation with the Taylor series
where , at time t, and is the frame difference of adjacent frames, we have:
Thus, the brightness change constraint equation becomes:
where, normalizing, .
Substitution of an approximate model (quadratic Taylor series)
into (19) gives:
as the non-weighted solution, where , , and .
yields a linear solution in parameters of the approximate model:
where F(x,t)=f(q(x)) at time t, , at time t, and is the frame difference of adjacent frames.
To construct a single floating-point image of increased spatial extent and increased dynamic range, first the images are spatiotonally registered (brought not just into register in the traditional `domain motion' sense, but also brought into the same tonal scale through homometric gain adjustment). This form of spatiotonal transformation is illustrated in Fig 11 where all the images are transformed into the coordinates of the first image of the sequence, and in Fig 12 where all the images are transformed into the coordinates of the last frame in the image sequence. It should be noted that the final homometric composite can be made in the coordinates of any of the images. The choice of reference frame is arbitrary since the result is a floating point image array (not quantized)! Furthermore, the final composite need not even be expressed in the spatiotonal coordinates of any of the incoming images. For example homometric coordinates (linear in the original light falling on the image array) may be used, to provide an array of measurements that linearly represent the quantity of light, to within a single unknown scalar constant for the entire array.
Although the response function, f(q), is fixed for a given camera, the `effective response function', depends on the exposure, , associated with frame, i, in the image sequence. By evaluating , we arrive at the so-called `certainty images' (Fig 9). Lighter areas of the `certainty images' indicate moderate values of exposure (mid-tones in the corresponding images), while darker values of the certainty images designate exposure extrema -- exposure in the toe or shoulder regions of the response curve where it is difficult to discern subtle differences in exposure.
The composite image may be explored interactively on a computer system (Fig 13).
It should be noted that the environment map was generated by a covert wearable apparatus, simply by looking around, and that no special tripod or the like was needed, nor was there significant conscious thought or effort required. In contrast to this proposed method of building environment maps, consider what must be done to build an environment map using QuickTime VR:
Despite more than twenty years photographic experience, Charbonneau needed to learn new approaches for this type of photography. First, a special tripod rig is required, as the camera must be completely level for all shots. A 35 mm camera ... with a lens wider than 28 mm is best, and the camera should be set vertically instead of horizontally on the tripod. ... Exposure is another key element. Blending together later will be difficult unless identical exposure is used for all views. [Campbell]The constraint of the QuickTime VR method and many other methods reported in the literature (e.g. [Sawhney][Kumar et al.], that all pictures be taken with identical exposure, is undesirable for the following reasons:
Once the final image composite, which reports, up to a single unknown scalar, the quantity of light arriving from each direction in space, it may also be reduced back to an ordinary (e.g. non-homometric) picture, by evaluating it with the function f. Furthermore, if desired, prior to evaluating it with f, a lateral inhibition similar to that of the human visual system, may be applied, to reduce its dynamic range, so that it may be presented on a medium of limited display resolution, such as a printed page (Fig 14).
I_c = f(L(f^-1(I)))
where is the output (or composite) image and L is the linear filtering operation. Images sharpened in this way tend to have a much richer, more pleasing and natural appearance, than those that are sharpened according to either a linear filter, or the variant of homomorphic filtering suggested by Stockham.
Perhaps the greatest value of homometric imaging, apart from its ability to capture high quality pictures that are visually appealing, is its ability to measure the quantity of light arriving from each direction in space. In this way, homometric imaging turns the camera into an array of accurate light meters.
Furthermore, the process of making these measurements is activity driven in the sense that areas of interest in the scene will attract the attention of the human operator, so that he or she will spend more time looking at those parts of the scene. In this way, those parts of the scene of greatest interest will be measured with the greatest assortment of ``rulers'' (e.g. with the richest collection of differently quantized measurements), and will therefore, without conscious thought or effort on the part of the wearer, be automatically emphasized in the composite representation. This natural foveation process arises, not because the Artificial Intelligence (AI) problem has been solved and built into the camera, so that it knows what is important, but simply because the camera is using the operator's brain as its guide to visual saliency. Because the camera does not take any conscious thought or effort to operate, it ``lives'' on the human host without presenting the host with any burden, yet it benefits greatly from this form of humanistic intelligence.
Section 3.2.3 may be further accentuated by building a camera system that is itself foveated.
Accordingly, the author designed and built a number of WearComp embodiments containing more than one electronic imaging array. One common variant, with a wide-angle camera in landscape orientation combined with a telephoto camera in portrait orientation was found to be particularly useful for humanistic intelligence: The wide camera provided the overall contextual information from the wearer's perspective, while the other (telephoto) provided close-up details, such as faces.
This `bi-foveated' scheme was found to work well within the context of the spatiotonal model described in the previous Section (3.2.3).
