Method and system for object detection

By providing signal data and pyramid hierarchical structures at multiple spatial scales, combined with machine learning models, the problem of excessive Doppler channel points caused by stationary objects in linear frequency modulation sequence radar is solved, improving the efficiency and accuracy of object detection, and reducing the Calculate the cost.

CN112241003BActive Publication Date: 2024-12-06APTIV TECHNOLOGIES AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010679468.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-07-17
Filing Date
2020-07-15
Publication Date
2024-12-06
Estimated Expiration
2040-07-15

AI Technical Summary

Technical Problem

In the signal processing of linear frequency modulation sequence radar, there are too many Doppler channel points caused by static objects, making it difficult to distinguish multiple reflectors, and high resolution technology is required, which has high computing costs.

Method used

By providing signal data at multiple spatial scales, signal data at different scales are provided to multiple detectors, and object detection is performed using pyramid hierarchical structure combined with machine learning models, including processing of distance, speed and angle information, and object attributes are used for neural networks. prediction and classification.

Benefits of technology

It improves the efficiency and accuracy of object detection, can effectively process objects of different sizes, reduces the amount of training data for machine learning models, enhances processing capabilities, and reduces computing costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112241003B_ABST
    Figure CN112241003B_ABST
Patent Text Reader

Abstract

The present application discloses a method and system for object detection. The computer-implemented method includes the following steps performed by computer hardware components: providing signal representation data including distance information, speed information, and angle information; for each spatial scale in a plurality of spatial scales, determining corresponding scale data of the corresponding spatial scale based on the signal representation data to obtain a plurality of scale data; providing the plurality of scale data to a plurality of detectors; and each detector performing object detection based on at least one scale data in the plurality of scale data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to methods for object detection, for example using radar data. Background Art

[0002] In conventional signal processing of chirp sequence radars, the goal is to create detection points by estimating the distance, radial velocity, and angle of physical points scattered in the measurement environment, which reflect the energy emitted by the radar transmitter back to the receiver. The amount of reflected energy is very dependent on the material, orientation, geometry, and size of the object. In this case, Fourier transforms are usually used to convert time / spatial data into frequency responses that can be interpreted by humans. A common approach is to create multiple range-Doppler maps, which indicate the radial distance and velocity of objects in the scene, respectively. These responses are then set as thresholds to filter out any responses that may be noise or interference and therefore not of concern. As a further step, the remaining detections are processed to extract the angle in the airspace (using a receiver array), which is then used together with the distance information of the reflection point to accurately locate the reflection point in the scene.

[0003] The problem with this technique is that for each range Doppler bin, the spatial information covers a fairly large field of view (e.g., -75 to +75 degrees), which may contain many scatterers. This is especially true if the scene being measured contains many stationary objects. If the autonomous vehicle (in other words: the vehicle on which the radar sensor is provided) is stationary, the stationary objects will cause many points in the zeroth Doppler bin to correspond to non-moving objects. In this case, distinguishing between multiple reflectors becomes very difficult, and computationally expensive high-resolution techniques are usually deployed to resolve the multiple reflection points.

[0004] Therefore, there is a need to provide systems and methods for efficient object detection. Summary of the invention

[0005] The present disclosure provides a computer-implemented method, a computer system, and a non-transitory computer-readable medium. Implementations are presented in the specification and drawings.

[0006] In one aspect, the present disclosure relates to a computer-implemented method for object detection, which includes the following steps performed by computer hardware components: providing signal representation data including distance information, speed information and angle information; determining corresponding scale data of the corresponding spatial scale based on the signal representation data for each spatial scale in a plurality of spatial scales to obtain a plurality of scale data; providing the plurality of scale data to a plurality of detectors; and each detector performing object detection based on at least one scale data in the plurality of scale data.

[0007] In other words, multiple scaled versions of the input signal data are provided to multiple detectors, and each detector uses at least one of the scaled versions for object detection.

[0008] Providing scale data of various scales to multiple detectors can enhance detection efficiency and detection accuracy, especially when dealing with objects of different sizes in the region of interest. The scale data can also be referred to as a spatial scale pyramid. For example, a deep pyramid level may include information that is more suitable for detecting large objects, while a shallow pyramid level may contain information that is more suitable for detecting small objects.

