TW201426466A - Optical touch system - Google Patents

Abstract

An optical touch system is disclosed to identify the location of an operating object in a three dimensional space. The system includes a first camera, a second camera and a processing unit. The first camera and the second camera obtains a first two dimensional data and a second two dimensional data from the three dimensional space respectively, and the processing unit identifies the location of the operating object by using the data. In detail, the processing unit executes a processing mechanism described as following to calculate the location. First, a three dimensional coordination is established by using the two dimensional data. Second, each of the gradient variation in the two dimensional data is applied to identify the location of the operating object. Third, each of the shape information in the two dimensional data is applied to identify the location of the operating object. Forth, the parallax pitch of the operating objects in the two dimensional data is applied to identify the actual location of the operating object in the three dimensional space.

Description

Optical touch system

The present invention relates to a touch technology, and more particularly to an optical touch technology.

Please refer to FIG. 1 , which is a schematic diagram of the operation principle of a conventional desktop optical touch device. In the first embodiment, the desktop optical touch device 100 mainly includes a projector 110, an infrared receiving lens 120 and an infrared emitter 130. The principle is as follows: First, the projector 110 defines an operation plane 101 on the desktop, and the operation plane 101 can display the multimedia content to be displayed by the projector 110. Then, the infrared emitter 130 lays a layer of infrared light curtain on the operation plane 101; next, when the operator 102, such as a finger or a pen, moves on the operation plane 101, the infrared rays hit the operator 102 and then reflect to the infrared rays. The lens 120 is received. Therefore, the infrared receiving lens 120 can obtain information that a specific position is an infrared bright spot on a plane, and the specific position is the position of the operating object 102. However, such a touch technology that detects the position of the object 102 by infrared light is susceptible to infrared interference in the external ambient light.

Therefore, the inventors of the present invention have collected the relevant materials in view of the above-mentioned deficiencies, and have designed and evaluated the optical touch system through continuous trial and modification through multi-party evaluation and consideration, and with years of experience accumulated in the industry. .

According to an embodiment of the present invention, an optical touch system is provided for identifying a position of an operator in a three-dimensional operation space, and includes a first photographic lens, a second photographic lens, and an arithmetic unit. The first photographic lens is used to obtain a first two-dimensional plane information from the three-dimensional operation space; the second photographic lens is used to obtain a second two-dimensional plane information from the three-dimensional operation space. The operation unit uses the first two-dimensional plane information and the second two-dimensional plane information to identify the position of the operation object; wherein the operation unit performs a processing mechanism to identify the position of the operation object, and the processing mechanism includes the following steps: Step 1 A two-dimensional plane information and a second two-dimensional plane information establish a three-dimensional coordinate system; in step two, the operator is identified by the gradient change in the first two-dimensional plane information and the second two-dimensional plane information; step three, the first two The dimension information and the linear information in the second two-dimensional plane information identify the operator; in step four, the position of the operator is defined by the parallax distance between the first two-dimensional plane information and the second two-dimensional plane information.

It should be noted that, in other embodiments of the present invention, when the processing unit performs the processing mechanism, the operation object is established in advance in the operation of identifying the gradient in the first two-dimensional plane information and the second two-dimensional plane information. One direction gradient map (HOG) model, and a direction gradient map (HOG) model is used as a filter to identify an operator in the first two-dimensional plane information and the second two-dimensional plane information, respectively.

For example, the above-described directional gradient map (HOG) model can be composed of a root filter and a plurality of partial filters. In an alternative design, multiple partial filters can each have a corresponding deformation cost to anchor (anchor) is around the root filter.

On the other hand, when the processing unit performs the processing mechanism, when the operator objects are identified by the first two-dimensional plane information and the linear information in the second two-dimensional plane information, a linear model of the operator may be established in advance to describe the operator. And contouring, and then identifying the operator in the first two-dimensional plane information and the second two-dimensional plane information, respectively. For example, when the processing unit processes the linear information in the first two-dimensional plane information and the second two-dimensional plane information, the plurality of broken edge chains may be repaired and connected into at least one contour segment by at least one section of the bridge, and then based on the operation. The linear model preset by the contour of the object performs a projection comparison of similar weights.

From another perspective, in another embodiment of the present invention, when the stereo coordinate system, the gradient change identification, and the linear information identification function are completed, when an object is involved in the first two-dimensional plane information and the second two-dimensional plane information, And detecting that the parallax distance of the object is between an upper limit value and a lower limit value, the gradient change identification and the linear information recognition function can be performed to determine whether the object is an operation object.

It is claimed that if the parallax distance of the object is lower than the lower limit value, the object non-operating object can be directly determined; if the parallax distance of the object is higher than the lower limit value, it can be directly determined that the object does not trigger a touch event, regardless of the operation. Things or not.

In order to achieve the above object and effect, the technical means and structure adopted by the present invention are described in detail in the embodiments of the present invention as follows.

