Method and device for recognizing an object, as well as machine learning model and computer program product

Abstract

The present disclosure relates to a method for recognizing an object. The method comprises the following steps: i) Comparing a detected object from a second frame with at least one detected object from a first frame using at least one of the following parameters: a) distance in terms of distance, in particular the Euclidean distance between the objects; b) Difference between the objects in their geometric dimensions, in particular the height and / or width of the objects; c) Difference between the objects in their preferably optical and / or acoustic appearance; d) Equality of the objects with respect to an object type; ii) Evaluating the results from step i) to determine whether the detected object of the second frame is the detected object of the first frame. The present disclosure further relates to a device (1) for recognizing an object, as well as a machine learning model and a computer program product.

Description

[0001] The present disclosure relates to a method and a device for recognizing an object and in particular a machine learning model and in particular a computer program product.

[0002] Object recognition, also known in technical circles as object tracking or at least as a subset thereof, is becoming increasingly relevant in a wide variety of computer-based applications. For example, object recognition is crucial for many machine vision applications. From analyzing pedestrian and traffic flows to analyzing movement patterns in sports, object recognition plays an indispensable role in gaining valuable insights.

[0003] In machine vision applications, object recognition is typically a subsequent step following object recognition. Object recognition involves identifying an object, such as a person, car, or other item. This is usually based on individual frames from a video sequence, for example, from a surveillance camera, in which a single object is detected. Object tracking then determines whether each subsequent frame represents the same initially identified object.

[0004] Analytical methods for object tracking use previously known movement patterns to predict the likely path of an object. These methods usually require a relatively large amount of computing time and often deliver an insufficient success rate in assessing whether a currently observed object corresponds to a previously recognized object.

[0005] Against this background, there is also a need to improve object recognition. This is based on the expectation that an improved analytical method will allow object recognition to be carried out more quickly and / or with a higher success rate.

[0006] Fundamental improvements in object recognition are offered by a method, in particular an analytical method, which includes or consists of the following steps: i) Comparing a detected object of a second frame with at least one detected object of a first frame, preferably created earlier in time, using at least one of the following parameters: a) distance in terms of distance, in particular the Euclidean distance between the objects; b) Difference between the objects in their geometric dimensions, in particular the height and / or width of the objects; c) Difference between the objects in their preferably optical and / or acoustic appearance; d) Equality of the objects with respect to an object type; ii) Evaluating the results from step i) to determine whether the detected object of the second frame is the detected object of the first frame.

[0007] The improved method is based on the idea of ​​using a procedure that a human would use to assess whether an object is one that has already been recognized. Such a procedure relies, for example, on parameters of the type described above. It has been shown that object recognition is facilitated when one of these parameters is already used. In particular, this results in advantages in the robustness of the method. In particular, it also results in advantages in the execution speed of the method.

[0008] The improved method can be designed such that step i) is executed using at least two of the parameters a) to d). It has been shown that this further improves object recognition. In particular, it offers further advantages in terms of the method's robustness and execution speed. These advantages are especially noticeable, for example, when step i) is executed using parameters a) and b).

[0009] The improved method can further be designed such that step i) is executed using at least three of the parameters a) to d). It has been shown that this further improves object recognition. In particular, it results in further advantages in the robustness of the method. It also results in further advantages in the execution speed of the method. These further advantages are particularly noticeable, for example, when step i) is executed using parameters a), b), and c).

[0010] The improved method can further be designed such that step i) is executed using all four parameters a) to d). It has been shown that this further improves object recognition. In particular, it results in further advantages in the robustness of the method. It also results in further advantages in the execution speed of the method.

[0011] The improved method can further be designed such that step i) is executed using at least one additional parameter. Such an additional parameter can be the degree of occlusion of at least one of the objects by a third object. Additionally or alternatively, such an additional parameter can be the degree of occlusion of the object in the second frame by a third object, preferably in the first frame. It has been shown that this further improves object recognition. In particular, this results in further advantages in the robustness of the method. It also results in further advantages in the execution speed of the method.

