Smart flap for animals

Abstract

The invention relates to a door (1000) for an animal, preferably a pet, such as cats, dogs or the like, in particular an animal flap, preferably a cat flap, comprising: • a housing (1100, 1200, 1300), • a flap (1400) having a lock, and • electronics (1500) having a first sensor, preferably a camera, which is designed to carry out at least one of the following functions, in particular during the day and / or night: • a) animal identification in order to identify the animal; and / or • b) prey identification in order to identify a prey of the animal; and / or • c) health identification in order to identify the health and / or well-being of the animal; and / or • d) unfamiliar-animal identification in order to identify a different animal.

Description

[0001] Smart pet door

[0002] The present invention relates to an intelligent animal door and an application program for such an animal door.

[0003] Pet doors, such as cat or dog doors, are generally known.

[0004] A cat flap (also called a cat door, or, in the case of slightly larger models, a dog door or dog flap) is a passageway for cats, dogs, or similar animals, allowing them to enter and leave premises such as houses, apartments, or indoor spaces independently of humans. These flaps are typically installed in doors, windows, or masonry.

[0005] Nowadays, there are pet doors that recognize your own pet using a chip, thus preventing access by other animals, such as cats or other animals, such as raccoons.

[0006] Animal doors typically consist of a simple flap that swings in one or two directions and is mounted in a frame. The frame is also referred to as the housing.

[0007] Many pet doors also feature a mechanism, a so-called 4-way lock, that regulates access through the pet door. A slider or lever allows you to adjust the direction in which the door opens. This allows you to set the door to be completely closed, open to both sides, or open to only one side, so that a pet can only enter or only exit.

[0008] Furthermore, controlled animal doors are increasingly being used. The animal in question wears, for example, a collar with an infrared transmitter, a magnetic key, or an implanted RFID transponder. A corresponding receiver in the animal door then only grants access to appropriately equipped or authorized animals. This prevents unauthorized animals from using the door. A controlled animal door is, for example, known from DE 10 2022 116 065 A1. This controlled animal door uses an RFID transponder and an additional jamming signal to prevent unauthorized animals from entering.

[0009] A disadvantage of previously known animal flaps is, for example, that the animals can enter the premises even with live prey and / or cannot enter the premises at all if they lose their collar and electronic identification.

[0010] The object of the present invention is therefore to provide an intelligent animal door that addresses the above-mentioned problems and is in particular designed to perform animal detection and / or prey detection and / or health detection.

[0011] According to the invention, a door for an animal, preferably a pet such as a cat, dog or the like, in particular an animal flap, preferably a cat flap, is proposed, comprising a housing, a flap with a locking mechanism, and electronics with a first sensor, preferably a camera, which is configured to perform at least one of the following functions, in particular during the day and / or at night: animal detection to recognize the animal and / or prey detection to recognize prey of the animal; and / or health detection to recognize the health of the animal; and / or foreign animal detection to identify another animal.

[0012] Such doors are also known as pet doors. The pet door described herein is preferably designed as a cat or dog door, i.e., for cats, dogs, or the like.

[0013] The animal flap essentially consists of a frame, a flap with a locking mechanism and electronics designed to operate the locking mechanism, in particular so that the flap is closed, open or open in one direction.

[0014] The animal flap is therefore designed as a complete system and preferably includes electronics that directly control the flap or its mechanism. The animal flap described herein does not use any jamming signals or the like, but opens or closes the flap directly. Preferably, the electronics comprise artificial intelligence or a neural network, or can access such via a data interface, particularly for animal detection, prey detection, health detection, and / or foreign animal detection.

[0015] Preferably, the electronics further comprise a control unit comprising: a timer for locking and / or unlocking the flap; and / or a motion sensor and / or infrared sensor to put the electronics into or out of standby mode, in particular to save energy; an optical and / or acoustic module to send warning signals, in particular to scare away foreign animals.

[0016] Particularly preferably includes mechanical and / or electronic switches to lock and / or unlock the flap, especially online and / or offline.

[0017] Particularly preferred, it includes mechanical and / or electronic switches to turn prey detection and / or foreign animal detection and / or health detection on and / or off, especially online and / or offline.

[0018] The housing is preferably designed as a multi-part frame and includes, for example, two covers and optionally a spacer that is arranged between the two covers.

[0019] Thanks to its multi-part design, the pet flap described herein can be retrofitted into any opening, be it in a wall, a front door, an apartment door, or the like. The pet flap is installed, for example, by placing the first cover against the opening from the outside and screwing the second cover to the first from the inside.

[0020] Preferably, the first cover is the one located on the outside, and the second cover is the one located on the inside, i.e., facing into the space. The covers are thus mounted against the opening from one side of the opening. The housing therefore has an outer (first) cover and an inner (second) cover, with the inner cover extending into the space and the outer cover being located outside the space.

[0021] Preferably, the covers and the optional spacer have corresponding mounting openings by means of which the elements can be assembled into the housing or frame. For example, the housing or frame can be assembled using screw or snap-fit ​​connections.

[0022] Preferably, the covers or spacers are made of weather- and UV-resistant PVC or other polymer.

[0023] In a particularly preferred embodiment, the covers or the spacer are also designed to be scratch-resistant.

[0024] The flap is designed, for example, as a double or single flap. Preferably, the flap(s) are movably, and in particular pivotably, mounted in the housing. This can be achieved, for example, by means of a pivot axis in the form of pins, which are arranged at the upper end of the flap and rotatably mounted in the housing. In one embodiment, for example, a first flap is arranged in the first cover and a second flap is arranged in the second cover. In another embodiment, exactly one flap is provided, for example, centrally in the area of ​​the spacer.

[0025] Preferably, the flap is made of weather- and UV-resistant PVC or another polymer and / or is transparent or translucent, in particular so that the animal can see through it. Preferably, the flap is scratch-resistant.

