Noise filtering device and method based on machine learning

Abstract

A machine learning-based noise filtering device according to one embodiment of the present disclosure may include a camera that acquires image data corresponding to a first range where an object is located, a radar that acquires radar data in a second range corresponding to the first range, at least one processor, and at least one memory including computer program code. The at least one memory and the computer program code may be configured so that the device, through the at least one processor, inputs the image data or the radar data into a filter model to determine whether the radar data related to the object is noise. The filter model may be a model that determines whether sensor fusion is possible based on first training data corresponding to the image data and second training data corresponding to the radar data, and learns the features of the first training data or the second training data that label the determined possibility of sensor fusion.

Description

Technology Field

[0001] The present disclosure relates to a machine learning-based noise filtering device and method. Background Technology

[0002] In road traffic involving both vehicles and pedestrians, the potential for accidents is always present, and indeed, numerous traffic accidents and dangerous situations occur nationwide every day. Particularly in high-speed driving environments such as highways or at night, drivers' ability to perceive their surroundings is likely to be significantly impaired, which can lead directly to major traffic accidents. While various technologies have been proposed to detect road conditions, conventional methods have been limited to detecting objects ahead using infrared cameras installed on roads or vehicles and notifying drivers; however, this approach suffers from the disadvantage of low detection efficiency. Recently, technologies utilizing radar to collect traffic information are being proposed to formulate traffic policies aimed at reducing time and costs associated with traffic congestion and enhancing the efficiency of road operations.

[0003] Radar sensors detect signals that are reflected back from objects based on the radio wave information they generate. Along with the signals actually reflected back from the vehicle, various noises enter the radar sensor; the sensor refines this into point data through a preprocessing step that retains only the meaningful signals. This point data may include false point data generated by some noise signals; however, the radar sensor filters out these noise signals using various filtering techniques and tracks multiple consecutively generated point data to identify a vehicle. Considering that multiple point data may be generated from a single vehicle, the tracking process includes a homogeneous data fusion algorithm that groups signals related to the vehicle's length and width.

[0004] However, radar sensors have a problem in that it is difficult to filter false point data in enclosed environments such as tunnels. The problem to be solved

[0005] One objective of the present disclosure is to filter noise from radar signals in an environment where accurate signal acquisition is difficult due to the occurrence of a large amount of noise.

[0006] In addition, another objective of the present disclosure is to filter noise by using image data from a camera in addition to radar, and to filter noise using a machine learning-based filtering model when the performance of the camera deteriorates.

[0007] However, the technical problems that this embodiment aims to solve are not limited to the technical problems described above, and other technical problems may exist. means of solving the problem

[0008] A machine learning-based noise filtering device according to one embodiment of the present disclosure may include a camera that acquires image data corresponding to a first range where an object is located, a radar that acquires radar data in a second range corresponding to the first range, at least one processor, and at least one memory including computer program code. The at least one memory and the computer program code may be configured so that the device, through the at least one processor, inputs the image data or the radar data into a filter model to determine whether the radar data related to the object is noise. The filter model may be a model that determines whether sensor fusion is possible based on first training data corresponding to the image data and second training data corresponding to the radar data, and learns the features of the first training data or the second training data that label the determined possibility of sensor fusion.

[0009] A machine learning-based noise filtering method according to one embodiment of the present disclosure may include the operation of acquiring image data or radar data from a camera that acquires image data corresponding to a first range where an object is located or from a radar that acquires radar data in a second range corresponding to the first range, and the operation of inputting the image data or the radar data into a filter model to determine whether the radar data is noise. The filter model may be a model that determines whether sensor fusion is possible based on first training data corresponding to the image data and second training data corresponding to the radar data, and learns the features of the first training data or the second training data that have labeled the determined possibility of sensor fusion. Effects of the invention

[0010] According to the machine learning-based noise filtering device and method of the present disclosure, the accuracy of vehicle detection can be improved by accurately distinguishing a signal detected by a radar into a true signal and a noise signal.