[0012] The improved procedure can further be designed such that the results from at least one comparison in step i) are used as metadata for step ii). This ensures that the comparison results are presented in a structured manner, thus facilitating data evaluation in step ii).

[0013] Furthermore, the improved method can be designed to utilize a preferably trained AI model to determine whether the object detected in the second frame is the same object detected in the first frame. In this disclosure, the term "AI" is an abbreviation for "artificial intelligence".

[0014] The AI ​​model enables the recognition device to draw conclusions about the identity of the detected objects based on at least one parameter. Furthermore, the AI ​​model allows for the recognition of specific patterns from a collection of datasets; in other words, the AI ​​model learns from the datasets it is trained on. This trainability improves object recognition performance. In particular, it facilitates faster object recognition with a higher success rate.

[0015] For example, the AI ​​model is a machine learning model or incorporates a machine learning model. This allows large amounts of data to be tracked and, based on this, trends and patterns to be identified that humans would not recognize. Therefore, faster and better decisions can be made when recognizing objects.

[0016] For example, the improved method uses a random forest algorithm to determine whether the object detected in the second frame is the same object detected in the first frame. This algorithm is based on an automated compilation of decision trees. Each of these decision trees goes through a structured decision hierarchy for each question to be answered. The combination of several differently structured decision trees forms a so-called random forest. Each tree thus contributes to the model's final decision. This achieves a combination of effectiveness and efficiency, so that answers regarding object recognition can be delivered in real time with the minimum possible use of computing power.

[0017] Additionally or alternatively, the improved method can use an XGBoost algorithm to determine whether the object detected in the second frame is the same object detected in the first frame. This algorithm is based on a sequence of stacked gradient-boosting functions. These functions allow for greater complexity, offer more parameters for influencing the model's behavior, and thus provide performance advantages in object recognition.

[0018] Furthermore, the improved procedure may include one or more validation mechanisms to ensure accuracy and efficiency. In particular, error handling and validation steps are integrated into the improved procedure, for example, into the AI ​​model, to detect and address unforeseen situations.

[0019] The individual images used as the basis for the improved method, in particular the first and second images, can be captured by a single camera and / or video camera. The individual images can be part of an image sequence, especially a video sequence. The camera and / or video camera can be stationary. In this case, the individual images are, for example, captured from the same perspective and / or depict the same area.

[0020] The individual images may also be captured by different cameras or video cameras, for example, covering different areas, and / or the camera or video camera may have changed position between capturing the individual images. Therefore, the data for the individual images also contains information about the respective position of the camera or video camera and / or about any changes in position between capturing the individual images, which are or can be taken into account for recognition.

[0021] Further improvements can be achieved through an AI model, particularly a machine learning model, trained using the improved method described above. This model can include a random forest algorithm, specifically the random forest algorithm described above, to determine whether a detected object in a second frame is a detected object in a first frame.

[0022] Additionally or alternatively, the model can include an XGBoost algorithm to determine whether the detected object of the second frame is the detected object of the first frame, or to determine whether a detected object of a second frame is a detected object of a first frame.

[0023] Further improvements are fundamentally achieved through a computer program product with program code stored on a computer-readable medium for carrying out the improved procedure described above, and in particular encompassing the model described above.

[0024] Further improvements can be achieved in principle by using a method for detecting and re-detecting an object. This method comprises, or consists of, the steps of providing a first image, detecting at least one object in the first image, providing a second image, detecting an object in the second image, and carrying out the improved method described above.

[0025] For example, the second image was taken at a later time than the first. Information about the detected objects, particularly the object detected in the first image and the object detected in the second image, is provided in the data. For example, the data contains information about the position of the detected object in the corresponding image. For example, the data contains information about the dimensions, particularly the height and / or width of the detected object. For example, the data contains information about the class type to which the detected object was assigned. This information may have been determined using a so-called bounding box.

[0026] To achieve the aforementioned improvements, a device for recognizing an object is proposed, which is hereinafter also referred to as the recognition device. The proposed recognition device is particularly suitable for carrying out the method described above. The proposed recognition device is also particularly suitable for implementing the computer program product described above.