[0026] The flap's locking mechanism is specifically designed to hold the flap in a predetermined position, closing the opening to and / or from the premises. For this purpose, the locking mechanism can, for example, comprise a mechanical, electrical, or electromechanical mechanism, preferably controlled electronically. The electronics themselves can calculate corresponding control commands from sensor data, receive them via a push button attached to the housing, or via a corresponding app. Preferably, the push button is located on the second cover, i.e., the cover facing inwards. The locking mechanism can therefore be controlled automatically by the electronics and corresponding sensors, or manually by the user, for example, by pressing a button or using an app.

[0027] The locking mechanism can be engaged and disengaged electronically or manually by the user, for example, via a push button or an app. For instance, the user may have locked the flap using the push button, preventing any animal from entering the room. If the user's pet then approaches the flap from outside, the electronic system uses a camera to identify the pet and unlocks the flap, allowing the pet to enter. The flap is then relocked electronically.

[0028] Preferably, the locking mechanism is designed as an electromechanical locking mechanism, in particular in the form of a 4-way locking mechanism.

[0029] The electronics, for example, consist of a computing unit and a multitude of interfaces and include at least one sensor which is configured to perform one of the functions described a) to d) herein, in particular animal detection, prey detection, health detection and / or foreign animal detection.

[0030] Animal detection is preferably carried out using a sensor, such as a camera, and a database and / or artificial intelligence.

[0031] Animal detection, prey detection, health detection, and / or foreign animal detection are preferably performed using a model, in particular a self-optimizing model or a retrainable model, which, for example, runs on a Neural Processing Unit (NPU). The NPU is specifically designed to accelerate the neural network, especially its evaluations.

[0032] Health detection is particularly preferably carried out by means of a Neural Processing Unit (NPU), especially on the Neural Processing Unit (NPU), which is particularly connected to and / or part of the electronics.

[0033] The Neural Processing Unit (NPU) significantly accelerates image analysis, enabling faster animal, prey, and / or foreign animal detection. Health assessment is preferably performed not on the Neural Processing Unit (NPU), but on a backend, such as a server – i.e., outside the animal hatch.

[0034] Health detection serves primarily to assess the health and / or well-being of the animal, for example, based on facial and / or movement recognition. For instance, certain characteristics of the animal, such as the distance between the eyes, ear position, or gait, are analyzed to evaluate its health and / or well-being.

[0035] Preferably, the electronics have at least one connection to a push button with which the locking of the flap can be actuated, an interface to the sensor and optionally further interfaces if further sensors are present, as well as further interfaces for example for an app or a database described herein.

[0036] Preferably, the electronics comprise at least two sensors to perform the functions described in a) to d) herein.

[0037] Preferably, the electronics are arranged within the frame, in particular in such a way that the electronics are protected from dust and water.

[0038] The sensor(s) can be of any design, provided they are suitable for performing one of the functions a) to d). For example, the sensor could be a camera, such as a photo or video camera, or a TOF (time-of-flight) sensor, a PMD (photonic mixing device) sensor, or the like.

[0039] Preferably, the sensor(s) are configured as a camera, in particular a digital camera. The camera is set up to capture individual images and / or videos. Where images are mentioned herein, these are individual images and / or frames from videos.

[0040] Optionally, infrared lighting can also be provided to ensure suitable night photography. It is specifically recommended that the infrared lighting only be activated at dusk and / or in darkness.

[0041] Preferably, the electronics comprise two cameras. A first camera, preferably arranged above the flap, and a second camera, preferably arranged below the flap. Preferably, the first camera is arranged in the first cover and the second camera in the second cover, i.e., the cover that is arranged inside, with both cameras facing outwards to perform one of the functions described herein a) to d).

[0042] Preferably, the camera(s) are subjected to image processing after the fact, thereby improving or optimizing image quality. This image processing can be performed either by the electronics of the pet door (edge) or externally (remotely), for example via a cloud service.

[0043] In the event that the image quality is too poor even after processing, i.e., too blurry, overexposed, or the like, it is suggested to create a new image using the camera.

[0044] The first camera is preferably positioned centrally above the flap and in the first cover, i.e., the cover that faces outwards.

[0045] The first camera is directed outwards in particular to perform one of the functions a) to d) outside the premises.

[0046] By positioning the first camera above the flap, it is particularly well-suited to capturing the animal's face. The first camera therefore essentially records the animal from above.

[0047] Preferably, the first camera has a wide-angle lens and is designed to take pictures of an animal both during the day and at night, especially regardless of weather conditions and / or twilight.

[0048] The first camera is also ideally suited for taking infrared images of an animal at night or at dusk. This can be achieved, for example, by the camera not having an infrared filter. The first camera is therefore optimized for night vision and specifically designed for capturing infrared radiation. For this purpose, the camera may, for instance, have an optical filter that is transparent to infrared light, enabling the effective capture and processing of infrared radiation.

[0049] Alternatively or additionally, the camera can be equipped with a highly sensitive image sensor specifically optimized for capturing images in low light conditions and in the near-infrared range. The second camera is preferably positioned centrally below the flap and within the second cover, i.e., the cover facing inwards. The second camera is specifically directed outwards to perform one of the functions a) to d) outside the premises.

[0050] The positioning of the second camera, particularly below the flap, makes it especially well-suited to capturing the animal's mouth and its prey. Essentially, the second camera records the animal from below.

[0051] If the first and second cameras are positioned offset from each other, such an arrangement also allows for a perspective shot of the animal, which can be used, for example, to recognize the animal's facial features.

[0052] It is particularly preferred that the second camera be positioned lower in the room than the first camera, especially so that when the animal approaches the flap from the outside, the second camera is further away from the animal than the first camera.

[0053] Preferably, the second camera has a wide-angle lens and is designed to capture images of an animal during the day and / or at night. Specifically, this camera is also equipped to take infrared images of an animal. This can be achieved, for example, by the camera not having an infrared filter. The second camera is therefore optimized for night vision and specifically designed to detect infrared radiation. For this purpose, it can, for instance, have an optical filter that is transparent to infrared light, enabling the effective detection and processing of infrared radiation.

[0054] Alternatively or additionally, the camera can be equipped with a highly sensitive image sensor that is specifically optimized for capturing images in low light conditions and in the near-infrared range.