Please refer to FIG. 2 , which is a schematic structural diagram of the optical touch control system 200 of the present embodiment. In FIG. 2 , the optical touch system 200 includes a first photographic lens 210 , a second photographic lens 220 , and a processing unit 230 . The first photographic lens 210 obtains a first two-dimensional plane information 211 from the three-dimensional space in which the operation plane 101 and the operator 102 are located, and the second photographic lens 220 obtains a second two-dimensional plane from the same three-dimensional space. Information 221. Then, the processing unit 230 detects the specific location of the operator 102 according to the first two-dimensional plane information 211 and the second two-dimensional plane information 221 . To further illustrate, the processing unit 230 must first establish a three-dimensional system to correspond to the stereoscopic operation space in which the operation plane 101 and the operator 102 are located, and then analyze the information of the first two-dimensional plane information 211 and the second two-dimensional plane information 221 The position of the operator 102 is confirmed to achieve the touch effect.

Specifically, the processing unit 230 must execute a processing mechanism 240. The processing mechanism 240 includes at least the following steps: First, step 1 is performed to establish a stereo coordinate system program 241: using the first two-dimensional plane information 211 and the second two-dimensional plane information 221 The difference between them is to create a three-dimensional coordinate system. For example, for the three-dimensional Cartesian coordinate system xyz, the first two-dimensional plane information 211 and the second two-dimensional plane information 221 are the same x-y two-axis plane information obtained from different z-axis height positions. Therefore, the above-mentioned stereo coordinate system can be established by using the aberration of the first two-dimensional plane information 211 and the second two-dimensional plane information 221 . A more detailed three-dimensional coordinate system creation technique will be exemplified later.

Then step 2 is performed to identify the operator program 242 with a gradient: after the three-dimensional coordinate system has been properly established, respectively, from the first two-dimensional plane with the operator 102 intervening In the face information 211 and the second two-dimensional plane information 221, the position of the operator 102 in the three-dimensional coordinate system is recognized. Here, the background picture on the operation plane 101 may be an active picture, such as a liquid crystal display, or a passive projection picture, such as an operation plane 101 having multimedia content defined by the projector 110. Taking the first two-dimensional plane information 211 obtained by the first photographic lens 210 as an example, when the operator 102 appears on the operation plane 101, the processing unit 230 needs to be utilized in the first two-dimensional plane information 211 in the present program. The color gradient of the object frame changes to define the position and shape of the individual objects; thereby, the object conforming to the shape of the preset operator 102 can be locked to obtain position information. Of course, the second two-dimensional plane information 221 is also processed in comparison; a more detailed gradient identification operator technique will be exemplified later.

Next, step 3 is executed to identify the operator program 243 in line: in the gradient identification operator program 242, when the object conforming to the shape of the preset operator 102 is defined by the gradient change, there is a considerable degree of misjudgment possibility. That is, the other objects in the background image on the operation plane 101 that are similar to the operator 102 are mistaken for the operator 102. Therefore, in the present program, the processing unit 230 further utilizes the linear shape of the object shape of each suspected operation object 102 in the foregoing program 242, and further determines the use of the preset operation object 102 linear information to exclude similar operations based on the gradient information. 102 objects. Of course, the second two-dimensional plane information 221 is also processed in comparison; a more detailed line identification operation technique will be exemplified later.

Finally, step 4 is executed to confirm the operator position program 244 with the parallax spacing: Please refer to FIG. 3 together. FIG. 3 is the embodiment of the present invention, and the operator is confirmed by the parallax distance. Schematic diagram of the location. After the position of the operator 102 is respectively taken out from the first two-dimensional plane information 211 and the second two-dimensional plane information 221, the processing unit 230 can define the true position of the operator 102 by using the parallax distance between the two in the program. . Specifically, as shown in FIG. 3 , when the finger is used as the operator 102 , the fingertip touches the projection surface on the left side and the fingertip does not touch the projection surface on the right side. Regardless of whether the fingertip touches the projection surface, the center of the fingertip is almost at the same position on the projection surface for the top view of the upper first photographic lens 210, but for the second photographic lens 220 below, the fingertip Touching the projection surface and the projection surface that has not touched the projection surface will cause the center of the fingertip to be in different positions on the projection surface; the distance generated between the two is the parallax distance. If the operator 102 touches the plane, the parallax spacing 201 is If the operator 102 is away from the plane, that is, the non-touch state, the parallax spacing 202 must be greater than the above threshold. Therefore, the program can thereby determine whether the operator 102 is in a touch state to determine whether to perform various touch functions using the position of the operator 102.