[0011] In addition, according to the machine learning-based noise filtering device and method of the present disclosure, additional distance-based filtering is possible using other sensor data such as a camera, and filtering may be possible even when the performance of the camera deteriorates. Brief explanation of the drawing

[0012] Figure 1 illustrates a problem situation according to the prior art. FIG. 2 is a configuration diagram of a machine learning-based noise filtering device according to one embodiment of the present disclosure. FIG. 3 is an input / output signal diagram of a filter model according to one embodiment of the present disclosure. FIG. 4 is a distribution of radar signals for an object according to one embodiment of the present disclosure. Figure 5a illustrates an image of a test situation. FIG. 5b illustrates a radar map in a test situation according to the prior art. FIG. 5c illustrates a radar map in a test situation according to one embodiment of the present disclosure. FIG. 6 illustrates a flowchart of a machine learning-based noise filtering method according to one embodiment of the present disclosure. Specific details for implementing the invention

[0013] Embodiments of the present invention are described below with reference to the attached drawings to enable those skilled in the art to easily implement the invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0014] Throughout this specification, when a component is described as being located "on" another component, this includes not only cases where a component is in contact with another component, but also cases where another component exists between the two components.

[0015] Throughout this specification, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0016] Throughout this specification, terms of degree such as “about,” “substantially,” etc., are used to mean at or near the stated value when inherent manufacturing and material tolerances are presented in the stated meaning, and are used to prevent unscrupulous infringers from unfairly exploiting the disclosure in which precise or absolute values ​​are mentioned to aid in understanding this specification. Throughout this specification, terms of degree such as “step” or “step of” do not mean “step for”.

[0017] Throughout this specification, the term “combination(s) of these” included in the Markush-type expression means one or more mixtures or combinations selected from the group consisting of the components described in the Markush-type expression, and means including one or more selected from the group consisting of said components.

[0018] Throughout this specification, the description "A and / or B" means "A or B, or A and B".

[0019] Hereinafter, embodiments and examples of the present invention will be described in detail with reference to the attached drawings. However, the present invention may not be limited to these embodiments and examples and drawings.

[0020] Figure 1 illustrates a problem situation according to the prior art.

[0021] Referring to FIG. 1, a radar device (not shown) can track multiple point data that occur in succession and recognize them as a vehicle. Considering that multiple point data may occur from a single vehicle, the tracking process may include a homogeneous data fusion algorithm that groups signals included in the length and width of the vehicle.

[0022] A radar device (not shown) that detects radar signals reflected back from an object by generated electromagnetic waves can receive various noises along with the reflected signal reflected back from the object.

[0023] Radar signals can be refined into point data through a preprocessing step that retains only meaningful signals. However, even the refined point data may contain false point data due to noise, and the phenomenon of such false point data occurring is called "ghosting." Ghost filtering methods to filter out false point data have existed previously.

[0024] Conventional ghost filtering methods primarily generate point data by passing through CFAR (Constant False Alarm Rate) during the preprocessing stage, and then undergo a radar-specific ghost filtering stage.

[0025] Ghost filtering techniques include Field of View (FOV) filters, Double Bounce filters, and Multi-Angle filters. The FOV filter is a logic that filters signals outside of radar performance standards; for example, in a radar with a 50-degree left-right field of view (FOV), if a signal is detected at an angle exceeding 50 degrees, it can be filtered out. The Double Bounce filter is a logic that determines when a single signal reflects off the same object multiple times and its velocity and distance values ​​are doubled; if there is a signal with doubled values ​​for both velocity and distance, it can filter out signals other than the closest one. The Multi-Angle filter is used when signals with the same velocity and distance are reflected off smooth surfaces and detected from multiple angles; it can retain signals with stable strength that are repeatedly detected across multiple frames while filtering out unstable signals.

[0026] However, unlike general road environments, it is difficult to filter out false point data using conventional ghost filtering methods in environments where accurate signal collection is challenging due to diffuse reflection, such as tunnels. Ghosting phenomena are more frequent in tunnels than in the outside due to their enclosed nature, and radar signals are easily reflected off the smooth surfaces inside the tunnel, making it difficult to distinguish between true and false point data. In particular, because false point data often exhibits complex characteristics—such as having no significant difference in signal strength between true and false data, or even possessing stronger strength than true data, and being detected repeatedly across multiple frames just like true data—it is difficult to filter out false point data using existing ghost filtering methods.

[0027] As shown in Fig. 1, when comparing the vehicles (T, G) marked with circles in the image and the radar map, it can be seen that only one vehicle (T) is observed in the image, whereas two vehicles (T, G) are detected in the radar map due to ghosting.