[0055] The use of cameras allows access based solely on the animal's appearance. This eliminates the need for the familiar RFID-chip collar, and the animal can still enter the premises even if it has lost its collar, for example, during a fight with another animal.

[0056] Preferably, the door, in particular the animal flap, also has other elements, such as a rubber seal for the housing, a brush seal for the flap, a power and / or battery connection for the electronics, a mounting adapter for the hole cover, cover glasses, preferably made of plastic, for infrared light from the cameras and / or speakers on PCB.

[0057] Alternatively or additionally, infrared lighting, for example in the form of infrared LEDs, is arranged on the housing, especially to increase the image quality for the first and / or second camera.

[0058] Alternatively or additionally, the animal flap or housing has a loudspeaker, buzzer or the like, especially to give acoustic feedback to the animal or the user.

[0059] Preferably, the electronics are powered by a battery and / or a rechargeable battery and / or a power connection, for example an alternating current and / or direct current connection.

[0060] Preferably, the electronics include, for example, a microcontroller or processor configured to execute an AI algorithm, a corresponding AI algorithm, a working memory configured to process images, non-volatile memory for an operating system and / or the storage of biometric data and / or for a neural network, and a power supply, for example in the form of a cable or a battery.

[0061] Alternatively or additionally, the electronics have a WLAN interface, a WLAN module or the like to configure the flap (open, closed) using an app or to update the software of the electronics.

[0062] Preferably, LEDs are also provided, particularly for emitting warning and / or indicator signals. In a particularly preferred embodiment, infrared LEDs are also provided, for example, on both sides of the first and / or second camera, to improve image quality for twilight and / or night photography. Preferably, the LEDs are located behind covers.

[0063] Preferably, a control panel is also provided, particularly on the second cover in an upper area. The control panel includes, for example, several push buttons with which the user can activate certain functions, such as locking the flap and / or animal detection and / or prey detection and / or another function of the animal flap or electronics. Preferably, the door also has a motion sensor configured to activate the first sensor.

[0064] Preferably, the motion sensor is also configured to activate other sensors, such as the first and second cameras described herein.

[0065] The use of the motion sensor allows, for example, the electronics to remain in a sleep (standby) mode until the motion sensor detects movement and then activates certain and / or all of the electronics' functions. Preferably, the electronics remain in a waiting mode until activated by the motion sensor, at which point they can perform animal detection, prey detection, health detection, and / or foreign animal detection.

[0066] The motion sensor therefore ensures, in particular, that the energy consumption of the electronics is kept as low as possible.

[0067] Preferably, the motion sensor is designed as a passive infrared motion sensor, in particular comprising one or more pyroelectric sensors, a Fresnel lens for focusing the infrared radiation and corresponding signal processing.

[0068] Preferably, the electronics are configured to lock and / or unlock the flap based on one of the functions described a) to d) herein and / or to output a signal or information about it, for example via an app to a user or to a cloud.

[0069] For example, the electronics recognize the animal and then unlock the flap so the animal can enter the premises. Alternatively, the electronics may recognize that the animal has prey in its mouth and keep the flap locked.

[0070] Preferably, animal detection is carried out using and / or a) artificial intelligence or b) a neural network.

[0071] Preferably, the housing includes a first cover that incorporates the first sensor for performing animal detection.

[0072] The first sensor is preferably a camera and is positioned centrally above the flap in the first cover, i.e., the cover facing outwards. The electronics are configured using this first sensor to perform animal recognition, for example, by creating a digital image of the animal and subsequently evaluating its corresponding biometric data. If the biometric data from the image of the animal in front of the flap matches the user's stored biometric data for their animal, the flap is unlocked.

[0073] Alternatively, animal detection can also be carried out by another or a second camera, which is preferably located centrally below the flap in one or the second cover.

[0074] Animal recognition can be performed by the electronics themselves, i.e., on device, or by other electronics, i.e., remotely.

[0075] Preferably, animal identification is performed on-site (edge) using biometric data; that is, the animal's biometric data is captured by the camera and compared with previously stored data. If the comparison is successful, the flap is unlocked. If the comparison is unsuccessful, the flap remains closed.

[0076] Preferably, the housing includes a second cover which has a second sensor to perform prey detection, for example by generating a digital recording of the animal and subsequent evaluation using object recognition, which is based, for example, on a neural network.

[0077] The second sensor is preferably designed as a camera and is located centrally below the flap in the second cover, i.e., the cover that faces inwards.

[0078] Prey detection is preferably carried out using electronics (edge), in particular to identify the animal's prey as quickly as possible, for example by means of artificial intelligence and / or a classification model and / or a detection model, such as object recognition, for example YOLO object recognition (You Only Look Once).

[0079] In one embodiment, prey detection serves only to identify whether the user's animal is carrying prey, i.e., simple prey detection. In another embodiment, prey classification is also performed – thus, it is determined not only "if" but also "what" the animal is carrying. Preferably, prey detection is performed by means of and / or using a) artificial intelligence or b) a neural network.

[0080] It is therefore specifically proposed that a camera be used to take a picture of the animal and that this picture be analyzed using artificial intelligence or a neural network. This can be done, for example, using so-called feature extraction. If the quality of the picture is not particularly good, it is further proposed to optimize the image quality using software or similar methods.

[0081] Preferably the housing comprises a first cover and a second cover, wherein the first cover has the first sensor and the second cover has a second sensor, in particular to perform health detection.

[0082] It is therefore specifically proposed here that health detection be carried out using at least two sensors. Preferably, the sensors are designed as cameras and enable perspective recording. Alternatively, health detection can also be carried out with only one camera that is configured to take perspective recordings.

[0083] Health assessment is preferably performed remotely, for example on a server or in the cloud. Furthermore, it is suggested that image series be used for health assessment; that is, multiple images from different times, for example over several days, are used to evaluate the animal's health status.

[0084] To determine an animal's state of health, various characteristics can be used, such as facial features, coat color, ear position, eye position, and eye cloudiness. In particular, facial features, ear position, and eye cloudiness can be used to detect certain levels of pain or suffering in the animal.

[0085] Preferably, health detection is performed using pain detection based on object tracking, feature extraction and / or artificial intelligence and / or deep learning and / or a database.