It should be noted that, from the state of the left and right sides shown in FIG. 3, it can be found that the center of the fingertip observed by the upper first photographing lens 210 and the second photographing lens 220 below the fingertip when the fingertip touches the projection surface There is still a slight difference in the position of the center of the fingertip on the projection surface. In fact, this is why the touch point is defined as the center of the fingertip without the tip of the fingertip. The center of the fingertip has a certain thickness, so when the fingertip touches the projection surface, the center of the fingertip determined by the upper first photographing lens 210 and the second photographing lens 220 below is still expected after the mapping operation (Homography). The drop exists because the center of the fingertip does not completely fall on the plane. Since the present embodiment only uses a visible light camera, the content of the projected image has a direct influence on the fingertip detection. When the projected image appears like a finger or the image contains a real finger, the system may be misjudged, so The center of the fingertip is used as a touch point, which can effectively avoid misjudgment caused by objects appearing in the picture, that is, the finger image in the picture may not be regarded as the operator 102 because the correspondence is too perfect.

In other words, the processing mechanism 240 may further include the following steps: using a difference between the images obtained by the first photographic lens 210 and the second photographic lens 220, a stereoscopic system may be suggested; although, in subsequent use, the two lenses are used. There is still a drop in the value of the same object, but it can be corrected. After correction, if the parallax spacing 201 of an object is zero, it means the non-operating object 102, but the image object on the background image. Considering that the correction correction cannot be 100%, the present embodiment can define that the parallax distance 201 of the operator 102 must be greater than the lower limit.

On the other hand, as the operation plane 101 and the stereoscopic operation space where the operator 102 is located vary in depth, the fixed angle may not be used because the first photographic lens 210 and the second photographic lens 220 have different oblique projection angles to the center of the fingertip. The value is used to determine the touch event. In other words, although the height of the fingertip from the projection surface is the same, different parallax occurs because the positions of the first photographic lens 210 and the second photographic lens 220 are different. Therefore, the present embodiment further proposes a dynamic threshold value scheme. Solve the problem; that is, determine whether the fingertip touches the projection surface is determined by the following equation (1): Among them, TH _low ( x , y ) and TH _high ( x , y ) are the upper and lower bounds of the dynamic critical value, which will change as the center of the fingertip appears at different positions on the projection surface. Ideally, TH _low ( x , y ) should be 0 at any position, because the mapping operation (Homography) can correspond to the points appearing on the plane, and there is almost no parallax between the two fields of view; but in reality, it must There may be some errors in the search for the center of the fingertip. If the parallax of the center of the fingertip between the two fields of view is D _touch ( x , y ) when the fingertip touches the projection surface, the embodiment adopts TH _low ( x , y )=0.5 D _touch ( x , y ) and TH _high ( x , y )=1.5 D _touch ( x , y ). In fact, the value of TH _high ( x , y ) affects the accuracy of the touch, that is, how close the fingertip is to the projection surface, it will be judged as a touch event; if the position of the center of the fingertip is more precise, TH _high can be ( x , y ) is changed to the value of D _touch ( x , y ) to provide a higher-accuracy touch determination. Therefore, when the center of the fingertip found by the upper field of view is mapped to the lower field of view by the mapping operation (Homography), if the difference D _ts ( x , y ) is between the critical values, compared with the center of the fingertip found by the lower field of view, Then you can determine that the fingertip touches the projection surface.

In practice, the establishment of the stereo coordinate system program 241 has established a correspondence between the first photographic lens 210 and the second photographic lens 220 in a common plane; therefore, it is only necessary to find an operation object in the field of view of the first photographic lens 210. 102 possible positions and define possible operating points, such as fingertips, through the mapping operation (Homography) can easily determine when the fingertip touches the projection surface, the fingertips in the second photographic lens 220 will fall on these possible In the relative position of the operating point, it is only necessary to find the position of the fingertip in the position of interest in the field of view of the second photographic lens 220 and determine whether the fingertip touches the projection surface. Because the practice only cares about the touch event, when the fingertip has not touched the projection surface, the relative position of the fingertip center determined by the first photographic lens 210 corresponding to the second photographic lens 220 does not appear at the fingertip. So straight It is judged that the fingertip has not touched the projection surface, thereby speeding up the overall calculation speed.

In other words, in the overall operation, if the parallax distance 201 obtained by the first photographic lens 210 and the second photographic lens 220 for an object is greater than an upper limit value, it indicates that the operating object 102 has not touched the operation plane. Therefore, the gradient recognition operator program 242 and the line recognition operator program 243 are not required to be executed in the loop timing. As a whole, at the beginning of the system setup, it is necessary to obtain the stereo coordinate system program 241, the gradient recognition operator program 242, and the linear recognition operator program 243 from the first photographic lens 210 and the second photographic lens 220. In the information, the operator 102 for establishing the stereo system and correcting various parameters is identified, thereby filtering out the tedious background information; after confirming the operator position program 244 by the parallax spacing, the position of the operator 102 is exactly defined, and the stereo system is Complete. For subsequent use, the object position program 244 is confirmed by the parallax spacing, and the object of the first two-dimensional plane information 211 and the second two-dimensional plane information 221 obviously has a parallax distance (the object of the background image does not obviously have parallax). Spacing, unless the system data is offset, whether the parallax distance is between the lower limit and the upper limit; if it is lower than the lower limit, it is defined as a background object or other interferent (such as a mosquito fly) ), instead of the operator 102; if it is higher than the upper limit, it means that no touch event has occurred regardless of whether it is the operator 102 or not, and no subsequent touch manipulation function is available.