[0027] The proposed recognition device comprises a comparison unit and an evaluation unit. The comparison unit is specifically configured to compare a recognized object from a second frame with at least one recognized object from a first frame, preferably one acquired earlier, based on at least one of the following parameters: a) distance in terms of distance, in particular the Euclidean distance between the objects; b) Difference between the objects in their geometric dimensions, in particular the height and / or width of the objects; c) Difference between the objects in their preferably optical and / or acoustic appearance; d) Equality of the objects with respect to an object type.

[0028] The evaluation device is specifically designed to evaluate the results of at least one comparison of the comparison device in order to determine whether the object of the second single image is the object of the first single image.

[0029] In the present disclosure, the term "device for recognizing an object" refers in particular to an electronic unit of computer hardware that controls certain processes and / or sequences. The device may include a digital processing unit, which may, for example, comprise a microprocessor unit (CPU). The CPU may be connected to a storage system and / or bus system via data and / or signals. The device may include one or more programs or program modules. The digital processing unit may be configured such that instructions implemented as a program stored in a storage system are executed, input signals are received from a data bus system, and / or output signals are sent to a data bus system. A storage system may include one or more, in particular different, storage media.The storage media can be, in particular, optical, magnetic, solid-state storage media and / or other, preferably non-volatile, storage media.

[0030] The proposed recognition device can be designed such that the comparison unit and the evaluation unit can communicate via preferably defined interfaces, in particular transmitting and / or processing data. The interfaces enable efficient data transmission and processing.

[0031] The proposed recognition device can be configured such that the comparison unit is set up to compare the recognized object of the second frame with the recognized object of the first frame based on at least two of the parameters a) to d). In particular, the comparison unit is set up to compare the recognized object of the second frame with the recognized object of the first frame based on parameters a) and b).

[0032] The proposed recognition device can further be configured such that the comparison unit is set up to compare the recognized object of the second frame with the recognized object of the first frame based on at least three of the parameters a) to d). In particular, the comparison unit is set up to compare the recognized object of the second frame with the recognized object of the first frame based on parameters a), b), and c).

[0033] The proposed recognition device can further be designed such that the comparison device is configured to compare the recognized object of the second single image with the recognized object of the first single image based on all four parameters a) to d).

[0034] The proposed recognition device can further be configured such that the comparison unit is set up to compare the recognized object of the second image with the recognized object of the first image based on at least one of the parameters a) to d) and at least one further parameter. The at least one further parameter can be the degree of occlusion of at least one of the objects by a third object. Additionally or alternatively, the degree of occlusion of the object of the second image by a third object of the first image can be used as a further parameter.

[0035] In particular, the recognition device is configured to provide and / or store the results from the at least one comparison of the comparison device in metadata. Furthermore, the recognition device is configured to provide the metadata to the evaluation device in order to perform the evaluation based on it.

[0036] Furthermore, the proposed recognition device can include an AI model. The AI ​​model is specifically designed to evaluate the results of the comparison device, for example, to determine whether the object detected in the second image is the same object detected in the first image.

[0037] In particular, the proposed recognition device includes a random forest algorithm, which is specifically designed to evaluate the results of the comparison device, for example, to determine whether the object detected in the second frame is the same as the object detected in the first frame. Additionally or alternatively, the proposed recognition device may include an XGBoost algorithm, which is specifically designed to evaluate the results of the comparison device, for example, to determine whether the object detected in the second frame is the same as the object detected in the first frame.

[0038] Furthermore, a device for recognizing an object may be provided. The device, hereinafter also referred to as the recognition device, is specifically designed to recognize an object in a single image. In particular, the recognition device is designed to recognize the aforementioned object in the first single image and the aforementioned object in the second single image.

[0039] In particular, the recognition device is configured to provide information about the at least one recognized object in the form of data. For example, the data is prepared in such a way that it can be processed by the recognition device.