[0086] Health detection is therefore primarily based on pain detection, which is mainly AI-based. Alternatively and / or additionally, health detection relies on so-called "feline grimace scale" (FGS) indicators. Preferably, animal detection and / or prey detection and / or health detection and / or foreign animal detection is carried out by means of electronics and / or artificial intelligence or a neural network.

[0087] Animal recognition is preferably carried out using a sensor, such as a camera, and a database and / or artificial intelligence or neural network, in particular in such a way that the actual recognition of the animal is carried out by means of an image comparison, such as an embedding comparison, for example by means of a multidimensional embedding vector.

[0088] Animal recognition is preferably achieved using artificial intelligence and / or object recognition and / or a trained neural network to generate biometric embeddings. The neural network is designed to extract a high-dimensional feature vector from images of an animal, representing characteristic, individually distinguishable biometric features. In contrast to conventional approaches, this eliminates the need for individual model retraining for each cat to be registered, as the system uses a pre-trained embedding model that maps generalized cat features into a metric embedding space. Registration of an authorized animal is accomplished, for example, by capturing one or more reference images, from which reference embeddings are generated and stored in a database.During identification, a query embedding is generated from the current camera image and compared with stored reference embeddings, preferably using cosine similarity, Euclidean distance, or other suitable distance metrics in the embedding space. The decision regarding the animal's identity is based on statistical thresholding methods, particularly using standard deviation, confidence intervals, adaptive thresholds, or probability distributions. Optionally, identification can be based on the aggregation of multiple temporally sequential images, thereby increasing recognition accuracy through temporal consistency checks.

[0089] Preferably, the artificial intelligence includes biometric embedding and / or is based on a Convolutional Neural Network (CNN) with a Vision Transformer. The artificial intelligence used here is therefore more than just a classic Convolutional Neural Network (CNN).

[0090] The artificial intelligence preferably employs an ensemble architecture, where multiple models are applied to a single frame (multi-stage approach). For example, a detection model is used for object detection or localization, a convolutional neural network (CNN) for classifying the frame, and a vision transformer for animal detection.

[0091] Preferably, the foreign animal detection system is designed as a multi-stage ensemble architecture, which combines various specialized neural network models in a cascaded processing pipeline.The architecture comprises in particular: a) an object detection module, preferably based on a region-based Convolutional Neural Network (R-CNN), a YOLO or similar architecture, for localizing and segmenting the animal in the image area and for extracting relevant image sections (region of interest, ROI); b) a classification module, preferably implemented as a Convolutional Neural Network (CNN) with ResNet, EfficientNet or comparable architecture, for determining the animal species, posture and / or the presence of prey objects; c) an identification module for individual animal recognition, preferably realized by a Vision Transformer (ViT) or a hybrid CNN-Transformer architecture, which transforms the extracted ROIs into high-dimensional biometric embeddings.

[0092] The object detection module and / or the classification module and / or the identification module can be arranged in parallel or sequentially, with the outputs of the individual modules preferably aggregated by a fusion layer, which makes a final identification decision, particularly using weighted confidence values, attention mechanisms, or learned weighting parameters. This ensemble architecture achieves greater robustness against single-model misclassifications, especially under varying lighting conditions, partial occlusion, or different viewing angles.

[0093] Optionally, the architecture can be extended by additional preprocessing modules, for example to compensate for motion blur using deblurring networks.

[0094] The artificial intelligence is preferably trained using a dataset based on real images of animals and / or cats. The dataset preferably comprises more than 10,000 real images.

[0095] The artificial intelligence is trained using a hybrid training dataset that includes both real and synthetically generated images. The training dataset thus comprises both a real data component and a synthetic data component, where a) the real data component is formed from, or includes, a dataset containing more than 40,000 annotated real images of cats in various environmental conditions, lighting conditions, and perspectives; b) the synthetic data component comprises synthetic images generated, for example, using Generative Adversarial Networks (GANs) or Diffusion Models, particularly for simulating rare scenarios such as cats with unusual prey, extreme lighting conditions (backlighting, IR images), rare disease symptoms for health detection, morphological variations of different cat breeds, or the like.

[0096] This hybrid data strategy allows, in particular, a larger dataset size for training images while simultaneously covering edge cases that are underrepresented in real datasets.

[0097] The artificial intelligence is preferably trained using a dataset to recognize the face and / or body of the animal, particularly taking biometric data into account. The artificial intelligence is thus trained to extract biometric data and / or features from images.

[0098] The artificial intelligence is preferably trained using a custom loss function, i.e., a control system that allows certain errors, such as: focus on few false negatives (FN): prey is almost always recognized, but the cat is more often falsely excluded, or focus on few false positives (FP): the cat is almost never falsely excluded, but prey recognition is somewhat weaker.

[0099] The artificial intelligence is preferably based on an actively learning and supervised approach (active learning, supervised machine learning approach). Therefore, the artificial intelligence is not based on an automatic approach, such as a retraining model, as shown, for example, in Fig. 11.

[0100] The artificial intelligence is not based on a purely passive retraining model with exclusively manual user feedback collection.

[0101] The Active Learning Pipeline preferably includes the following process steps: a) Video selection, whereby relevant video sequences are selected, for example using three complementary methods, such as user-feedback-based selection, where users can mark videos for review via a user interface; automatic uncertainty detection, whereby the system independently identifies video sequences where the model prediction has increased uncertainty (increased epistemic and / or aleatory uncertainty); random sampling to ensure representative coverage of the input space.b) Frame extraction from selected videos, whereby relevant individual frames are extracted from the selected video sequences, for example using the following methods: a Monte Carlo dropout procedure to identify frames with increased model uncertainty (increased epistemic and / or aleatory uncertainty); an ensemble analysis, whereby several trained models are applied to the same frames and, in the event of a discrepancy in the predictions (disagreement) or identified misclassifications (false positives / false negatives), corresponding frames are extracted; manual verification by human reviewers for targeted frame selection. c) Embedding-based diversity selection, whereby the extracted frame candidates are transformed into a high-dimensional n-dimensional metric embedding space, in particular using a pre-trained embedding model.Based on semantic distances in the embedding space, redundant frames, especially those with high similarity within the same video sequence, are filtered out, thereby maximizing the diversity of the training dataset. d) AI-assisted annotation with human verification, whereby the selected, non-redundant frames are annotated, in particular using an AI-supported pre-classification procedure, and subsequently verified by human reviewers (human-in-the-loop), thus ensuring high annotation quality with reduced manual effort. e) Model retraining, in particular based on the verified annotations, is carried out by retraining the artificial intelligence, preferably using incremental learning methods that enable continuous learning without loss of previously learned skills.