More specifically, an arm hand grasps the fist to extend the index finger, and intervenes in the established stereo operating system as an example. The first two-dimensional plane information 211 and the second two-dimensional plane information 221 obtained from the first photographic lens 210 and the second photographic lens 220 can roughly identify several objects having parallax spacing, and the palms or fists between the two images. The parallax distance of the part is greater than the preset upper limit, so it is not processed; and the index finger between the two images should be roughly touched between the upper limit and the lower limit when the touch operation plane 101 is touched. Further, the gradient recognition operator program 242 and the line recognition operator program 243 are activated to confirm that it is the preset operator 102. On the other hand, if the arm clenches the fist, does not extend the index finger, and falls on the operation plane 101, it must have at least one parallax distance between the upper limit value and the lower limit value, so it is necessary to further identify the operator program 242 with the gradient. The operator program 243 is identified in a line shape, excluding it as an eligible operator 102.

Next, the technical details of the above-described stereoscopic coordinate system program 241, the gradient recognition operator program 242, and the line recognition operator program 243 are introduced one by one. Please refer to FIG. 4, which is a schematic diagram of the operation steps of establishing the stereo coordinate system program 241. First, in the first step of the upper left panel, points C and C' are two known camera centers, and the physical meanings of the representatives include, but are not limited to, the first photographic lens 210 and the second photographic lens 220 described above. The plane E and the plane E' are image planes corresponding to points C and C', respectively. The X point is a reference point in the three-dimensional space, which together with the C point and the C' point of the camera center constitutes a pair of axis plane π . Moreover, the line connecting the C point and the C' point of the camera center constitutes a reference line, and the reference line passes through the image planes E and E' respectively, forming the pair of axis points e and e'. In addition, the C and C' points of the camera center are respectively connected to the reference point X, and each of them passes through the image planes E and E' to form another set of plane reference points x and x'. The pair of pivot points e and e' form a pair of axes I and I' with the plane reference points x and x', respectively.

Next, in the second step of the upper right panel, when X moves in the three-dimensional space, the axis plane π is only rotated around the reference line determined by the center of the camera. For the C and C' points of the camera center, X falls at the x position of the E plane and the x' position of the E' plane. Therefore, when X in the three-dimensional space moves along the c → x direction, the reference point X still falls at the same x position for the camera center C point, but for the camera center C' point, the reference point X will Moving on the axis I' of the intersection of the axis plane π and the image plane E'; that is, when a point in the three-dimensional space and a center of the two cameras jointly determine a plane, X moves on the axis plane π Then, as long as the field of view E of the camera C determines a point, it can correspond to a line in the field of view E' of the camera C'. Such a relationship is called a pair-axis geometric relationship.

Therefore, suppose that only a point x is seen from the camera C without knowing its true position X in the three-dimensional space, and it can be known from the axis geometry that x will correspond to a line in the image plane E' of the camera C' field of view. I'll next, the stereo coordinate system program 241 establishes the correspondence between points and points between the image planes E and E' by limiting the position where X moves in the three-dimensional space. Thereby, the first two-dimensional plane information 211 and the second two-dimensional plane information 221 can be organized into one three-dimensional coordinate.

The field of view of the camera C cannot distinguish the position where X moves in the c → x direction, so the stereo coordinate system program 241 first cuts c → x by using a plane that does not pass through the camera center points C and C'. Under the Γ limit of the plane, when the camera C sees the position X of a three-dimensional space at the position of the image plane x, since X is limited to the plane Γ, it cannot move freely along c → x , and it can be known that the point will be x' falling in the field of view of camera C' is no longer a pair of axes I'. In other words, a point-to-point relationship between two cameras can be established through a plane.

Please refer to the third step of the lower right image to establish the correspondence between the two cameras from a plane in the three-dimensional space. Such a relationship is called a plane correspondence. In the plane correspondence, when the field of view of camera C sees x, the intersection of ray c → x and plane Γ can be seen that x is generated by x _Γ on the plane of the ;; that is, it can be x = H _{Γ C} x description of _Γ. On the other hand, x _Γ falls on x' for the field of view of camera C', meaning that it can be described by x' = H _{Γ C'} x _Γ . Combining the relationship between the two cameras to a common plane, the plane correspondence between the two image planes can be established as x' = H _{Γ C'} H _{C Γ} x = Hx ; and, because H _{Γ C} and H _{Γ C '} All are 1-to-1 mapping relationships, so H is a reversible matrix, ie x = H ^-1 x' .