[0028] Therefore, with the existing ghost filtering method, false point data (G) is not filtered out, and consequently, false point data (G) may be incorrectly recognized as a vehicle, which could degrade road control performance.

[0029] FIG. 2 is a configuration diagram of a machine learning-based noise filtering device (200) according to one embodiment of the present disclosure. FIG. 3 is an input / output signal diagram of a filter model (M) according to one embodiment of the present disclosure.

[0030] Referring to FIGS. 2 and FIGS. 3, a machine learning-based noise filtering device (200) according to one embodiment may include a camera (210), a radar (220), a processor (230), a memory (240), and a communication module (250).

[0031] The camera (210) can capture still images and video. According to one embodiment, the camera (210) may include one or more lenses, image sensors, image signal processors, or flashes. The camera (210) may be positioned to acquire image data corresponding to a specified range (e.g., a first range).

[0032] The radar (220) can detect an object and determine its direction, distance, speed, etc. by measuring the reflected waves that return after the electromagnetic waves irradiated by the object strike it. According to one embodiment, the radar (220) may include antennas corresponding to a transmitter that transmits electromagnetic waves and a receiver that receives reflected waves, respectively. The radar (220) may be positioned to acquire radar data corresponding to a designated range (e.g., a second range). For example, the second range may be the same as the first range, but there may be some differences depending on the difference in the field of view (FOV) of the camera (210) and the radar (220).

[0033] The processor (230) can control at least one other component (e.g., hardware or software component) of the machine learning-based noise filtering device (200) connected to the processor (230) by executing software (e.g., a program), and can perform various data processing or operations. According to one embodiment, as at least part of the data processing or operations, the processor (230) can store instructions or data received from other components in volatile memory, process the instructions or data stored in volatile memory, and store the resulting data in non-volatile memory. According to one embodiment, the processor (230) may include a main processor (e.g., a central processing unit or an application processor) or an auxiliary processor (e.g., a graphics processing unit, a neural processing unit (NPU), an image signal processor, a sensor hub processor, or a communication processor) that can operate independently or together with the main processor. For example, the auxiliary processor may be configured to use lower power than the main processor or to be specialized for a specified function, and the auxiliary processor may be implemented separately from the main processor or as part thereof.

[0034] An auxiliary processor may control at least some of the functions or states associated with at least one component of the machine learning-based noise filtering device (200), for example, on behalf of the main processor while the main processor is in an inactive (e.g., sleep) state, or together with the main processor while the main processor is in an active (e.g., application execution) state. According to one embodiment, the auxiliary processor (e.g., image signal processor or communication processor) may be implemented as part of another functionally related component. According to one embodiment, the auxiliary processor (e.g., neural network processing unit) may include a hardware structure specialized for processing an artificial intelligence model. The artificial intelligence model may be generated through machine learning. Such learning may be performed, for example, within the machine learning-based noise filtering device (200) itself where the artificial intelligence model is executed, or through a separate server. The learning algorithm may include, for example, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but is not limited to the examples described above. The artificial intelligence model may include multiple artificial neural network layers. The artificial neural network may be a deep neural network (DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), a restricted Boltzmann machine (RBM), a deep belief network (DBN), a bidirectional recurrent deep neural network (BRDNN), deep Q-networks, or a combination of two or more of the above, but is not limited to the examples described above.Artificial intelligence models may include software structures, either additionally or as a substitute, in addition to hardware structures.

[0035] The memory (240) can store various data used by at least one component (e.g., processor (230)) of the machine learning-based noise filtering device (200). The data may include, for example, software (e.g., a program) and input data or output data for related instructions. The memory (240) may include volatile memory or non-volatile memory. The program may be stored in the memory (240) as software and may include, for example, an operating system, middleware, or an application.

[0036] The communication module (250) can support the establishment of a direct (e.g., wired) communication channel or a wireless communication channel between a machine learning-based noise filtering device (200) and an external electronic device or server, and the performance of communication through the established communication channel. The communication module (250) may include one or more communication processors that operate independently of the processor (230) (e.g., application processor) and support direct (e.g., wired) communication or wireless communication. According to one embodiment, the communication module (250) may include a wireless communication module (e.g., cellular communication module, short-range wireless communication module, or GNSS (global navigation satellite system) communication module) or a wired communication module (e.g., LAN (local area network) communication module, or power line communication module).