[0102] According to the invention, a computer program product is further proposed, in particular for a door, as described herein.

[0103] The computer program product includes, for example, commands which, when executed on a computer, cause it to perform animal detection and / or health detection and / or prey detection and / or foreign animal detection for a door described herein.

[0104] Alternatively and / or additionally, the computer program product includes commands which, when executed on a computer, cause it to send and / or receive a push notification, send and / or receive an alarm, monitor (track) the health of an animal, provide a selection of images taken by a camera of the door described herein, collect statistics on the health of the animal, open and / or close the flap, and monitor or query the status (open / closed) of the flap.

[0105] According to the invention, a database, in particular a decentralized and / or AI-based database, is further proposed which is set up to provide data for a computer program product as described herein and / or to exchange data with an animal flap described herein.

[0106] In particular, the database is run as a cloud and / or set up to edit images, process images, process image sequences, execute AI algorithms, for example for health and / or pain detection, store and / or process images from cameras or from animal detection and / or prey detection and / or health detection and / or foreign animal detection.

[0107] Preferably, the data in the database, as well as communication with the database (e.g., via an app), is encrypted. Furthermore, the database is preferably configured to allow authorized third parties, such as verified veterinarians or verified veterinary medical systems, to access it, its data, and algorithms.

[0108] Preferably, the AI ​​algorithms on the database can be trained externally and / or updated.

[0109] An interface is a part of an electrical system that serves for communication. The term originates from the natural sciences and refers to the physical phase boundary between two states of a medium. It metaphorically describes the property of a system as a black box, of which only the "surface" is visible; communication is only possible via this surface. Two adjacent black boxes can only communicate with each other if their surfaces "match." The word also implies an "intermediate layer": for the two boxes involved, it is irrelevant how the other internally processes the messages and how the responses are generated. The description of the boundary is part of the boundary itself, and the black boxes only need to know the side facing them to ensure communication. This corresponds to the Latin etymology inter "between" and facies "appearance," "form," from the English word face.

[0110] When analyzing any given "system" as a whole, one will "dissect" this overall system into subsystems. The points of contact or connection between these subsystems (through which communication takes place) then represent the interfaces. Using these interfaces, the subsystems can be reassembled into a larger whole. They then serve as seams.

[0111] An application programming interface (API), also known as an application programming interface, is a part of a software system that allows other programs to connect to it. Unlike a binary interface (ABI), an API defines only the program connection at the source code level. Providing such an interface typically includes detailed documentation of the interface functions and their parameters, either in print or as an electronic document.

[0112] Besides accessing databases or hardware such as hard drives or graphics cards, a programming interface can also enable or simplify the creation of graphical user interface components. For example, the Windows Application Programming Interface (WAPI) of the Windows operating system allows external companies to develop software for this operating system.

[0113] Nowadays, many online services also provide programming interfaces; these are then called web services. In a broader sense, the interface of any library is referred to as a programming interface. This type of functional programming interface should be distinguished from the many other interfaces used in programming—for example, the parameters that are agreed upon and passed when calling subroutines.

[0114] Artificial intelligence (AI) is a subfield of computer science that deals with the automation of intelligent behavior and machine learning. The term is difficult to define, as there is already a lack of a precise definition of intelligence itself. One attempt at defining intelligence is that it is the property that enables a being to act appropriately and proactively in its environment; this includes the ability to perceive environmental data, i.e., to have sensory impressions and react to them, to receive, process, and store information as knowledge, to understand and generate language, to solve problems, and to achieve goals. Practical successes of AI are quickly integrated into application areas and then no longer fall under the umbrella of AI.

[0115] In this context, a neural network refers specifically to an artificial neural network. Artificial neural networks, also known as ANNs, are networks of artificial neurons inspired by the networks formed by biological neurons in the brain. An ANN is formed by interconnected artificial neurons, typically organized in layers. ANNs are used in machine learning. They enable computers to solve problems that are too complex to be described with rules, but for which there is a wealth of data that can serve as examples of the desired solution. ANNs form the basis for deep learning, which has allowed for significant advances in the analysis of large datasets since 2006. Successful applications of deep learning include image recognition and speech recognition.KNNs (K-nearest neighbors) are the subject of research in both machine learning, a subfield of artificial intelligence, and the interdisciplinary field of neuroinformatics. Replicating a biological neural network of neurons is more the domain of computational neuroscience. To improve the image quality of cameras, image processing can be performed after the camera has taken the picture. Image processing involves altering photographs, negatives, slides, or digital images. It is distinct from image processing, which involves manipulating the content of images. Image processing is often used to correct errors that can occur during photography or other image capture processes. These include, for example, over- and underexposure, blurriness, low contrast, image noise, red-eye, and converging lines.These errors often make images appear too dark, too bright, too blurry, or otherwise flawed. The causes can be technical problems or poor quality of the recording equipment (digital camera, lens, scanner), its incorrect operation, unfavorable working conditions, or inadequate source material. The two images on the right illustrate some image editing possibilities: The top image appears overexposed, has a color cast, the text is blurry, the object shows a light reflection at the top and is off-center. The lower, corrected, and now color-pure image, on the other hand, looks much clearer and sharper. This is because the subject is emphasized, as it is larger, corrected for distortion, and centered in the frame; the aspect ratio has been slightly adjusted.

