Method for determining the transmission coefficient of clothing

Abstract

The invention relates to a method for determining the transmission coefficient (T) of clothing worn by a person (2, 3) with respect to massage movements emanating from a seat (4, 5) with a massage function. The method according to the invention is characterized in that the person (2, 3) is detected via an image acquisition device and an image of the person (2, 3) is captured, after which the clothing is segmented and identified in the captured image of the person (2, 3), and a number of clothing layers are determined. A transmission coefficient (T) is then determined for each layer of clothing, and from this, an overall transmission coefficient (T) of the clothing worn by the person (2, 3) is determined. The specified transmission coefficient (T) can be used to control a massage function (13) in a vehicle seat (4, 5).

Description

[0001] The invention relates to a method for determining the transmission coefficient of clothing according to the type defined in more detail in the preamble of claim 1. The invention also relates to the use of such a method.

[0002] German patent application DE 10 2022 105 007 A1 discloses a method for operating a seat massage device in a motor vehicle. The massage function is adjusted based on various parameters. One of these parameters can be the person's clothing, which is at least indirectly detected by a control unit of the seat.

[0003] The aforementioned German document acknowledges that clothing affects the massage function, allowing this to be taken into account. However, the problem with the document lies in the lack of a detailed description of the data collection via the seat's control unit, which is highly imprecise.

[0004] The object of the present invention is therefore to provide an improved method for determining the transmission coefficient of clothing for mechanical and / or electromagnetic waves, for example, for the mechanical impulses emanating from a massage function. In this process, either the damping characteristics of the clothing or its transmission coefficient can be taken into account. The damping coefficient and the transmission coefficient always add up to 1.

[0005] According to the invention, this problem is solved by a method with the features in claim 1, and in particular in the characterizing part of claim 1. Advantageous embodiments and further developments are described in the dependent claims. Furthermore, the problem can also be solved by using the method according to the invention.

[0006] The inventive method thus determines the transmission coefficient of clothing, which is related to the attenuation coefficient of this clothing as explained above, as precisely as possible. According to the invention, the person is detected by an image acquisition device, i.e., their presence is recorded. Subsequently, an image of the person is captured, whereby a sequence of images or a video would also be conceivable in principle. The captured image of the person is then segmented to identify the clothing and determine the number of clothing layers. Following this, the transmission coefficient of each clothing layer is determined, which, according to a particularly advantageous embodiment, can be done using a trained artificial intelligence (AI)-based model (AI model). Finally, an overall transmission coefficient of the clothing worn by the person is determined from the transmission coefficients of the individual clothing layers.

[0007] Based on this transmission coefficient, or the attenuation caused by the clothing, appropriate settings can be made. For example, the heat intensity of a seat heater can be increased accordingly if the transmission coefficient is very low, so that the person feels the same warmth regardless of the thickness of their clothing. A similar principle can be used to adjust a seat's massage function, ensuring that the desired massage intensity is felt regardless of the clothing worn.

[0008] According to a highly beneficial training course, it can be implemented that image areas not belonging to the person are masked and ignored during person detection. A deep learning model can be used for this purpose, according to another highly beneficial training course.

[0009] According to a further advantageous embodiment, the segmentation can be carried out into distinct sub-areas of the clothing in order to better identify clothing layers. According to a particularly favorable embodiment of the inventive method, it can also be provided that, when detecting clothing layers, clothing parts belonging to the clothing are grouped into a segment, and segments not belonging to the clothing are removed accordingly, after which the remaining segments are considered the number of clothing layers. In this way, those segments that definitely do not belong to the clothing can be separated from the segmented image, and other segments that belong to the same garment based on texture, color, or the like, even if they are not connected (for example, a jacket open at the front), can be grouped together as a single garment. The individual segments, orGroups of segments then automatically represent the individual layers of clothing.