[0037] A machine learning-based noise filtering device (200) (e.g., processor (230)) according to one embodiment can filter noise by utilizing multiple sensors as well as radar data acquired from a radar (220).

[0038] A machine learning-based noise filtering device (200) according to one embodiment can determine whether the camera (210) is in a degraded state. For example, the camera (210) may experience degradation, such as reduced image classification performance, due to contamination of the lens, etc., caused by environmental factors such as smoke or dust in a closed space like a tunnel. The machine learning-based noise filtering device (200) can determine whether the camera (210) is in a degraded state based on image data obtained from the camera (210).

[0039] A machine learning-based noise filtering device (200) according to one embodiment can filter noise in radar data based on image data and radar data when the camera (210) is in an undegraded state.

[0040] In one embodiment, the machine learning-based noise filtering device (200) can acquire image data from a camera (210) and acquire radar data from a radar (220). In another embodiment, the machine learning-based noise filtering device (200) can receive at least one of image data or radar data from the outside through a communication module (250).

[0041] A machine learning-based noise filtering device (200) according to one embodiment can determine whether radar data can be fused with image data and sensor data through a filter model (M). The machine learning-based noise filtering device (200) can determine radar data that can be fused with image data and sensor data as not being noise, and determine radar data that cannot be fused with image data and sensor data as being noise.

[0042] In one embodiment, the filter model (M) can be trained to determine whether sensor fusion is possible by determining whether the distance between the coordinates corresponding to the image data and the coordinates corresponding to the radar data is less than or equal to a preset first distance.

[0043] For example, a machine learning-based noise filtering device (200) can recognize image data acquired from a camera (210) as a unique ID through an algorithm (e.g., YOLOv5), and then convert and display the coordinates of the vehicle on the image plane into a Bird-eye View. In addition, the machine learning-based noise filtering device (200) can acquire color information and vehicle type information of the vehicle through the camera (210).

[0044] For example, a machine learning-based noise filtering device (200) can collect radio wave information reflected from a vehicle from radar data acquired from a radar (220), and through signal processing, can represent the coordinates of a vehicle with a unique ID in a Bird-eye View coordinate system identical to an image.

[0045] FIG. 4 is a distribution of radar (220) signals for an object according to one embodiment of the present disclosure.

[0046] Referring further to FIG. 4, a preset first distance related to the distance error between the coordinates corresponding to the image data and the coordinates corresponding to the radar data can be set based on the size of the object to be detected. For example, if the object is a vehicle (e.g., a sedan), the distance from the center of the front of the vehicle (2D coordinates where the camera (210) detects the vehicle) to the top corner of the rear (a limit point that can be detected by the radar (220)) can be set as the first distance.

[0047]

[0048] Here, w is the width of the front of the car, and l is the length of the side of the car.

[0049] Additionally, the signal for an approaching vehicle detected by the radar (220) consists mostly of signals reflected from the front hood and windshield, and 99.9% of the radar (220) signals can be collected within 3.6 [m]. Assuming that the amount of data decreases exponentially with distance, the value of λ of the exponential distribution can be calculated as follows.

[0050]

[0051] Using λ=1.918 calculated using the above formula, the distance (ε) at which 99% of radar data can be detected probabilistically can be calculated as 2.5 [m] as follows. For example, the first distance can be set to 2.5 [m].

[0052]

[0053] In one embodiment, the filter model (M) can be trained to determine whether sensor fusion is possible, that the distance between the coordinates corresponding to image data from the camera (210) or the coordinates corresponding to radar data from the radar (220) is less than or equal to a preset second distance.

[0054] Here, the preset second distance is preset according to the performance of the camera (210) or radar (220), and can be set to, for example, 120 [m].

[0055] In one embodiment, the filter model (M) can be trained to determine that sensor fusion is possible when both radar data and image data satisfy the two conditions.