[0040] Further details and features will become apparent from the following description of at least one exemplary embodiment with reference to the drawing. The drawing shows... Fig. 1. A first exemplary embodiment of a device for recognizing an object in a schematic representation, Fig. 2 a second exemplary embodiment of a device for recognizing an object in a schematic representation, Fig. 3 a third exemplary embodiment of a device for recognizing an object in a schematic representation, Fig. 4 a fourth exemplary embodiment of a device for recognizing an object in a schematic representation, Fig. 5 an exemplary tree structure of a trained model which, in at least one of the embodiments of the device for recognizing an object according to the Fig. 1 to 4 can be used, Fig. 6 an enlarged section of the exemplary tree structure of the Fig. 5, Fig. 7 an exemplary environment in which a device for recognizing an object is, for example, of one of the embodiments of the Fig. 1 to 4 can be used, Fig. 8 a first exemplary single image from a stationary video camera, and Fig. 9 a second exemplary single image from the stationary video camera, taken at a later time than the first exemplary single image.

[0041] Fig. Figure 1 shows a first exemplary embodiment of a device 1 for recognizing an object, illustrated by a simplified block diagram. The exemplary device 1, hereinafter also referred to as the recognition device, is suitable for use in machine vision applications. For example, the exemplary device 1 can be used in the analysis of pedestrian and traffic flows. The exemplary device 1 can also be used to analyze movement patterns, for example, during sporting activities.

[0042] The exemplary device 1 preferably uses a digital processing unit with a microprocessor unit (CPU) that is data- and / or signal-connected to a storage system and / or bus system. The storage can comprise optical, magnetic, solid-state storage media or other non-volatile storage media.

[0043] Preferably, the exemplary device 1 is configured to process provided data containing information about at least one detected object O1 from a first single image E1 and one detected object O2 from a second single image E2. For example, the first single image E1 was created earlier in time than the second single image E2. For example, the data also contains information about a detected object class, i.e., whether it is, for example, a "vehicle", a "person", or the like.

[0044] The individual images E1 and E2 may have been captured by a single camera and / or video camera. They may also be part of an image sequence, particularly a video sequence. The camera and / or video camera may be stationary. In this case, for example, the individual images E1 and E2 may have been captured from the same perspective and / or be directed at the same area.

[0045] The individual images E1 and E2 may have been captured by different cameras or video cameras, for example, pointed at different areas, and / or the camera or video camera may have changed position between the capture of individual images E1 and E2. Therefore, the data for individual images E1 and E2 also contain information about the respective position of the camera or video camera and / or about any changes in position between the capture of individual images E1 and E2.

[0046] The exemplary device 1 comprises, for example, a comparison device 2 and an evaluation device 3. Preferably, the provided data for the individual images E1, E2 can be processed by the comparison device 2 and / or the evaluation device 3.

[0047] The comparison device 2 is in particular configured to compare the detected object O2 of the second single image E2 with the detected object O1 of the first single image E1, preferably on the basis of the parameters, also to be referred to as metrics, a) distance in terms of distance, in particular Euclidean distance between objects O1, O2, b) Difference between objects O1, O2 in their geometric dimensions, in particular the height and / or width of objects O1, O2, c) Difference between objects O1, O2 in their preferably optical and / or acoustic appearance, and d) Equality of objects O1, O2 with respect to an object type.

[0048] In this disclosure, the term "object type" is to be understood as synonymous with the term "object class." Both "object type" and "object class" in this disclosure refer to a classification of objects, for example, with regard to certain properties, in order to organize them. For instance, objects belonging to the same object class or object type are those that, based on their properties, fit into that class or type. A classification can be made, for example, according to "person," "car," "truck," "motorcycle," and / or the like. In this respect, "person," "car," "truck," and "motorcycle" then constitute an object class or object type.

[0049] Parameter a) is based on the idea that when observing, for example, cars, it's clear to a person that a car doesn't teleport abruptly. Objects of a class, such as objects of the class "Car," will have typical distances between their appearance in the current frame and their appearance in the previous frame. Cars will have significantly different distances than, for example, a person or a bicycle, which move comparatively slower than cars. Therefore, the distance between a car in the current frame and its counterpart in the previous frame is significantly different than the distance between the two objects.