[0010] According to a further very advantageous embodiment, it can also be provided that optically hidden layers of clothing are detected by means of a deep learning model, wherein the model has a certain semantic understanding that typically further layers of clothing are present under a winter jacket.

[0011] According to another highly advantageous embodiment, in which the AI ​​model is used to estimate the transmission coefficients per layer, the estimation of the transmission coefficients per layer can be based on the material and presumed thickness of the detected clothing layer, for which the trained AI model is used. This model can, for example, estimate different materials based on surface appearance, such as fur, leather, wool, fleece, corduroy, cotton, denim, or the like. It can then estimate the various effects on the damping of vibrations and / or heat according to its training data.

[0012] The transmission coefficient of the clothing as a whole can then be calculated from the product of the transmission coefficients of the individual layers.

[0013] As already mentioned, it is particularly advantageous to use the inventive method to control a massage function for the detected person in a vehicle seat accordingly, whereby the application can of course also be used for massage devices which are used outside of a vehicle.

[0014] Further advantageous embodiments of the method according to the invention also result from the exemplary embodiment, which is described in more detail below with reference to the figures.

[0015] This shows: Fig. 1. A schematic representation of two persons in a vehicle interior to illustrate the procedure in several individual steps; and Fig. 2 a schematic representation of a possible procedure.

[0016] Massage functions, such as those found in vehicle seats, as described in the aforementioned state of the art, have the problem that the perceived massage intensity depends significantly on the thickness of the clothing. For example, a person wearing a thick winter jacket will hardly notice the massage, whereas someone wearing only a T-shirt will feel it very clearly. In practice, different massage programs have different intensities. For instance, the massage intensity in an "Energizing" mode is typically much stronger than in a "Relaxing" mode. However, the "Energizing" mode would no longer be truly invigorating when wearing a thick winter jacket due to the cushioning properties of the clothing.In contrast, the “Relaxing” mode might be perceived as too strong by people who are only wearing a T-shirt or shorts, and would therefore have a more invigorating than relaxing effect.

[0017] In order for the individual massage programs to achieve their desired effect, it is essentially necessary that the passenger wears the "correct" clothing expected by the programmers or applicators of the massage function.

[0018] In principle, the thickness of clothing, as described in the aforementioned prior art, can now be taken into account to enable the use of various massage functions in a vehicle, i.e., a massage mode for relaxation or revitalization. To make this possible regardless of the person's clothing, the massage intensity of the seats should now automatically adjust to the thickness of the clothing.

[0019] To achieve this, a method for capturing the attenuation characteristics or transmission coefficient T of clothing worn by a person is required. For this purpose, an interior camera, typically found in vehicles, is used to capture and analyze the clothing. Such cameras are often already present in vehicle interiors for fatigue detection or video telephony, so the method usually requires no additional hardware. If the application is to be independent of a vehicle, for example in a smart home system with massage seats, a smartphone camera, a laptop camera, or a surveillance camera installed in a room could also be used in a similar way.

[0020] The invention will, however, be described below using the example of a [unclear] in Fig. The vehicle interior shown in section 1 and labelled with 1 is explained without limiting it to this.

[0021] Vehicle interior cameras can be differentiated according to the type of recordings they take: a) normal color images, also called RGB images, as well as b) Infrared images (IR images) as grayscale images.

[0022] At night in the dark, hardly anything is visible in RGB images, but it is in IR images. The grayscale mode works just as well during the day in bright light. Therefore, it is particularly suitable for the present invention because it always works in the same way, regardless of the ambient brightness. For example, a wide-angle camera positioned on the dashboard can be used to capture the... Fig. 1a) Capture the depicted image. It shows the vehicle interior 1, in which two people 2, 3 are located, sitting on the front seats 4, 5. These seats are to have a controllable massage function.