[0056] A machine learning-based noise filtering device (200) according to one embodiment can filter noise in radar data based on radar data when the camera (210) is in a degraded state. The machine learning-based noise filtering device (200) can determine whether there is noise in radar data related to an object by inputting radar data acquired from a radar (220) in normal operation into a filter model (M). The machine learning-based noise filtering device (200) can determine whether there is noise in radar data by extracting features from the input radar data and determining whether sensor fusion is possible for radar data corresponding to the extracted features based on the filter model (M).

[0057] In one embodiment, the filter model (M) may be a model that determines whether sensor fusion is possible based on first training data corresponding to image data and second training data corresponding to radar data, and learns the features of the first training data or second training data that label the determined possibility of sensor fusion.

[0058] The filter model (M) receives first training data and second training data as input to determine whether sensor fusion is possible, and can learn by labeling data where sensor fusion is possible as true (1) and data where sensor fusion is impossible as false (0). The filter model (M) can be trained to determine that sensor fusion is possible if the separation distance between the first training data and the second training data is less than or equal to a preset first distance, and at the same time, the distance from the camera (210) or the radar (220) is less than or equal to a preset second distance.

[0059] The filter model (M) can learn the absolute characteristics of the data and the relative characteristics with respect to other data as features of the first training data or the second training data. The absolute characteristics of the data may include the distance (r_i), velocity (v_i), or azimuth (θ_i) from the camera (210) or radar (220) corresponding to the data. The relative characteristics with respect to other data may include the distance difference (r_near - r_i), velocity difference (v_near - v_i), or azimuth difference (θ_near - θ_i) with respect to the data closest to the data (near) among the data included in the frame.

[0060] Here, the data closest to the corresponding data (near) refers to the point data closest to the i-th point data among all point data stored in the frame of the corresponding data. For example, the data closest to the corresponding data (near) may be the point data where the Euclidean norm is minimized.

[0061] The filter model (M) of the present disclosure can learn the distance difference and azimuth angle with other data containing location information because there is a high probability that false point data due to tunnel reflection will occur at a location close to the true point data. Additionally, the filter model (M) can learn the speed difference because false point data occurs at a similar speed around the true point data.

[0062] In one embodiment, the filter model (M) may be a model learned by normalizing the features of the first training data or the second training data. The machine learning-based noise filtering device (200) may normalize the features of the radar data as at least part of its operation to determine whether the radar data is noise, and input the normalized features to the filter model (M).

[0063] Data normalization eliminates dependency on specific locations, thereby allowing a model learned from a radar (220) installed in a specific tunnel to be applied to a radar (220) installed at a different location. In other words, independence from the sensor installation location can be ensured by scaling the variation in measurement values ​​that occurs depending on the sensor installation location. For example, as a normalization method, a min-max method can be used to scale specific data based on the minimum and maximum values ​​of the entire data.

[0064] In one embodiment, the filter model (M) may be a machine learning trained model that uses the Random Forest algorithm and the Adaptive Boosting algorithm together.

[0065] Random Forest is an ensemble technique that combines multiple decision trees to improve normalization performance. Compared to decision trees, it offers relatively higher accuracy and is robust against noise and outliers.

[0066] AdaBoost (Adaptive Boosting) can be adopted as an ensemble method different from Random Forest. Unlike Random Forest, which reduces the variance of the learning error, AdaBoost focuses on reducing the bias of the learning process and can reduce the error of the entire model by iteratively creating models.

[0067] Therefore, the filter model (M) can improve both the accuracy and generalization of learning by combining the advantages of high-dimensional analysis and noise removal, as well as high-difficulty pattern learning, by using Random Forest and Ada Boost together.

[0068] The filter model (M) can optimize the hyperparameters of Random Forest and Ada Boost. For example, the hyperparameters can be optimized by utilizing cross-validation and grid search methods.

[0069] Random Forest was tested with maximum depth values ​​of 3, 5, and 7, and minimum sample in a leaf values ​​of 5, 10, and 15, while AdaBoost was tested with estimator values ​​of 10, 15, and 20, and learning rates of 0.1, 0.5, and 1.0, respectively. Based on the test results, maximum depth of 7, minimum sample in a leaf of 5, estimator value of 10, and learning rate of 0.1 can be adopted as optimal values.

[0070] When noise filtering is tested according to the machine learning-based noise filtering device (200) of the present disclosure, it can be evaluated as a confusion matrix as follows.