[0050] Parameter b) is based on the idea that objects in the background have different dimensions than objects in the foreground. When looking at cars, for example, it's clear to a person that a car in their field of vision doesn't change its dimensions arbitrarily and abruptly. If an object in the current frame is the same object as in the previous frame, its dimensions are likely to have changed less significantly or differently compared to all other objects.

[0051] Parameter c) is based on the idea that when a person looks at a car, it's clear that the same car won't suddenly change fundamentally, for example, in its color. An object will show a higher similarity to itself in previous images than to other objects.

[0052] Parameter d) is based on the idea that when observing, for example, road traffic, a person clearly distinguishes between a pedestrian and a car. They also understand that a car is a passenger vehicle or a truck. Humans thus assign objects to individual classes, such as "pedestrian," "car," "passenger vehicle," and "truck." If an object in a current frame is the same as the object in the previous frame, then the current object should belong to the same class as the object in the previous frame.

[0053] Preferably, the comparison device 2 is configured to perform the comparison based on a combination of parameters a) to d). Preferably, the comparison device 2 is configured to perform the comparison based on all of parameters a) to d). By comparing based on parameters a) to d), the exemplary device 1 relies on a procedure that a human intuitively uses to assess whether an object is one that has already been recognized. In this respect, the exemplary device 1 reflects the human intelligent behavior involved in recognizing an object.

[0054] A possible comparison by the comparison facility 2 can, for example, take place by combining the following aspects based on the above parameters a) to d): • Is the distance between the current object, i.e., object O2 of the second frame E2, and a known object, i.e., object O1 of the first frame E1, small enough? • Is the current object O2 similar enough to a known object O1 in terms of its geometric dimensions, such as its height and / or width? • Is the current object O2 visually similar enough to a known object O1? • Does the current object O2 belong to the same object class as the known object O1?

[0055] Preferably, the evaluation unit 3 is configured to evaluate the results of at least one comparison by the comparison unit 2 in order to determine whether object O2 of the second frame E2 is object O1 of the first frame E1. The result of the evaluation is then available, for example, as output information A. The evaluation unit 3 and the comparison unit 2 can be integrated into a single functional unit. Alternatively, the comparison unit 2 and the evaluation unit 3 can be separate functional units.

[0056] Preferably, the exemplary device 1 is configured to provide and / or store the results from the at least one comparison of the comparison device 2 in metadata. Preferably, the exemplary device 1 is further configured to provide the metadata to the evaluation device 3 in order to perform the evaluation based on it.

[0057] Preferably, the exemplary device 1 uses and / or includes a preferably trained AI model to determine, based on parameters a) to d), whether the recognized object O2 of the second frame E2 is the recognized object O1 of the first frame E1. This implements a further measure to orient itself towards human-like intelligent behavior when recognizing an object.

[0058] The AI ​​model can be or include a machine learning model. For example, the model might include a random forest algorithm and / or an XGBoost algorithm to determine whether the detected object O2 in the second frame E2 is the same as the detected object O1 in the first frame E1. In other words, the AI ​​model provides an assessment of object equality based on the object comparison performed using the metrics or parameters.

[0059] For example, the Random Forest algorithm uses or consists of an automated compilation of decision trees. Each tree traverses a structured decision hierarchy for each question to be answered. Similarly, the XGBoost algorithm uses or is based on a sequence of stacked gradient boosting functions. These functions offer greater complexity and parameters for customizing the model's behavior.

[0060] Fig. Figure 2 shows a second exemplary embodiment of a device 1.1 for recognizing an object, based on a simplified block diagram. The exemplary device 1.1 differs from the exemplary device 1 of the Fig. 1. This is achieved by using an additional parameter e) alongside parameters a) to d). This parameter e) relates to the degree of occlusion of at least one of the objects O1, O2 by a third object.