Finally, as shown in the fourth step of the lower left panel, after the correspondence between the two cameras in one plane in three-dimensional space is established by the plane correspondence relationship H, it can be further explored that the reference point X in the three-dimensional space is far from the previous definition. The parallax produced when the plane is flawed. It is stated that after the plane correspondence H between the camera C and the C' is established through the plane ,, for any point x appearing in the image plane E, it can be found in the other image plane E' through the plane correspondence H. Corresponding point =Hx. When X is not on the plane 且 and still falls on the ray c → x , x and x _Γ still correspond to the same point X for C. Therefore, through the plane correspondence H will map X to Above; however, for C', X should actually fall on another point x' to axis I', The difference from x' is the parallax derived from the plane Γ.

Next, the present embodiment describes the technical details of the gradient recognition operator program 242 in FIGS. 5-9. Please refer to FIG. 5 first, and FIG. 5 is a schematic diagram of the operation of the operator 102 on the background screen. It can be observed from Fig. 5 that there is a multi-media on the operation plane 101. The body background picture exists; at this time, if the multimedia background picture is formed by projection, it is bound to directly generate the light and shadow interference 103 on the operator 102, which affects the judgment of the color block. More specifically, the operator 102 to be detected, that is, the finger, is interfered by the image projected by the projector, so in addition to the change of the background, the operator 102 may have a different content depending on the projected image. The change in color. In the dynamic background and the foreground where the color changes with the background, the more stable information is the edge; therefore, the next step is how to accurately find the position of the finger from the edge information.

In practice, in addition to the ambient light source, the projector also produces a dynamic light source, so it will encounter high dynamic range brightness changes. In order to use the edge information effectively, the feature descriptor is used here. Histogram of Orientation Gradients (HOG), which is used to make the edge of an object significantly different in intensity depending on the contrast. The HOG descriptor treats the image as a combination of many small regions and collects the features of these regions. The direction is quantized to handle the simple deformation produced by the object, through normalization between each small region (normalization) To reduce the intensity of the different edges produced by high contrast images, as detailed below.

Please refer to Fig. 6 and Fig. 7 together, which are explanatory diagrams of the direction gradient feature description operation. The HOG descriptor is mainly used to calculate the gradient of the image, and then counts the values in each cell block after the direction quantization, so as to effectively describe the appearance and shape of the object through quantization and local statistics. In the directional gradient characterization operation, gradients are first calculated for each pixel (pixel) of the image using differential filters. Let r(x, y) and θ (x, y) denote the magnitude and orientation of the intensity gradient at the pixel pixel(x, y), respectively. The gradient orientation of each pixel is quantized into p bins using the following equation (2) or (3) (here p=9 is illustrated): Among them, B ₁ represents the quantization mode of contrast sensitivity, and B ₂ represents the quantization mode of contrast insensitive. Then, let b {0,1,..., p -1}, the feature map of the image is done by the following equation (4) (B value means B ₁ or B ₂ ):

As stated above, the function F indicates that an edge is seen as a particular direction voting to the corresponding bin according to the strength of the edge. After processing the map for each pixel, the information is then integrated to divide the image into a combination of many small cells. Statistics through these edges can effectively describe the characteristics of an object. For example, a small cell is an 8x8 pixel, and each region can be viewed as a vector of dimension p. In order to reduce the occurrence of objects happening between the cut areas, the edge intensity of each pixel is interpolated to other adjacent four areas in addition to the area where the position is contributed to itself. Since the value of the gradient has a great relationship with the contrast of the image, normalization is also performed between the regions, that is, a small cell and three adjacent small regions are regarded as one region. Block.

Next, let δ , γ {-1, 1} The normalization factors are as shown in the following equation (5): Among them, each region and several different adjacent regions are normalized, thus producing four values after naturalization, as shown in the following equation (6): T _α ( v ) in the equation represents that each element in the vector v is selected to be smaller after comparison with α .

Through these processes, the corresponding feature map can be obtained for any of the input images. Finally, the HOG model 104 of the operator 102 can be obtained by using a Support Vector Machine (SVM) to perform a supervised learning method, as shown in FIG.

Finally, using the known HOG model 104, the position of the operator 102 is detected using a sliding-window on any newly input image by treating the HOG model 104 as a filter ( Filter F), this filter contains vectors of different weights. The detection process uses the filter (filter F) to start the sub-window of the (x, y) from the upper left corner of the image feature map H, and calculate the score by using the inner product. The following equation (7) shows: After calculating the score for each position of the complete image, the object to be detected can be found through a simple threshold value.