[0071] Predicted Positive Predicted Negative Actual Positive 0.974 0.0255 Actual Negative 0.246 0.757

[0072] Referring to Table 1 above, it can be confirmed that noise is conservatively identified, as the probability that actual positive point data is judged as false (Predited Negative) is very low at 0.0255. Additionally, it is confirmed that the probability that actual negative point data is judged as true (Predicted Negative) is relatively high at 0.246, but from a filtering perspective, it has the performance to filter out more than 3 / 4 of false point data. In one embodiment, the machine learning-based noise filtering device (200) may not use radar data in the object tracking algorithm if it determines that the radar data is noise. In the case where the machine learning-based noise filtering device (200) determines that the radar data is not noise, it may track the object based on the radar data.

[0073] FIG. 5a illustrates an image of a test situation. FIG. 5b illustrates a radar map in a test situation according to the prior art. FIG. 5c illustrates a radar map in a test situation according to an embodiment of the present disclosure.

[0074] Referring to FIGS. 5a to 5c, as shown in FIG. 5a, only one vehicle (T) is identified in the area marked with a circle in the image captured by the camera (210). When using a filtering technique according to the prior art, as shown in FIG. 5b, the radar data corresponds to one vehicle (T) and is identified as two vehicles (T, G) by true radar data (T) and false radar data (G).

[0075] In contrast to this, according to the machine learning-based noise filtering device (200) according to one embodiment of the present disclosure, as illustrated in FIG. 5c, radar data is identified as one vehicle (T). That is, it can be confirmed that false radar data (G) is filtered by the machine learning-based noise filtering device (200) according to one embodiment of the present disclosure.

[0076] FIG. 6 illustrates a flowchart (600) of a machine learning-based noise filtering method according to one embodiment of the present disclosure.

[0077] Referring to FIG. 6, a machine learning-based noise filtering device (200) according to one embodiment can generate a filter model (M) that learns the features of the first training data or the second training data in operation 610. For example, the machine learning-based noise filtering device (200) may learn the filter model (M) by continuously input data, or may obtain a filter model (M) that has been previously learned.

[0078] A machine learning-based noise filtering device (200) according to one embodiment can acquire image data or radar data from a camera (210) or radar (220) in operation 630.

[0079] A machine learning-based noise filtering device (200) according to one embodiment can determine whether the camera (210) is in a degraded state in operation 650.

[0080] A machine learning-based noise filtering device (200) according to one embodiment can determine whether there is noise in the radar data by inputting image data or radar data into a filter model (M) in operation 670.

[0081] For example, if the machine learning-based noise filtering device (200) determines that the camera (210) is not degraded, it can input image data and radar data into a filter model (M) to determine whether there is noise in the radar data.

[0082] For example, when a machine learning-based noise filtering device (200) determines that the camera (210) is in a degraded state, it inputs radar data into a filter model (M) to determine whether the radar data is noise.

[0083] A machine learning-based noise filtering device (200) according to one embodiment can track an object based on radar data when, in operation 690, radar data is determined not to be noise.

[0084] In the machine learning-based noise filtering device (200) described above, the machine learning-based noise filtering method may also be implemented in the form of a computer program stored on a computer-readable recording medium executed by a computer or in the form of a recording medium containing instructions executable by a computer. Additionally, in the machine learning-based noise filtering device (200) described above, the machine learning-based noise filtering method may also be implemented in the form of a computer program stored on a computer-readable recording medium executed by a computer.

[0085] A computer-readable recording medium may be any available medium accessible by a computer and includes both volatile and non-volatile media, and both removable and inremovable media. Additionally, a computer-readable recording medium may include a computer storage medium. A computer storage medium includes both volatile and non-volatile, removable and inremovable media implemented by any method or technique for storing information such as computer-readable instructions, data structures, program modules, or other data.

[0086] The function realized by the component described herein may be realized in a general-purpose processor (230), a specific-purpose processor, an integrated circuit, an ASIC (Application Specific Integrated Circuits), a CPU (Central Processing Unit), a circuit, and / or a combination thereof, which is programmed to realize the described function. The processor (230) includes transistors or other circuits and is considered to be a circuit or a processing circuit. The processor (230) may be a programmed processor that executes a program stored in memory (240).

[0087] In this specification, circuits, parts, units, and means are hardware programmed to perform or execute the described functions. Such hardware may be any hardware disclosed in this specification or any hardware known to be programmed or execute the described functions.