One realization of the apparatus comprised two cameras concealed in a pair of ordinary eyeglasses, is depicted in Figure 15.
Now suppose that a filter, L, is inserted into the `reality stream' by virtue of the appropriate signal processing on the incoming images prior to display on h:
I_m = h^-1(L(f^-1(I_c)))
In this context, L is called the `visual filter'mann260, and may be more than just a linear spatial filtering operation. As a trivial but illustrative example, consider L such that it operates spatially to flip the image left-right. This would make the apparatus behave like the left-right reversing glasses that Kohler and Dolezal made from prisms for their psychophysical experiments. (See Fig 17 (VR).)
One feature of this wearable tetherless computer-mediated reality system is that the wearer can choose to allow others to alter his or her visual perception of reality over an Internet connected wireless communications channel. An example of such a shared environment maps appears in Figure 18). This map not only allows others to vicariously experience our point of view (e.g. here a spouse can see that the wearer is at the bank, and send a reminder to check on the status of a loan, or pay a forgotten bill), but can also allow the wearer to allow the distant spouse to mediate the perception of reality. Such mediation may range from simple annotation of objects in the `reality stream', to completely altering the perception of reality.
. Perhaps the most interesting example of synthetic synesthesia was the addition of a new human sensory capability based on miniature wearable radar systems combined with intelligent signal processing. In particular, the author developed a number of vibrotactile wearable radar systems in the 1980s, of which there were three primary variations:
Such simple systems as these suggest a future in which intelligent signal processing, through the embodiment of humanistic intelligence, becomes environmentally aware. It is misleading to think of the wearer and the computer with its associated input/output apparatus as separate entities. Instead it is preferable to regard the computer as a second brain, and its sensory modalities as additional senses, which through synthetic synesthesia are inextricably intertwined with the wearer's own biological sensory apparatus.
Furthermore, the apparatus forms a symbiotic relationship with its host (the human), in which the high-level intelligence arises on account of the existence of the host (human), and the lower-level computational workload comes from the signal processing hardware itself.
The emphasis of this paper was on Personal Imaging, to which the application of HI gave rise to a new form of intelligent camera system. This camera system was found to be of great use in both photography and documentary video making. Its success arose from the fact that it (1) was simpler to use than even the simplest of the so-called ``intelligent point and click'' cameras of the consumer market (many of which embody sophisticated neural network architectures), and (2) afforded the user much greater control than even the most versatile and fully-featured of professional cameras.
This application of HI took an important first step in moving from the `point and click' metaphor, toward the `look and think' metaphor -- toward making the camera function as a true visual memory prosthetic which operates without conscious thought or effort, while at the same time affording the visual artist a much richer and complete space of possibilities.
A focus of HI was to put the human intellect into the loop but still maintain facility for failsafe mechanisms operating in the background. Thus the personal safety device, which functions as a sort of ``black box'' monitor, was suggested.
What differentiates HI from environmental intelligence (ubiquitous computing, reactive rooms, and the like, is that there is no guarantee environmental intelligence will be present when needed, or that it will be in control of the user. Instead, HI provides a facility for intelligent signal processing that travels with the user. Furthermore, because of the close physical proximity to the user, the apparatus is privy to a much richer multidimensional information space than that obtainable by environmental intelligence.
Furthermore, unlike an intelligent surveillance camera that people attempt to endow with an ability to recognize suspicious behaviour, WearComp takes its task from the user's current activity, e.g. if the user is moving, it's taking images; if the user is still, it's not taking in new orbits based on the premise that the viewpoint changes, etc.
Systems embodying HI are:
Dr. Carter volunteered freely of his time to help in the design of the interface to WearComp2 (the author's 6502-based wearable computer system of the early 1980s), and Kent Nickerson similarly helped with some of the miniature personal radar units and photographic devices involved with this project throughout the mid 1980s.
Much of the early work on biosensors and wearable computing was done with, or at least inspired by work the author did with Dr. Nandegopal Ghista, and later refined with suggestions from Dr. Hubert DeBruin, both of McMaster University. Dr. Max Wong of McMaster university supervised a course project in which the author chose to design an RF link between two 8085-based wearable computers which had formed part of the author's ``photographer's assistant'' project.
Much of the inspiration towards making truly wearable (also comfortable and even fashionable) signal processing systems was inspired through collaboration with Jeff Eleveld during the early 1980s.
Bob Kinney of US Army Natick Research Labs assisted in the design of a tank top, based on a military vest, which the author used for a recent (1996) embodiment of the WearComp apparatus worn underneath ordinary clothing.
Additional Thanks to VirtualVision, HP labs, Compaq, Kopin, Colorlink, Ed Gritz, Miyota, Chuck Carter, and Thought Technologies Limited for lending or donating additional equipment that made these experiments possible.