[0061] The exemplary device 1.1 has a comparison unit 2.1, which preferably has the same function as the comparison unit 2 of the exemplary device 1 and is additionally configured to compare the detected object O2 of the second single image E1 with the detected object O1 of the first single image E1 based on parameters a) to d) and e). The exemplary device 1.1 has an evaluation unit 3.1, which preferably has the same function as the evaluation unit 3 of the exemplary device 1 and is additionally configured to further evaluate the results of the comparison unit 2.1, taking parameter e) into account, in order to make an assessment of whether objects O1 and O2 are possibly identical.

[0062] A possible comparison by the comparison device 2.1 can, for example, take place by combining the following aspects based on the above parameters a) to e): • Is the distance between the current object, i.e., object O2 of the second frame E2, and a known object, i.e., object O1 of the first frame E1, small enough? • Is the current object O2 similar enough to a known object O1 in terms of its geometric dimensions, such as its height and / or width? • Is the current object O2 visually similar enough to a known object O1? • Does the current object O2 belong to the same object class as the known object O1? • Is the current object in the current frame, i.e., object O2 of the second frame E2, partially or completely obscured by a third object?

[0063] Fig. Figure 3 shows a third exemplary embodiment of a device 1.2 for recognizing an object, illustrated by a simplified block diagram. The exemplary device 1.2 differs from the exemplary device 1 of the Fig. 1. This is achieved by using an additional parameter e') alongside parameters a) to d). This parameter e') relates to the degree of occlusion of object O2 of the second frame E2 by a third object of the first frame E1.

[0064] The exemplary device 1.2 has a comparison unit 2.2, which preferably has the same function as the comparison unit 2 of the exemplary device 1 and is additionally configured to compare the detected object O2 of the second single image E1 with the detected object O1 of the first single image E1 based on parameters a) to d) and e'). The exemplary device 1.2 has an evaluation unit 3.2, which preferably has the same function as the evaluation unit 3 of the exemplary device 1 and is additionally configured to further evaluate the results of the comparison unit 2.2, taking parameter e') into account, in order to make an assessment of whether objects O1 and O2 are possibly identical.

[0065] A possible comparison by the comparison device 2.2 can, for example, take place by combining the following aspects based on the above parameters a) to e'): • Is the distance between the current object, i.e., object O2 of the second frame E2, and a known object, i.e., object O1 of the first frame E1, small enough? • Is the current object O2 similar enough to a known object O1 in terms of its geometric dimensions, such as its height and / or width? • Is the current object O2 visually similar enough to a known object O1? • Does the current object O2 belong to the same object class as the known object O1? • Is the current object in the current frame, i.e., object O2 of the second frame E2, partially or completely obscured by a third object in the first frame E1?

[0066] Fig. Figure 4 shows a fourth exemplary embodiment of a device 1.3 for recognizing an object, illustrated by a simplified block diagram. The exemplary device 1.3 differs from the exemplary device 1 of the Fig. 1 by using, in addition to parameters a) to d), the parameter e) and the parameter e') described above.

[0067] The exemplary device 1.3 has a comparison unit 2.3, which preferably has the same function as the comparison unit 2 of the exemplary device 1 and is additionally configured to compare the detected object O2 of the second single image E1 with the detected object O1 of the first single image E1 based on parameters a) to d) and e) and e'). The exemplary device 1.3 has an evaluation unit 3.3, which preferably has the same function as the evaluation unit 3 of the exemplary device 1 and is additionally configured to further evaluate the results of the comparison unit 2.3, taking into account parameter e) and parameter e'), in order to make an assessment of whether objects O1 and O2 are possibly identical.

[0068] Fig. Figure 5 shows an exemplary tree structure 50 of a trained model, which can be used in at least one of the embodiments of the devices 1, 1.1, 1.2, 1.3 described above. The model can be the AI ​​model described above.