[0088] If the hardware is a processor (230) considered to be of a circuit type, the circuit, the part, means, or unit is a combination of the hardware and the software used to configure the hardware and / or the processor (230).

[0089] The foregoing description of the present disclosure is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the present disclosure. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0090] The scope of the present disclosure is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present disclosure. Explanation of the symbols

[0091] 200 : Noise filtering device 210 : Camera 220 : Radar 230 : Processor 240 : Memory 250: Communication module M: Filter Model

Claims

Claim 1 A machine learning-based noise filtering device comprises: a camera that acquires image data corresponding to a first range in which an object is located; a radar that acquires radar data in a second range corresponding to the first range; and at least one processor. A machine learning-based noise filtering device comprising at least one memory including computer program code, wherein the at least one memory and the computer program code are configured through the at least one processor to input the image data or the radar data into a filter model to determine whether there is noise in the radar data related to the object, wherein the filter model is a model that determines whether sensor fusion is possible based on first learning data corresponding to the image data and second learning data corresponding to the radar data, and learns the features of the first learning data or the second learning data that label the determined possibility of sensor fusion, and wherein the filter model is learned to determine that sensor fusion between the radar data and the image data is possible if the distance between the coordinates corresponding to the image data and the coordinates corresponding to the radar data is less than or equal to a first distance set based on the size of the object, and the distance between the camera and the coordinates corresponding to the image data or between the radar and the coordinates corresponding to the radar data is less than or equal to a preset second distance. Claim 2 A machine learning-based noise filtering device according to claim 1, wherein the at least one memory and the computer program code are configured to operate through the at least one processor to determine whether the device is noise in the radar data, at least as part of the operation, to determine whether the radar data is capable of sensor fusion based on the characteristics of the radar data, and to determine whether the radar data is noise based on whether the determined radar data is capable of sensor fusion. Claim 3 delete Claim 4 A machine learning-based noise filtering device according to claim 1, wherein the filter model is a model that learns the absolute characteristics of the data and the relative characteristics with respect to other data as a characteristic of the first training data or the second training data. Claim 5 A machine learning-based noise filtering device according to claim 4, wherein the absolute characteristics of the corresponding data include a separation distance, speed, or azimuth angle corresponding to the corresponding data, and the relative characteristics with respect to other data include a difference in distance, speed, or azimuth angle with the data closest to the corresponding data among the data included in the frame. Claim 6 In claim 1, the filter model is a model learned by normalizing the features of the first training data or the second training data, and the at least one memory and the computer program code are configured to normalize the features of the radar data and input the normalized features into the filter model as at least part of the operation to determine whether the radar data is noise through the at least one processor, a machine learning-based noise filtering device. Claim 7 A machine learning-based noise filtering device according to claim 1, wherein the filter model is a machine learning-trained model that uses the Random Forest algorithm and the Adaptive Boosting algorithm together. Claim 8 A machine learning-based noise filtering device according to claim 1, wherein the at least one memory and the computer program code are configured through the at least one processor to: determine whether the camera is in a degraded state, and based on the result of determining whether the degradation is, input the radar data into the filter model. Claim 9 In claim 1, the machine learning-based noise filtering device configured to track an object based on radar data when the radar data is determined to be non-noise, through the at least one memory and the computer program code, the device through the at least one processor. Claim 10 A machine learning-based noise filtering method comprises: an operation of acquiring image data or radar data from a camera that acquires image data corresponding to a first range in which an object is located, or from a radar that acquires radar data in a second range corresponding to the first range; and an operation of inputting the image data or the radar data into a filter model to determine whether the radar data is noise, wherein the filter model is a model that determines whether sensor fusion is possible based on first training data corresponding to the image data and second training data corresponding to the radar data, and learns the features of the first training data or the second training data that label the determined possibility of sensor fusion, and wherein the filter model is a model that learns to determine that sensor fusion is possible between the radar data and the image data when the distance between the coordinates corresponding to the image data and the coordinates corresponding to the radar data is less than or equal to a first distance set based on the size of the object, and the distance between the camera and the coordinates corresponding to the image data or between the radar and the coordinates corresponding to the radar data is less than or equal to a preset second distance.

Images