However, although the HOG model 104 can detect the position of the operator 102, it is quite sensitive to the rotation of the object. Taking Fig. 7 as an example, when there are many similar appearances 105 in the background picture, the HOG model 104 is susceptible to object rotation, and misjudgment is prone to occur. Specifically, when the gradient orientation is voted into the nearest bin, the result of the voting generated when the object is rotated may be different from what was originally expected. In real-world applications, the operators 102 often have slight rotations, and the feature maps produced by these rotations are almost all different. When calculating the score, only objects in the same direction as the HOG model 104 can obtain a higher score, thereby identifying it as the operator 102. When the angle of rotation differs from the HOG model 104, the result of calculating the score will also get a lower score. Therefore, as long as the angle of rotation of the operator 102 is too large, the calculated score is insufficient to separate the operator 102 from the background, thereby producing a number of false alarms at the similar appearance 105.

In order to overcome the above drawbacks, the present embodiment further introduces a partial model based HOG operation method in the gradient identification operator program 242; in detail, since the operator 102 usually has a slight variation, the original HOG model is excluded. In addition to 104, some movable partial models are additionally added to change position as part of the operating object 102 changes. For details, please refer to FIG. 8. FIG. 8 is a schematic diagram of the HOG operation mode based on the partial model. The original HOG model 104 is used as the root filter, and a plurality of partial filters 107 are added. ) to describe the part where the object changes greatly, and the partial filter 107 moves and Not limited, each partial filter 107 has a corresponding deformation cost 108.

Specifically, the filter 106 (part filter) and the partial filter 107 (part filters) are respectively calculated at different resolutions, that is, the filter 106 (root filter) finds the approximate object position and then passes through the high resolution. Part of the filter 107 (part filters) to cover a small part of the object changes. At this time, although the appearance of the operator 102 may change, the variable portion is not unrestricted movement, and each partial filter 107 has an anchor position to ensure these portions. The movement is reasonable.

In other words, the anchor is the most likely position for the variable part to appear. The farther the part is from the anchor, the higher the cost will be, as shown in the following equation (8). Said:

Where p ₀ denotes the present filter 106, and p ₁ , . . . , p _n denotes n partial filters 107, function F '. ( H , p _i ) denotes a feature vector in a sub-window in which the upper left corner of the feature map H is p; and F ' is a transpose after F is lined up. Each partial model consists of three parts, denoted by ( F _i , v _i , d _i ), where F _i is the filter of the i-th part; v _i is a two-dimensional vector representing the i-th The fixed anchor position of the part; d _i is a four-dimensional vector used to represent the coefficient of the quadratic function, and its purpose is to represent the cost function corresponding to the partial leaving anchor. When d _i = (0, 0, 1, 1), the deformation cost of the i-th part is only the distance between the part and the fixed two-dimensional space between the anchors.

Next, in order to calculate the position of the body corresponding to the best part at each position, it can be expressed by the following equation (9): The remaining steps are the same as the steps of the HOG operation described above, and the position of the object can be detected by the selection of the threshold value.

Finally, the present embodiment describes the technical details of the linear recognition operator program 243 in the manner of Figures 9-11. As mentioned above, although the HOG model of the filter 106 (part filter) and the partial filter 107 can accurately find the position of the operator 102, the concept of operation based on a small cell may be insufficient. The information of the operator 102 is processed to a pixel level.

Taking the operator 102 as a finger as an example, the accuracy of the center of the fingertip directly affects the accuracy of the touch point position and the determination of the touch; in order to improve the accuracy of the system operation, the accuracy of the fingertip center is improved. Demand. Therefore, in the linear identification operator program 243, the present embodiment describes the outline of the finger in a line model, and uses the contour known by the finger to detect the malfunction of the HOG model 104 including the partial model. (false alarms) Remove and use the outline of the finger to find the center of the fingertip. The feature of the shape model is that in addition to detecting objects, the shape of the object can be described. Since the shape model is constructed in pairs of adjacent segments (PAS), it is necessary to teach the local contour features (P). The basic principle of AS detection is that the contour segment network (CSN) is as follows: First, because the edge detection is easily affected by image brightness and contrast, an edge is often cut. A few pieces of broken edge chains. Therefore, the way to connect the edge chain is to record these cuts. For example, if there is a point in the edge chain C2 falling within the search range of the edge chain C1, the edge chain C1 will be connected with the edge chain C2; here, refer to Figure 9, which is the two edge chains. Connection diagram. It is worth noting that, as shown in Fig. 9, the significance of the search range 109 using the ladder type is that as the distance from the end point of the edge chain is further away, the uncertainty of the position of the other edge chain also increases.

Next, after joining the edge chains, the edge chains must be segmented into contour segments that approximate the line. At this point, the segment bridging of the contour segment defines its own tangent-continuous connection, and segment bridging can effectively repair the broken edge. The contour network segment (CSN) is mainly constructed on the use of edge chain links and contour segments, which record the missing edge information by connecting the edge chain links. The entire network segment is reorganized based on the contour segments using line segments that approximate the line. For details, please refer to Figure 10, which is a schematic diagram of the operation of the contour network segment, which depicts the following six patching rules:

(1) As shown in Rule 1 in Figure 10: on the same edge chain One of the line segment front points is connected to the back end of the other line segment.