[0069] The exemplary tree structure 50 is intended to illustrate the functionality of the model. The model functions like a very large and very complex decision tree. Evaluation unit 3, 3.1, 3.2, or 3.3 performs its evaluation, for example, using such a tree structure 50 to assess the equality of objects O1 and O2. This assessment is based on an "if-then" comparison. This can be done by equating an object in the current frame, i.e., object O2 in the second frame E2, with an object in the previous frame, i.e., object O1 in the first frame E1, if the comparison of the difference, for example, in height, width, distance, visual similarity, or historical and current overlap, yields certain results.

[0070] During its training, the model learns through hundreds of thousands of such evaluations in which ratio, combination and threshold values ​​the "if-then" comparisons must take place in order to lead to the desired results. Fig. Figure 6 shows, for example, an enlarged section of the exemplary tree structure 50 of the Fig. 5.

[0071] Fig. Figure 7 shows an exemplary environment 100 in which at least one embodiment of the devices 1, 1.1, 1.2, 1.3 described above can be used. For example, the environment 100 comprises a transmission system 110, in particular a streaming system, for example a real-time streaming system. For example, the environment 100 further comprises a video / image archive 120 and an analysis machine 130.

[0072] For example, the transmission system 110 comprises a video server 111 and one, several, or a multitude of video / image capture devices 112, which may be installed at multiple locations. Video / image capture devices 112 include, for example, CCTV (Closed Circuit Television) cameras, HD (High Definition) cameras, non-HD cameras, handheld cameras, and other video / image capture devices.

[0073] For example, the video server 111 receives dynamic images or video recordings from the video / image recording devices 112 and transmits the associated data in the form of input frames to the analysis machine 130. The video server 111 can process the input images before transmission and, if necessary, refine them, for example, by detecting and removing noise.

[0074] The video / image archive 120 can be a data storage device configured to store recorded or archived videos and / or images. The video / image archive 120 can consist of or contain a multitude of local or remote databases. The video / image archive 120, like the transmission system 110, can transmit data in the form of input images, in particular single images, to the analysis machine 130. Preferably, the transmission system 110, the video / image archive 120, and the analysis machine 130 are connected via a network 140.

[0075] Network 140 can be any suitable wired network, wireless network, a combination of both, or any other conventional network. Examples include a LAN connection, a WLAN connection, an internet connection, a point-to-point connection, or any other network connection or combination thereof. Network 140 can also be a local, regional, or global communications network, such as a corporate telecommunications network, the internet, a global mobile communications network, or any combination of similar networks.

[0076] For example, the analysis machine 130 is a component of at least one of the following systems: surveillance system, security system, retail system, control system, and business intelligence-based system. Preferably, the analysis machine 130 receives the input frames, in particular individual images, from the transmission system 110 and / or the video / image archive 120.

[0077] Preferably, the analysis machine 130 comprises a recognition device 131 for recognizing objects, an embodiment of the recognition devices 1, 1.1, 1.2, 1.3 described above, and a database 132. For example, the input frames are stored in the database 132, in particular temporarily stored.

[0078] The recognition device 131 serves, for example, to recognize objects on the input frames or individual images. The recognition device 131 performs tasks such as locating the object on the individual image and assigning an object class to the object. The information about recognized objects determined by the recognition device 131 is provided as data and stored, for example, on the database 132, and is available for further processing or evaluation in the recognition device 1, 1.1, 1.2, or 1.3.

[0079] The results generated by the analysis machine 130 can be remotely monitored and / or automatically sent to a variety of user devices 150 as push messages or notifications. For example, the analysis machine 130 is connected to the user devices 150 via wired or wireless communication means. These user devices 150 could include a smartphone, mobile device, telephone, personal digital assistant (PDA), computer, workstation, notebook, mainframe, or any equivalent device capable of processing, sending, and receiving data.

[0080] Fig. Figure 8 shows an example of a single image with a marked object within it, which is defined by a so-called bounding box. In this case, the object is assigned a bounding box and an object class, here the object class "Person". The bounding box is characterized by the coordinates of "top left" and "bottom right".