[0116] Object recognition is a subfield of image processing, or computer vision, where the goal is to identify individual objects within images. An image is divided into regions that form meaningful units, which are then further analyzed for specific features to assign the image region to a class of objects. Preliminary object recognition usually follows this pattern: Divide an image into smaller, fixed-sized sections (windows) and then apply a classification algorithm to each window. While it is relatively easy for us humans to assign individual objects, such as a poodle or a German Shepherd, to an abstract category—in this case, the category "dog"—it is extremely difficult to teach a computer to do so. This task becomes increasingly challenging as more classes of objects need to be identified. Therefore, a specific algorithm is typically applied to the image, which, for example,can only recognize faces.

[0117] The animal flap or electronics described herein preferably operate without an RFID transponder or do not have an RFID transponder.

[0118] The animal flap and its electronics described herein preferably operate without interference (jamming) or similar sources, such as a (high-)frequency generator. The flap's mechanism is preferably controlled exclusively via image recognition, which reduces the number of electronic components required, making the animal flap less prone to failure and more cost-effective. The animal flap described herein is a complete product and not an extension or retrofit for existing animal flaps.

[0119] The present invention is explained in more detail below with reference to the accompanying figures.

[0120] Fig. 1 shows an animal flap in one embodiment in one view.

[0121] Fig. 2 shows a first cover of an animal flap in one view.

[0122] Fig. 3 shows an animal flap in a third view.

[0123] Fig. 4 shows an animal flap in one embodiment in a further view.

[0124] Fig. 5 shows an animal recognition system.

[0125] Fig. 6 shows prey detection.

[0126] Fig. 7 shows a health detection system.

[0127] Fig. 8 shows the process of animal detection.

[0128] Fig. 9 shows the process of prey detection.

[0129] Fig. 10 shows the process of a health screening.

[0130] Fig. 11 shows the training of an artificial intelligence.

[0131] Figure 1 shows an animal flap 1000 in one embodiment.

[0132] The animal flap 1000 includes a first cover 1100, an optional spacer 1200, a second cover 1300 and a flap 1400.

[0133] The first cover 1100, the spacer 1200 and the second cover 1300 together form the housing.

[0134] An opening is arranged in the center of the housing, which is concealed by the flap 1400. The flap 1400 is pivotally mounted in the housing, for example by means of two pins arranged at the upper end of the flap, in particular such that an animal, preferably a cat, can pass through the opening, provided the flap is unlocked.

[0135] Preferably, the second cover 1300 has a control panel 1350 with which the user can control the locking of the flap 1400.

[0136] Figure 2 shows a first cover 1100 of an animal flap 1000, in particular as shown in Figure 1.

[0137] The first cover 1100 is essentially designed like a square frame with rounded corners and has an opening in the middle through which an animal, especially a cat, can pass.

[0138] Within the first cover 1100, in particular in the middle on the sides, mounting openings 11 10 are provided, by means of which the first cover 1100 and the second cover 1300 can be screwed to the housing around an opening in a wall, door or the like.

[0139] Furthermore, the first cover 1100 has a first camera 1510 above its opening (in the installed state above the flap 1400), by means of which the functions described herein a) to d) can be carried out.

[0140] In addition to the camera 1510, there is also an aperture 1512 arranged to the right and left of it, behind which an LED is located, for example to give warning signals and / or emit infrared light.

[0141] Figure 3 shows a second cover 1300 of an animal flap 1000, in particular as shown in Figure 1.

[0142] The second cover 1300 is essentially designed like a square frame with rounded corners and has an opening in the middle through which an animal, especially a cat, can pass.

[0143] Within the second cover 1300, particularly in the center of the sides, mounting openings 1310 are provided, which are designed as complementary mounting openings to the mounting opening 1110, as shown in Figure 2, in order to screw the second cover 1300 to the first cover 1100 to form the housing. Furthermore, the second cover 1300 has a second camera 1520 below its opening (in the installed state below the flap 1400), by means of which the functions described a) to d) herein can be performed.

[0144] In addition to the camera 1520, there is also an aperture 1522 arranged to the right and left of it, behind which an LED is located, for example to give warning signals and / or emit infrared light.

[0145] The second cover 1300 also includes a spacer 1200, which is an integral part of the second cover. Alternatively, the spacer 1200 can also be an integral part of the first cover 1100 or a separate component.

[0146] Figure 4 shows an animal flap in one embodiment, in particular when the first cover 1100, as shown in Figure 2, and the second cover 1300, as shown in Figure 3, are joined to form the housing.

[0147] Both the first camera 1510, which is housed in the first cover 1100, and the second camera 1520, which is housed in the second cover 1300, have the same recording direction, namely outwards, i.e. in the direction of the first cover 1100, which is located outside the premises.

[0148] The first camera 1510 and the second camera 1520 are positioned offset from each other in the recording direction, specifically such that the upper camera 1510 is further forward in the recording direction than the lower camera 1520. This results in the second camera 1520 being further away from the animal than the first camera 1510. Consequently, the second camera 1520 has a longer recording time, which can be used to process the data from the second camera 1520 for more complex calculations and / or for remote processing, for example, on a server.

[0149] Fig. 5 shows animal recognition, in particular using the example of a cat, by means of object recognition.

[0150] Animal recognition is carried out, for example, using the first camera. For this purpose, the first camera 1510 generates an image of the animal, for example, a cat 2000 standing in front of the cat flap.

[0151] The cat is then identified using object recognition. If it is the user's cat, the lock opens. If it is a strange cat or another animal, the flap remains closed. Object recognition is a subfield of image processing or computer vision, where the goal is to identify individual objects in images. An image is divided into regions or pixels that form meaningful units, which are then further analyzed for features in order to assign the image region or pixels to a class of objects.

[0152] Animal recognition is carried out, for example, by the electronics and the first camera.

[0153] In this case, for example, the interpupillary distance of 2100, defined by pixels 14 and 17, is used to identify the animal. However, other and / or multiple characteristics can also be used to identify the animal.

[0154] Fig. 6 shows prey detection, in particular using the example of a cat 2000 by means of object recognition.

[0155] Although individual pixels 1,2 of the cat 2000 can be recognized, the image region 2200(B) in which, for example, the pixels of the mouth or the mouth characteristics would have to be identified, is obscured.