[0009] According to another aspect, the signal representative data is based on at least one of a radar signal and an ultrasonic signal. For example, the signal representative data may be information (e.g., energy or frequency) based on a reflection of a received transmitted signal. Based on the energy of the received reflection, and based on the time distribution of the energy of the received reflection (in other words: the received reflected signal), the distance of the object reflecting the transmitted signal may be determined (e.g., by determining the time between the emission of the signal and the reception of the reflected signal). According to another aspect, the signal representative data comprises frequency domain radar data. Based on a phase shift that can be derived from the received signal and its frequency, the speed of the object may be determined (e.g., based on a Doppler shift of the frequency of the reflected signal).

[0010] According to another aspect, the signal represents data based on a signal received by an antenna array.By determining the time differences of arrival between reflected signals received at the various antennas, the direction of arrival (in other words: angle) can be determined.

[0011] It will be understood that although the signal representation data includes distance information, speed information and angle information, this information (distance information, speed information and angle information) is not necessarily explicitly present in the signal representation data (in other words: determined or decoded), but rather this information can be derived from the signal representation data.

[0012] According to another aspect, signal representation data is determined as dense data based on sparse input data, preferably based on a transformation that preserves the spatial order of the sparse input data. It has been found that processing based on sparse data can hinder machine learning applications because it may significantly increase the amount of training data required to effectively train a machine learning model. Therefore, converting sparse data to dense data can enhance processing.

[0013] According to another aspect, multiple spatial scales are provided in a hierarchy from fine spatial resolution to coarse spatial resolution. According to various embodiments, data of various spatial scales (in other words: data with various spatial resolutions) are determined based on sharing information horizontally (i.e., within one level of the pyramid; in other words: between data of the same spatial resolution) and vertically (i.e., from one level of the pyramid to another level of the pyramid; in other words: between data of different spatial resolutions). Each level in the feature pyramid according to various embodiments may include multiple feature vectors, each feature vector describing a radar response from a different range. When determining data for coarse pyramid level features, features from several fine distances may be combined into a single coarse distance. Features from different distances may be combined, i.e., data may be combined in one dimension (1D). It will be understood that data of different pyramid levels may be determined using various methods, such as averaging, maximum pooling, neural networks, or any other method suitable for adjusting data to a coarser scale or a finer scale. Features in each level may be determined based on features in other levels and based on received radar data (e.g., a pattern of energy of received (e.g., reflected) radar signals).

[0014] According to another aspect, multiple spatial scales may be associated with the distance information of the signal representative data. In other words: the signal representative data is adjusted relative to the distance information. It has been found that this may enhance object detection of objects of varying sizes (e.g. when objects of small size and objects of much larger size are to be detected).

[0015] According to another aspect, each detector performs object detection in a respective predetermined spatial region. According to another aspect, the respective predetermined spatial regions of the plurality of detectors are provided over a predetermined range in polar coordinates.

[0016] The respective predetermined areas of the various detectors may vary in size. For example, there may be a small predetermined area and a large predetermined area. There may be multiple areas of different sizes.

[0017] According to another aspect, each detector provides a respective confidence level of the object detection performed. This may enhance further processing of the detection results, since detection results with a high confidence level may be treated differently from detection results with a low confidence level.

[0018] According to another aspect, each detector predicts the properties of the detected object when detecting the object. The properties of the detected object include at least one of the position of the detected object, the size of the detected object, the orientation of the detected object, the speed of the detected object, the shape of the detected object, or the category of the detected object. Detecting the category can classify the detected object into one of a plurality of categories. For example, it can be detected whether the detected object is a truck, a car, a bicycle, or a pedestrian. These properties can then be used for further processing, for example, in controlling an autonomous driving car.

[0019] According to another aspect, at least one of the steps is performed by a neural network. The neural network can be trained based on labeled data, the labeled data including reference object information and data, based on which reference signal representation data can be obtained. This can allow automatic learning of a complete object detection model without manually designing features, modules or encoding. The neural network can be, for example, a convolutional neural network (CNN), for example, a deep convolutional neural network (deep CNN).