[0023] Depending on the positioning of the interior camera, different viewing angles are possible. For example, with a lower positioning in the vehicle, the thighs can also be visible, allowing legwear to be identified. The clothing visible to the camera in this way allows for a very good assessment of its transmission coefficient or damping properties. Only the buttocks area is not captured; however, it exhibits significantly less variance in terms of damping caused by clothing than the upper body, since most people simply wear underwear and trousers (or a skirt). Therefore, the buttocks area is not explicitly described in the further description of the invention.

[0024] The following explains the individual steps involved in the procedure for determining the attenuation or transmission coefficient of clothing. These steps are shown in the flowchart of the Fig. 2 summarized again.

[0025] In the first step (A), all persons 2, 3 in the image are detected. For example, the algorithm yolo (= you only look once), e.g., in its current version 11, is used for this purpose. The name comes from the fact that it doesn't analyze video sequences, but only individual images / frames. The algorithm therefore doesn't incorporate information from previous frames, but always performs a new, independent analysis. The yolo algorithm is a deep learning model based on a convolutional neural network (CNN).

[0026] If at least one person (2, 3) is recognized, the rest of the image is masked, as described in Fig. 1b) is shown. The masked part 6 of the image is ignored. The masking thus removes the background of the image to avoid unnecessary information, so as not to confuse the other AI models.

[0027] It is conceivable to automatically black out the eye area in an additional step, should anonymization be necessary.

[0028] In the next step, B, the image area containing at least one person (2, 3) is segmented into sub-areas. The goal is to distinguish the individual layers of clothing (e.g., a T-shirt and a jacket over it). The FastSAM algorithm (Fast Segment Anything Model) can be used for this purpose (https: / / arxiv.org / pdf / 2306.12156.pdf). This is an open-source model published in summer 2023, based on the powerful SAM transformer from Meta (formerly Facebook). FastSAM performs real-time instance segmentation and divides the image area into the following: Fig. 1c) Exemplary sub-sections or segments shown. Steps A and B could also be referred to together as Pre-Processing C.

[0029] The following then takes place in block D: Fig. 2. The actual determination of the damping properties of the clothing. In a first sub-step E, the individual layers of clothing are to be identified. However, this requires several sub-steps.

[0030] In substep E1, related clothing items are grouped together.

[0031] Visually separate but related clothing sections should be recognized and grouped as a single unit. For example, if someone is wearing a jacket open, the left and right halves are visually separated. However, these two segments should be treated as one layer of clothing, not two. A suitably trained AI model is used to group individual segments of the same garment. This is also in Fig. 1c) is shown by marking related segments with the same hatching.

[0032] In step E2, segments not belonging to the clothing are removed.

[0033] Person(s) 2, 3 were previously subdivided into their surface segments in steps A and B, which included the head and hands. In step E, however, only the clothing is of interest. Therefore, all segments not belonging to the clothing are to be removed. For this purpose, a computer model was developed that can recognize skin, hair, and accessories such as the suspenders, seatbelt, and steering wheel depicted on person 2, since these objects may be present in the masked sections of the person but are not of interest.

[0034] The segments that do not belong to the clothing are removed, so that the remaining segments only represent the layers of clothing, as shown in Fig. 1d) is shown. Ultimately, the most important factor is the cushioning effect of the clothing between the body and the massage seat 4, 5, i.e., the buttocks and back area. Therefore, all accessories should be removed, as they generally do not extend into this area.

[0035] In step E3, the clothing layers are now counted.

[0036] For each person (2 or 3), the remaining segments are counted. This then also corresponds to the number of clothing layers per person (2 or 3). The number of clothing layers is an indicator of the level of cushioning provided by the clothing, as each additional layer increases the cushioning. In the Fig. 1e) and Fig. 1f) The segments corresponding to the individual clothing layers are shown separately from each other.

[0037] In a further optional step, not shown here, it would also be conceivable to detect additional hidden layers of clothing.