[0156] From this, it is deduced that the animal has something in its mouth, which is most likely prey 7000, and the flap remains locked accordingly, and a warning signal is issued to the animal via LEDs.

[0157] Fig. 7 shows a health detection system, in particular using the example of a cat, by means of object recognition.

[0158] Health detection is performed, for example, using the first and second cameras. The first camera, 1510, takes a picture of the animal, for example, a cat, 2000, standing in front of the cat flap. The second camera, 1520, also takes a picture simultaneously.

[0159] The images thus generated are then superimposed, for example using software, so that a perspective evaluation of the facial features can be carried out based on image points 1 , ... , 7, in particular based on facial features based on image points 1 , ... , 7.

[0160] It is therefore specifically proposed that, for health monitoring using object recognition, the facial features of the cat in 2000 be analyzed. The facial features thus determined can then, for example, be saved and / or compared with already saved facial features. Should the cat's facial features change over time, this can be used as an indication that the cat's health is changing, especially deteriorating.

[0161] Fig. 8 shows the process of an animal detection 3000.

[0162] For animal detection, for example, the first camera 1510 and / or the second camera 1520 of the animal flap 1000 is used, in particular as shown in Figures 1 to 4.

[0163] In the first step (3100), an image of the animal is taken using a camera, specifically the first camera. This can be done, for example, by activating the first camera and optionally an infrared light with a motion sensor, and then taking an image of the animal with the camera.

[0164] In a next step, this image is segmented 3200 times using object recognition, and then certain features are extracted from the image, as shown, for example, in Figures 5 or 6.

[0165] Subsequently, the extracted characteristics are classified in a further step 3300, for example using a biometric database in which the data of the user's animal is stored.

[0166] In the next step, a decision algorithm (3400) is used to determine whether the animal belongs to the user or is someone else's. This can be done, for example, by comparing the extracted characteristics with a database containing the relevant characteristics of the user's animal, such as coat color, eye spacing, or similar data. Preferably, this process is performed by the electronics of the pet door itself (edge).

[0167] If the photographed animal belongs to the user, the flap opens ("open"). If the photographed animal belongs to someone else, the flap remains closed ("closed").

[0168] In step 3400, a signal is generated specifically for directly opening the flap. Step 3400 does not generate a signal for jamming or similar actions. The flap or closing mechanism is therefore controlled directly based on image recognition. Additionally, the user can be informed about the process and / or the process can be saved, including the image captured by the camera.

[0169] Optionally, a warning signal, such as a visual and / or acoustic signal, can even be issued, especially if it is not the user's animal, i.e., a stranger's animal.

[0170] The animal recognition system described herein can also be used to locate missing animals, for example, by comparing the images captured by the animal recognition system with a database of missing animals. It is therefore also suggested that the door be used for the detection of foreign animals or missing animals, and that any found animals be reported accordingly.

[0171] Fig. 9 shows the process of prey detection 4000.

[0172] Prey detection 4000 is preferably performed downstream of animal detection 3000, as shown in Figure 7. This means, in particular, that prey detection is only carried out once the user's animal has been identified ("animal=yes").

[0173] In a first step 4100, an image of the animal is provided; for this purpose, for example, the image from animal recognition 3100 can be used and / or a new image can be taken using the first and / or second camera.

[0174] In the next step, this image is segmented using object recognition (4200), and then specific features are extracted from the image, as shown in Figures 5 and 6. Prey recognition focuses particularly on the segment of the image that includes the animal's mouth. Alternatively, classification can be performed instead of object recognition.

[0175] Subsequently, the extracted features are classified in a next step (4300) using a neural network for prey. This neural network was trained, for example, using extensive image data of prey animals and the "ground truth" technique. Preferably, this process is performed by the electronics of the animal flap itself (edge).

[0176] In the next step, step 4400 uses a decision algorithm to assess whether the animal is carrying prey or not. If the animal is not carrying prey, the flap is opened ("open"). If the animal is carrying prey, the flap remains closed ("closed").

[0177] Optionally, a warning signal can be issued to the animal if it is carrying prey, for example a visual and / or acoustic signal.

[0178] Additionally, the user can be informed about the process and / or the process can be saved, in particular this also includes the image captured by the camera.

[0179] In another embodiment, the prey detection is positioned upstream of the animal detection 3000, as shown in Figure 7.

[0180] Fig. 10 shows the process of a health detection 5000.

[0181] The health detection 5000 is preferably performed independently of the animal and / or prey detection. However, the health detection is only carried out for the user's animal.

[0182] The health detection 5000 is preferably performed after the animal detection 3000. This means, in particular, that prey detection is only carried out once the user's animal has been identified ("animal=yes").

[0183] Furthermore, the health detection 5000 is preferably based on perspective images using two cameras, as shown in particular in Figure 7.

[0184] In a first step 5100, an image of the animal is provided; for this purpose, for example, the image from animal recognition 3100 can be used and / or a new image can be taken using the first and second cameras, especially in the form of a perspective view.

[0185] In a next step, this image is segmented 5200 times using object recognition, and then certain features are extracted from the image, as shown in Figure 6.

[0186] Health assessment relies primarily on the so-called "feline grimace scale" (FGS) indicator. This means that specific features are extracted, which can then be used to run an FGS algorithm, such as ear position, eye area, jaw tension, whisker changes, or similar characteristics. Alternatively or additionally, another algorithm can be used, such as a landmark-based approach that focuses on specific landmarks.

[0187] Subsequently, the extracted features are classified in a next step using a neural network for prey based on a deep learning model. This neural network was trained, for example, using extensive image data of animals and the "ground truth" technique. Features crucial to the FGS algorithm, such as ear position, eye area, mouth tension, whisker changes, and the like, were labeled accordingly. Statistical calculations can also be performed, for example, to measure ear angles.

[0188] In the next step 5400, the process is then repeated and / or compared with already generated data, in particular so that the health detection is based on image series of the animal and a final assessment, a health assessment, can be carried out.

[0189] The assessment can be based, for example, on a health factor and / or stored data and / or an analysis of changes over time. For example, it may be observed that the FGS indicators deteriorate over time, indicating poor health.