[0020] In another aspect, the present disclosure relates to a computer system configured to perform some or all of the steps of the computer-implemented method described herein.

[0021] The computer system may include a plurality of computer hardware components (e.g., a processing unit (in other words: a processor), at least one memory unit (which may also be referred to as a memory) and at least one non-transitory data storage device, such as a hard drive or a solid-state drive). It will be understood that other computer hardware components may be provided and used to perform the steps of the computer-implemented method in the computer system. The non-transitory data storage device and / or the memory unit may include a computer program for instructing a computer, for example, using a processing unit and at least one memory unit to perform several or all steps or aspects of the computer-implemented method described herein.

[0022] In another aspect, the present disclosure relates to a non-transitory computer-readable medium including instructions for performing some or all of the steps or aspects of the computer-implemented methods described herein. The computer-readable medium may be configured as an optical medium such as a compact disc (CD) or a digital versatile disc (DVD); a magnetic medium such as a hard disk drive (HDD); a solid-state drive (SSD); a read-only memory (ROM) such as a flash memory, etc. In addition, the computer-readable medium may be configured as a data storage device accessible via a data connection such as an Internet connection. The computer-readable medium may be, for example, an online data repository or a cloud storage device.

[0023] The present disclosure also relates to a computer program for instructing a computer to execute several or all steps or aspects of the computer-implemented method described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Exemplary embodiments and functions of the present disclosure are described herein in conjunction with the following drawings which are schematically shown:

[0025] Figure 1 is a diagram illustrating an overview of the data flow from radar through a neural network structure to a final output of object detection according to various embodiments;

[0026] Figure 2A is an illustration of respective detectors having respective small areas, the detectors performing object detection in the respective small areas, according to various embodiments;

[0027] Figure 2B is an illustration of respective detectors having respective large regions, the detectors performing object detection in respective small regions, according to various embodiments;

[0028] Figure 3 is a flow chart illustrating a computer-implemented method for object detection according to various embodiments. DETAILED DESCRIPTION

[0029] Figure 1 An overview diagram 100 is depicted of the flow of data from radar through a neural network structure to a final output of object detection in accordance with various embodiments.

[0030] According to various embodiments, detection and classification of multiple objects may be performed directly from frequency domain radar sensor data, for example using machine learning methods. Multiple object detectors (e.g., N object detectors, for example, a first object detector 108, a second object detector 110, other object detectors 112 such as represented by dots, and an Nth object detector 114) may be provided. Multiple object detectors 108, 110, 112, 114 may each be tuned to a corresponding specific region of space, so that all of these detectors 108, 110, 112, 114 share a common input (e.g., a feature vector, for example including a specific scale pyramid 106) and a preprocessing chain.

[0031] Each of these detectors 108, 110, 112, 114 can determine an output (for example, the first detector 108 can determine a first output 116, the second detector 110 can determine a second output 118, other detectors 112 can determine other outputs 120 represented by dots, and the Nth detector 114 can determine an Nth output 122). In more detail, each of the detectors 108, 110, 112, 114 can locate an object within its associated area and can classify the object into one of multiple categories. In addition, each of these detectors 108, 110, 112, 114 can predict various attributes of the detected object, such as size, orientation, or speed. Example 124 and Example 126 show regions of interest with detection results of detector 108 and detector 118, respectively, and will be referred to below. Figure 2A and Figure 2B The other detectors 112 and the Nth detector may provide other detection results in the region of interest, as shown at point 128 .

[0032] A radar 102 (eg, a chirp radar) may provide radar data as input data to generate a feature vector (in other words: to compress into a feature vector) 104 .

[0033] Various embodiments may include a radar that includes an antenna array. Such a radar may provide signals corresponding to various possible combinations of distance and relative speed for each antenna. All of these signals from the different antennas may be combined for further processing, for example, by a neural network.