[0038] This step is challenging, but in principle, it would be conceivable to implement functionality that anticipates hidden layers. For example, if someone is wearing a closed winter jacket, only a single layer is visually visible on their upper body. However, it is almost certain that there is at least one more hidden layer of clothing underneath. Accurately identifying such concealed layers is difficult, but for the use case considered here, it is already helpful if hidden layers can be detected in as many cases as possible.

[0039] Advanced image recognition models sometimes also possess semantic understanding, for example, when they are linked to a Large Language Model (LLM), as is the case with chatGPT-4o. If such a model recognizes a winter jacket, for instance, it "understands" that a winter jacket is not normally worn without additional layers of clothing underneath.

[0040] Other trainable indicators include closed buttons or button plackets and closed zippers, as these often appear on outer layers of clothing, where another layer can be expected underneath. Clothing segments protruding from under the sleeves or at the collar are also features that a cognitive model can be trained on. If one of these features is present, the layer counter is incremented by 1.

[0041] However, this step is very complex and offers comparatively little added value. This is because only an average transmission rate can be assumed for the hidden layers, as further analysis is not really possible.

[0042] In step F in the process of Fig. 2. The transmission levels per layer are now estimated.

[0043] In the previous step E, the individual clothing layers per person 2, 3 were identified. Now these will be analyzed in more detail. The goal is to estimate the transmission coefficient of massage waves, i.e., how well massage waves can pass through the clothing. The overall transmission coefficient T of the clothing is the product of the individual transmission coefficients T1 to T2. n of the respective clothing layers. The transmittance is a material property and is defined as the quotient between wave intensity before passing through the clothing (= I). Sitz) and after passing the clothing (= I Haut ): T=SkinISeat The transmittance is therefore a measure of the intensity "transmitted" and takes values ​​between 0 and 1 or 0% and 100%. The attenuation is consequently 1 - T.

[0044] The transmittance T depends on: - from the material / fabric of the clothing layer, - from the thickness of the clothing layer, - from the frequency of the massage wave (which is approximately assumed to be constant), - from the angle of incidence of the massage wave (which is approximately assumed to be constant), - from the pressure exerted by the clothing on the seat (which is approximately assumed to be constant).

[0045] It is not possible to accurately determine the transmission coefficient based solely on camera images. However, an estimate based on the material and thickness of the clothing layers is quite feasible. For the present application, a distinction between three categories is sufficient, each assigned a general value for the transmission coefficient: - low transmission coefficient = high vibration damping - average transmission coefficient = average vibration damping - High transmission rate = low vibration damping

[0046] A computer simulation model is trained to classify each layer of clothing into these three categories. This is done based on the visually identifiable thickness and type of material / fabric. The different types of fabric can be classified as follows based on the three categories above:

[0047] High vibration damping: 1. Fur: • Visual characteristics: Fur has an irregular, soft texture and is often thick and voluminous. It can exhibit a variety of colors and patterns, including solid, spotted, or patterned. In image recognition, fur can be identified by its fluffy and irregular surface, typical fur appearance, and characteristic hair structure. • Vibration damping: Very high, due to the dense structure and natural fibers. • Examples: fur coats, fur jackets. 2. Lined materials: • Optical characteristics: Lined materials often have a smooth outer surface made of polyester, nylon, or other synthetic fabrics. The lining can vary in type and thickness and may consist of down, synthetic insulation materials, or other insulating fabrics. Image recognition can detect the presence of a lining by the characteristic volume and structure of the garment, as well as, where applicable, seams, quilting, or special lining patterns. • Vibration damping: Very high, depending on the type of lining and outer layer. • Examples: down jackets, synthetically lined jackets, skiwear. 3. Leather: • Visual characteristics: Leather has a smooth and robust surface, often thick and heavy. It is available in various colors, mostly black, brown, beige, and other natural tones. In image recognition, leather can be identified by its glossy surface, characteristic leather texture with fine creases in the surface, and, where applicable, visible seams. • Vibration damping: Very good, as leather is dense and heavy. • Examples: Leather jackets, leather coats. 4. Wool: • Visual characteristics: Wool has a soft and often thicker texture with a light fiber structure due to its animal origin. It can be available in various colors, usually solid, heathered, or striped. In image recognition, wool can be identified by its typical light fiber structure. Wool is also frequently used for knitwear and is then easily recognizable by the visible knit pattern. • Vibration damping: Good, as wool absorbs vibrations well due to its thickness and fiber structure. • Examples: Knitwear, wool sweaters, wool coats. 5. Fleece: • Visual characteristics: Fleece has a fluffy, voluminous texture and is often thick and soft. It can be solid-colored or patterned, with various surface textures such as smooth, crinkled, or structured. In image recognition, fleece can be identified by its soft, fluffy surface and characteristic texture. • Vibration damping: Good, as fleece is voluminous and fluffy and therefore effectively dampens vibrations. • Examples: Fleece jackets, sweaters.