[0081] Fig. Figure 9 shows an example of a single image, which was taken at a later time than the single image of the Fig. 8 was recorded, but the same area is captured there. This is shown in the single image of the Fig. The 8 detected objects have shifted in the single image of the Fig. 9 moved further. This is exemplified in the Fig. 9 is represented by the second bounding box present there. The coordinates of the moving object have also been determined and an object class assigned. Reference symbol list 1 Recognition device 1.1, 1.2 Recognition device 1.3, 1.4 Recognition device 2 Comparative institution 2.1, 2.2 Comparative Institution 2.3, 2.4 Comparative Institution 3 Evaluation unit 3.1, 3.2 Evaluation unit 3.3, 3.4 Evaluation unit 50 tree structure 100 surroundings 110 transmission system 111 video servers 112 Video / image recording devices 120 video / image archive 130 analysis machine 131 Detection device 132 database 140 network 150 user devices a), b), c) Parameter d), e), e') Parameter

Claims

A method for recognizing an object, comprising the steps: i) comparing a recognized object of a second frame with a recognized object of a first frame based on at least one of the following parameters: a) distance, in particular Euclidean distance between the objects; b) difference between the objects in their geometric dimensions, in particular the height and / or width of the objects; c) difference between the objects in their preferably optical and / or acoustic appearance; d) similarity of the objects with respect to an object type; ii) evaluating the results from step i) to determine whether the recognized object of the second frame is the recognized object of the first frame. Method according to claim 1, wherein step i) is performed using at least two of the parameters a) to d). Method according to claim 1 or 2, wherein step i) is performed using parameters a) and b). Method according to one of the preceding claims, wherein step i) is carried out using at least three of the parameters a) to d). Method according to one of the preceding claims, wherein step i) is performed using parameters a) and b) and c). Method according to one of the preceding claims, wherein step i) is performed using all four parameters a) to d). Method according to one of the preceding claims, wherein step i) is performed using a further parameter: • Degree of occlusion of at least one of the objects by a third object. Method according to one of the preceding claims, wherein step i) is performed using a further parameter: • Degree of occlusion of the object of the second single image by a third object of the first single image. Method according to one of the preceding claims, wherein the results from the at least one comparison of step i) in metadata are used as the basis for step ii). Method according to one of the preceding claims, wherein the method uses an AI model to determine whether the detected object of the second frame is the detected object of the first frame. Method according to one of the preceding claims, wherein the method uses a random forest algorithm to determine whether the detected object of the second frame is the detected object of the first frame. Method according to one of the preceding claims, wherein the method uses an XGBoost algorithm to determine whether the detected object of the second frame is the detected object of the first frame. Method for detecting and re-detecting an object, comprising steps for: • providing a first single image; • detecting an object on the first single image; • providing a second single image; • detecting an object on the second single image; • performing a method according to any one of claims 1 to 12. Method according to claim 13, wherein the second single image was created at a later time than the first single image. Method according to claim 13 or 14, wherein information about the detected objects is provided in data. Machine learning model trained by a method according to any one of claims 1 to 12. Model according to claim 16, comprising a random forest algorithm to determine whether a detected object of a second frame is a detected object of a first frame. Model according to claim 16 or 17, comprising an XGBoost algorithm to determine whether a detected object of a second frame is a detected object of a first frame. Computer program product comprising program code stored on a computer-readable medium for carrying out a method according to any one of claims 1 to 12 and in particular comprising a model according to any one of claims 16 to 18. Device (1) for recognizing an object, comprising a comparison device (2) configured to compare a recognized object (O2) of a second frame (E2) with a recognized object (O1) of a first frame (E1) based on at least one of the following parameters: a) distance, in particular Euclidean distance between the objects; b) difference between the objects in their geometric dimensions, in particular the height and / or width of the objects; c) difference between the objects in their preferably optical and / or acoustic appearance; d) similarity of the objects with respect to an object type; an evaluation device (3) configured to evaluate the results of at least one comparison by the comparison device (2) in order to determine whether the object (O2) of the second frame (E2) is the object (O1) of the first frame (E1).

Images