(2) As shown in Rule 2 in Figure 10: When two edge chains (C1, C2) are connected at the end points, the line segment of the edge chain C1 is connected to the line segment of the edge chain C2.

(3) As shown in Rule 3 in Figure 10: When considering the connection of the T-junction, the line segment of the edge chain C1 is connected to the nearest end point of the two line segments of the edge chain C2.

(4) As shown in Rule 4 (Rule 4): Let s be the segment bridge connecting the edge chain C1 and the edge chain C2, the segment bridge s will be after the front end point of the edge chain C1 line segment, and at the edge chain Before the rear end of the C2 line segment.

(5) As shown in Rule 5 in Figure 10: Two line segments on the same edge chain are connected together if there are consecutive endpoints.

(6) As shown in Rule 6 (Rule 6): Consider a segment bridge s without a front-end point connection, because it covers the front end point of the edge chain C2 and continues to the back segment of the edge chain C2, so The edge chain C2 is connected to the other edge chain C3, and the segment bridge s is also connected to the line segment of the edge chain C3.

As mentioned above, because each edge chain is likely to be connected to several other edge chains, the contoured network segment (CSN) is a rather complex mesh structure, and the local features can describe a pair of connected segments. The relationship between the two, which in turn simplifies the complexity of the entire problem. That is, the line segment between the two is considered to be a partial contour feature (PAS). Next, explain how to use the mathematical equation to describe the local contour feature (PAS) as follows: local contour feature (PAS) P = ( x , y , s , e , d ) is a position (x, y), a step The parameter s, an intensity parameter e and a descriptor d = ( θ ₁ , θ ₂ , l ₁ , l ₂ , The composition is as follows; please refer to Fig. 11, which is a schematic diagram of the operation of the local contour feature. In Fig. 11, since the local contour feature (PAS) connects the gaps between the edge chains by segments, the local contour feature (PAS) can detect the object stably. Since the local contour feature (PAS) contains only two separate line segments, it can effectively and simply restore the boundary of the object and remove the cluttered background.

The difference between the local contour features (PAS) is calculated by D(P, Q), and D(P, Q) represents the degree of dissimilarity between the two PAS descriptors dp, dq, as shown in the following equation (10):

The first term of equation (10) is to calculate the difference between the relative positions of the two line segments, and the second term D _θ [0, π /2] calculates the difference in orientation between line segments, and the third term calculates the difference in length. For example, a finger in an operation may have a slight rotation, so Equation (10) can be rewritten to the following equation (11) as needed to account for the slight difference between the PAS caused by a slight rotation: In equation (11), the relative orientation difference between the two PAS features is further considered.

Therefore, the linear shape of the operator program 243 is determined by predefining an object to be detected, such as the operator 102, and the PAS features contained therein are repeatedly displayed in the training material. A code book C={ t _i } is created by collecting these PAS features. This code book contains the PAS descriptor of the format T{ t _i }, and V _i is the projection space corresponding to t _i (voti ng space). Each voting space is a three-dimensional space in which two dimensions are used as positions and the other dimension is feature size. Each PAS is assigned to many formats T as long as their PAS variability is small enough, as shown in the following equation (12): γ : T = { t _j | D ( d , t _j ) < γ } (12) The weight of the projection is determined by the difference between the PAS and the format. The smaller the difference, the higher the weight, as described in the following equation (13): e . (1- D ( d , t _j )/ γ ) (13)

However, the above description is only for the various embodiments of the present invention, and is not intended to limit the scope of the present invention. Therefore, the simple modifications and equivalent structural changes of the present invention and the contents of the drawings should be included in the same manner. Within the scope of the patent of the present invention.

100‧‧‧Tabletop optical touch device

110‧‧‧Projector

120‧‧‧Infrared receiving lens

130‧‧‧Infrared emitter

101‧‧‧Operational plane

102‧‧‧Operator

103‧‧‧Light and shadow interference

104‧‧‧HOG model

105‧‧‧similar appearance

106‧‧‧This filter

107‧‧‧Partial filter

108‧‧‧Transformation costs

109‧‧‧Search range

200‧‧‧Optical touch system

210‧‧‧ first photographic lens

211‧‧‧ first two-dimensional information

220‧‧‧second photographic lens

221‧‧‧Second 2D Plane Information

201, 202‧‧‧ Parallax spacing

230‧‧‧Processing unit

240‧‧ ‧ treatment mechanism

241-244‧‧‧Steps

Figure 1 is a schematic diagram of the operation principle of a conventional desktop optical touch device.

FIG. 2 is a schematic structural view of the optical touch system 200 of the present embodiment.

Fig. 3 is a schematic view showing the principle of confirming the position of the operator by the parallax distance in the present embodiment.