[0190] Other visual health indicators can also be used for assessment, such as coat condition, pupil dilation, tongue position, tail posture, or the like.

[0191] These image series and / or health data may have been generated and / or stored over several days, weeks, or months. Step 5400 is preferably performed using an external database 6000 or cloud storage (remotely).

[0192] If the health monitoring system detects a deterioration in the animal's health, a corresponding alarm is issued, for example as a push notification to a corresponding app.

[0193] The health monitoring system proposed here offers continuous, and notably non-invasive, monitoring of the animal. The animal's health can therefore be assessed without stressing it, as would be the case, for example, during a visit to the veterinarian. Furthermore, the data obtained through this health monitoring system can be used to assist the veterinarian in assessing the animal and / or supporting treatment decisions.

[0194] Furthermore, health detection offers a time- and cost-effective alternative for ongoing health assessment of the animal.

[0195] Furthermore, diseases in animals can be detected at an early stage using the neural network or AI.

[0196] Fig. 11 shows the training of an artificial intelligence, specifically in the form of an active learning pipeline. The animal flap therefore undergoes active and supervised training, primarily based on images or videos of numerous animal flaps.

[0197] All cat flaps continuously send videos to the system, which is stored on a server, for example. Depending on the number of cat flaps, this can amount to several thousand videos or more.

[0198] In a first step (STAGE 1: Intelligent Video Selection), the system selects from the large number of videos those that offer the best learning content or are most valuable to the system, primarily in three different ways: user feedback, where users can report videos in an app if the AI ​​has made an incorrect decision (e.g., "That wasn't prey" or "That was my cat, not a stranger"); auto-detection, where the AI ​​independently identifies videos it is unsure about, especially without notifying the user; and random sampling, where random samples are taken to ensure that "normal" cases remain in the training and the model is not only trained on problem cases.

[0199] Videos are preferably selected from all three methods, or a large number of selected videos are combined. Optionally, the large number of selected videos can be further filtered according to additional criteria, particularly to narrow down the selection. In a next step (STAGE 2: Video Analysis), the most informative individual frames (frame candidates) are extracted from the selected videos, for example, through: model uncertainty analysis, where, for example, a Monte Carlo dropout analysis identifies specific frames where the model is uncertain; and / or ensemble analysis, where, for example, multiple models analyze the same frames; and / or human verification, where, for example, humans selectively mark or label frames in particularly difficult cases.

[0200] The process described herein therefore only evaluates specific images or frames (frame candidates) and not entire videos, whereby the images are filtered out accordingly through the steps described above (STAGE 1 and 2).

[0201] In the next step (STAGE 3: Embedding-based diversity selection), a special embedding model checks how similar the frame candidates are, since some of them come from the same video.

[0202] Very similar frames (e.g., 5 almost identical pictures of the same cat in the same situation (perhaps from the same video)) are filtered out, leaving only diverse, information-rich frames.

[0203] The various, information-rich frames are then labeled using AI-assisted annotation. The AI ​​suggests pre-labels, and humans only need to verify them instead of labeling everything from scratch.

[0204] Finally, a model retraining is performed, i.e., the model is retrained with the verified, diverse data, using incremental learning, so the model does not forget the old data, but learns from it.

[0205] The steps described above are then carried out in a continuous improvement loop; that is, the improved model of a cat flap goes back into production and the cycle begins again – but now with a better model that makes fewer errors. (Reference list)

[0206] 1, 2, ... pixels

[0207] 1000 pet doors, especially cat doors

[0208] 1100 first cover, especially for the outside

[0209] 1110 Mounting opening

[0210] 1200 spacers

[0211] 1300 second cover, especially for the inside

[0212] 1310 Mounting opening

[0213] 1350 Control panel

[0214] 1400 flap

[0215] 1500 Electronics

[0216] 1510 first camera

[0217] 1512 Aperture including LED

[0218] 1520 second camera

[0219] 1522 Cover including LED

[0220] 1530 network connection

[0221] 2000 animals, especially cats

[0222] 2100 characteristics, especially of a cat

[0223] 2200 facial features, especially of a cat

[0224] 2200(B) Area of ​​facial features

[0225] 3000 animal recognition

[0226] 4000 prey detection

[0227] 5000 health detection

[0228] 6000 Cloud

[0229] 7000 loot

Claims

Claims 1. Door (1000) for an animal, preferably a pet, such as cats, dogs or the like, in particular an animal flap, preferably a cat flap, comprising: a housing (1100, 1200, 1300), a flap (1400) with a locking mechanism, and electronics (1500) with a first sensor, preferably a camera, which is configured to perform at least one of the following functions, in particular during the day and / or night: a) animal detection to recognize the animal; and / or b) prey detection to recognize prey of the animal; and / or c) health detection to recognize the health and / or well-being of the animal; and / or d) foreign animal detection to identify another animal.

2. Door (1000) according to claim 1, wherein it has a motion sensor configured to activate the first sensor.

3. Door (1000) according to claim 1 or 2, further comprising: the electronics being configured to lock and / or unlock the flap based on one of the functions a) to d) and / or to output a signal and / or information.

4. Door (1000) according to one of the preceding claims, wherein the housing comprises a first cover (1300) having the first sensor to perform animal detection.

5. Door (1000) according to one of the preceding claims, wherein the housing comprises a second cover (1300) which has a second sensor to perform prey detection.

6. Door (1000) according to claim 5, wherein prey detection is carried out by means of: a) an artificial intelligence; and / or b) a neural network.

7. Door (1000) according to any one of the preceding claims, wherein the housing comprises a first cover and a second cover, the first cover having the first sensor and the second cover having a second sensor for performing health detection.

8. Door (1000) according to claim 7, wherein the health detection is performed by means of pain detection based on object tracking and / or artificial intelligence and / or deep learning and / or a database.

9. Computer program product for a door, comprising commands which, when executed on a computer, cause the computer to perform animal detection and / or health detection and / or prey detection and / or foreign animal detection for a door according to any one of claims 1 to 8.

10. Database, in particular a decentralized and / or AI-based database, configured to provide data for a computer program product according to claim 9.

Images