[0048] Medium vibration damping: 1. Denim: • Visual characteristics: Denim has a robust, medium-weight texture and is often available in various shades of blue. It has a characteristic jeans structure with visible seams and, where applicable, abrasions or washes. In image recognition, denim can be identified by its typical jeans structure, characteristic color, and visible seams. • Vibration damping: Medium, as denim is tightly woven but not as thick as wool or leather. • Examples: Jeans, denim jackets. 2. Cord: • Visual characteristics: Corduroy has a thick, textured surface with visible grooves or ribs. It is available in various colors and can have a soft or stiff structure, depending on the type of corduroy weave. In image recognition, corduroy can be identified by its characteristic ribbed structure and typical corduroy appearance. • Vibration damping: Medium, as corduroy fabrics have a relatively dense structure which helps to effectively dampen vibrations. • Examples: Corduroy trousers, corduroy jackets.

[0049] Low vibration damping: 1. Cotton: • Visual characteristics: Cotton has a soft, often thinner texture and is available in various colors. It can be smooth or slightly textured, with a somewhat matte effect. Unlike wool, cotton is of plant origin. In image recognition, cotton can be identified by its soft surface and generally finer texture. • Vibration damping: Low, as cotton is light and thin and therefore absorbs vibrations less effectively. • Examples: T-shirts, light shirts. 2. Polyester (unlined): • Optical characteristics: Polyester has a smooth, synthetic texture and is often lightweight and thin. It can be glossy or matte. In image recognition, polyester can be identified by its smooth surface, fine synthetic structure, and any visible seams. • Vibration damping: Low, especially in thin form, as it does not have enough mass to effectively dampen vibrations. • Examples: Lightweight rain jackets. 3. Linen: • Visual characteristics: Linen has a light, textured feel and is often prone to wrinkling. It is frequently available in natural colors and can be smooth or slightly textured. In image recognition, linen can be identified by its characteristic light texture, its mostly natural coloring, and its high susceptibility to wrinkling. • Vibration damping: Very low, as linen is relatively thin and light. • Examples: Summer shirts, light trousers. 4. Silk: • Optical characteristics: Silk is smooth and shiny, light and thin, and available in various colors. In image recognition, silk can be identified by its smooth surface, its light and supple texture, and its typical high-gloss appearance. • Vibration damping: Very low, as silk is thin and light and hardly absorbs vibrations. • Examples: Blouses, dresses.

[0050] To create an AI model that can reliably identify substance classes, a very large number of images are needed for training. These can be obtained from the internet. However, the images must also be labeled. Manual labeling would be extremely time-consuming. Therefore, the images can be semi-automatically labeled by another AI model that has already been pre-trained. Alternatively, one could also use labeled image databases.

[0051] In the following step G, the overall transmission level of the clothing is determined.

[0052] The estimated transmission values ​​T per clothing layer determined in the previous step F are now to be combined to obtain an overall transmission value. This is done simply by multiplying the individual values.