Fig. 4 is a schematic diagram showing the operation steps of the stereo coordinate system program 241.

Figure 5 is a schematic diagram of the operation of the operator 102 on the background screen.

Figure 6 is an explanatory diagram of the direction gradient feature description operation.

Figure 7 is an explanatory diagram of the directional gradient characterization operation.

Figure 8 is a schematic diagram of a HOG operation based on a partial model.

Figure 9 is a schematic diagram of the connection of two edge chains.

Figure 10 is a schematic diagram of the operation of the contour network segment

Figure 11 is a schematic diagram of the operation of the local contour feature.

110‧‧‧Projector

101‧‧‧Operational plane

102‧‧‧Operator

200‧‧‧Optical touch system

210‧‧‧ first photographic lens

211‧‧‧ first two-dimensional information

220‧‧‧second photographic lens

221‧‧‧Second 2D Plane Information

230‧‧‧Processing unit

240‧‧ ‧ treatment mechanism

241-244‧‧‧Steps

Claims (10)

An optical touch system for identifying a position of an operation object in a three-dimensional operation space, comprising: a first photographic lens for acquiring a first two-dimensional plane information from the three-dimensional operation space; a second photographic lens for acquiring a second two-dimensional plane information from the three-dimensional operation space; and an operation unit for utilizing the first two-dimensional plane information and the second two-dimensional plane information, Identifying the location of the operator; wherein the computing unit performs a processing mechanism to identify the location of the operator, the processing mechanism includes the following steps: Step 1: using the first two-dimensional plane information and the second two-dimensional plane The information establishes a three-dimensional coordinate system; step two: identifying the operation object by using the first two-dimensional plane information and the gradient change in the second two-dimensional plane information; and step three: using the first two-dimensional plane information and the first The linear information in the two-dimensional plane information identifies the operator; and the fourth step: defining the parallax distance between the first two-dimensional plane information and the second two-dimensional plane information The position of the operator. The optical touch system of claim 1, wherein the operation unit is configured to perform a step gradient map (HOG) of the operator in advance. And using the directional gradient map (HOG) model as a filter, respectively identifying the operator in the first two-dimensional plane information and the second two-dimensional plane information. The optical touch system of claim 2, wherein the directional gradient map (HOG) model is composed of a root filter and a plurality of partial filters. The optical touch system of claim 3, wherein the plurality of partial filters each have a corresponding deformation cost to anchor an anchor in the filter (root filter) Around it. The optical touch system of claim 1, wherein the computing unit is configured to perform a step of the third step to establish a linear model of the operating object to describe the contour of the operating object, and then respectively The two-dimensional plane information and the second two-dimensional plane information identify the operator. The optical touch system of claim 5, wherein the computing unit is configured to process a plurality of broken edges when processing the linear information in the first two-dimensional plane information and the second two-dimensional plane information. The chain is patched into at least one contour segment with at least one section of the bridge. For example, the optical touch system described in claim 1 of the patent scope, the processing mechanism is more The method includes: when an object is involved in the first two-dimensional plane information and the second two-dimensional plane information, and detecting that the parallax distance of the object is between an upper limit value and a lower limit value, performing step 2 and step three To determine whether the object is the operator. The optical touch system of claim 7, wherein the processing mechanism further comprises: determining that the object is not the operator when the parallax spacing of the object is lower than the lower limit. The optical touch system of claim 7, wherein the processing mechanism further comprises: when the parallax spacing of the object is higher than the lower limit, directly determining that the object does not trigger a touch event. An optical touch system for identifying a position of an operation object in a three-dimensional operation space, comprising: a first photographic lens for acquiring a first two-dimensional plane information from the three-dimensional operation space; a second photographic lens for acquiring a second two-dimensional plane information from the three-dimensional operation space; and an operation unit for utilizing the first two-dimensional plane information and the second two-dimensional plane The surface information identifies the location of the operator; wherein the computing unit performs a processing mechanism to identify the location of the operator, the processing mechanism includes the following steps: Step 1: using the first two-dimensional plane information and the second The two-dimensional plane information establishes a three-dimensional coordinate system; step two: pre-establishing a direction gradient map (HOG) model of the operator, and using the direction gradient map (HOG) model as a filter, respectively in the first two-dimensional Identifying the operator in the plane information and the second two-dimensional plane information; Step three: pre-establishing a linear model of the operator to describe the contour of the operator, and then respectively in the first two-dimensional plane information and the second Identifying the operator in the two-dimensional plane information; Step 4: defining the position of the operator by the parallax distance between the first two-dimensional plane information and the second two-dimensional plane information; Step 5: When one The object intervenes the first two-dimensional plane information and the second two-dimensional plane information, and detects that the parallax distance of the object is between an upper limit value and a lower limit value, and performs step two and the third step to determine Determine if the object is the operator.