[0053] Even if individual layers of clothing are worn open at the front, such as an open jacket, this is irrelevant because all layers are normally closed and overlapping in the back area, and there are usually no openings there. Since only the back area rests on the seat, only this area is relevant for the massage function. The reason for combining the individual values ​​into an overall transmission value is to ensure that the massage function's control is only extended with one additional scalar value and not complicated by too many new parameters.

[0054] Steps A, B, E, F, and G are based on the YOLO principle (you only look once). This means that only one frame is evaluated individually—not a video sequence, but individual images that are evaluated completely independently. However, to reduce the potential for errors, it is recommended to perform several evaluations sequentially and calculate the average (averaging). This can be done in a final post-processing step H, where, for example, ten evaluations are performed consecutively in step I, with a time interval of, say, 5 seconds, as indicated by arrow 10, and the average of these individual values ​​is used as the result.

[0055] It is recommended to perform the procedure when the massage function is activated. However, it is possible that a vehicle occupant may change their clothing status during the journey (e.g., remove their jacket). Therefore, restarting and repeating the process at a set interval, e.g., every 5 minutes, is advisable.

[0056] The described procedure is carried out individually for each identified person, as indicated by arrow 11.

[0057] The described method can now be used to determine the overall transmission level of the clothing. This value should, as described in Fig.Figure 3 is used to improve the massage function. The starting point of a massage control 12 is the intended massage intensity 13. To achieve this intensity for at least one person 2, 3, regardless of the clothing symbolized by box 14 and its overall transmission coefficient T, the overall transmission coefficient T is determined via box 15 in the sense described above. The massage control 12 is then extended by a correction term 16, which compensates for the damping effect of the clothing. This correction term corresponds to the factor 1T with a transmission coefficient T range of 0 < T ≤ 1. Consequently, a massage intensity of 17 is always only increased by the correction term – except in the limiting case T = 1, which corresponds to a naked person without clothing. In this case, the massage intensity of 17 would remain the same.

[0058] The massage intensity 18 perceived by person 2, 3 therefore always corresponds exactly to the massage intensity 13 intended by the massage control 12.

Claims

[1] Method for determining the transmission rate (T) of clothing worn by a person (2, 3) for massage movements emanating from a seat (4, 5) with massage function, characterized by , that the person (2, 3) is detected via an image acquisition device and an image of the person (2, 3) is captured, after which the clothing in the captured image of the person (2, 3) is segmented, identified and a number of clothing layers are determined, after which a transmission coefficient (T) for each clothing layer (8) is determined, and from this an overall transmission level (T) of the clothing worn by the person (2, 3) is determined. [2] Method according to claim 1, characterized by , that the transmittance (T) for each clothing layer (8) is determined using a trained AI model. [3] Method according to claim 1 or 2, characterized by, that when detecting the person (2, 3), image areas (6) not belonging to the person are masked and ignored. [4] Method according to claim 3, characterized by , that the image areas not belonging to the person (2, 3) are identified using a deep learning model. [5] Method according to any one of claims 1 to 4, characterized by that the segmentation into distinct sub-areas of clothing takes place. [6] Method according to any one of claims 1 to 5, characterized by , that when recording clothing layers, clothing items belonging together are grouped into a common segment and segments not belonging to the clothing are removed, after which the number of clothing layers is derived from the remaining segments. [7] Method according to any one of claims 1 to 6, characterized by, that the existence of optically hidden clothing layers is estimated using a large-language model. [8] Method according to any one of claims 2 to 7, characterized by , that the estimation of the transmission coefficients (T) per clothing layer is based on the material and the expected thickness of the detected clothing layer, for which the trained AI model is used. [9] Method according to any one of claims 1 to 8, characterized by , that the total transmittance (T) is formed as the product of the individual transmittances (T). [10] Use of the total transmission level (T) determined according to a method according to one of claims 1 to 9 for controlling a massage function (13) for the detected person (2, 3) in a vehicle seat (4, 5).

Images