[0034] The signals received in the different antennas of the radar's array are slightly offset in time and phase. According to various embodiments, this information can be used to extract the direction of arrival of the signals reflected by the objects in the scene in order to infer their relative positions. In addition, the pattern of reflections can be used to predict the class of the object (in other words: identity; for example, indicating whether the object is a truck, car, bicycle or pedestrian) and its further properties (such as size and orientation).

[0035] According to various embodiments, a neural network method is provided that involves combining range information, velocity information, and angle information (e.g., based on phase shifts of signals received by multiple antennas) into a single hybrid representation (which may be referred to as a compressed data cube). An anchor-based neural network for object detection on a radar compressed data cube is provided.

[0036] In various available radar implementations, data is transmitted in a compressed and / or sparse format in order to overcome the limited bandwidth between the radar sensor and the central processing unit. However, processing based on sparse data may hinder machine learning applications because it may greatly increase the amount of training data required to effectively train the machine learning model. According to various embodiments, if a radar implementation with compressed and / or sparse data is used, a compression step can be applied as part of a neural network, which can transform the sparse signal into a dense signal while preserving its spatial order. For example, a convolution calculation can preserve the spatial order by applying a sliding window on the spatial dimension. The general order of the spatial dimension can remain unchanged. For example, with a step size (step size) of 2, a kernel [1 1] on the data [1 0 0 0 1 1] can output [1 0 2], which [1 0 2] can still preserve the spatial arrangement of the data.

[0037] The mixed dense representation of the radar signal can be stored as a feature vector (104). The feature vector can be gradually compressed to smaller and smaller spatial resolutions using a neural network, thereby generating a feature pyramid 106, where the compression is only applied along the distance dimension of the radar data. The spatial scale pyramid 106 (in other words: feature pyramid) can then be subsequently expanded back to larger and larger resolutions. The spatial scale pyramid can maintain different information resolutions within each pyramid level. Such resolution can improve object detector performance for processing objects of different sizes in an image. For example, deep pyramid levels can include information for large objects, while shallow pyramid levels can include information for small objects.

[0038] The compressed features in the increased spatial scale (in other words: the feature pyramid 106) may then be provided to a plurality of detectors 108, 110, 112, 114, each of which is capable of detecting objects within a specified region of space and determining various properties of the detected objects.

[0039] The designated areas may be different for different detectors 108, 110, 112, 114; in other words: each of the detectors 108, 110, 112, 114 may have its own designated area, which may be different from the designated areas of all other detectors. Although pairwise different, corresponding designated areas of different detectors 108, 110, 112, 114 may overlap. The corresponding designated areas may have different sizes, but for each size there may be more than one designated area (but with different locations). The corresponding designated areas may also be referred to as anchors.

[0040] Each detector 108, 110, 112, 114 is provided with data from multiple spatial scales (in a compressed pyramid, in other words: the spatial scale pyramid 106), and can attempt to detect objects in all scales. According to various embodiments, the detectors 108, 110, 112, 114 can be distributed on the polar coordinates of the scene, and the same group of detectors 108, 110, 112, 114 can be shared between multiple spatial ranges (in other words: the same group of detectors can be applied to data at multiple scales in the spatial scale pyramid 106). Although each detector 108, 110, 112, 114 can be optimized for objects of a specific scale, each detector 108, 110, 112, 114 can detect objects of any scale in the spatial scale pyramid 106.

[0041] According to various embodiments, a predetermined number of detectors 108 , 110 , 112 , 114 may be provided and these detectors 108 , 110 , 112 , 114 may be assigned to angular ranges of the spatial domain in which the object is to be determined.

[0042] When provided with an input radar signal (in the form of a spatial scale pyramid 106), each detector 108, 110, 112, 114 outputs the detector's confidence in whether an object is present within its assigned spatial region. In addition, each detector 108, 110, 112, 114 predicts the exact location of the object within the region as well as various properties of the object. The confidence of the detectors 108, 110, 112, 114 can be used to provide further processing, such as removing duplicate detections or object tracking.

[0043] The neural network performs processing according to various embodiments, including initial compression 104, determining data at multiple spatial scales (in other words: determining a spatial scale pyramid 106), and object detectors 108, 110, 112, 114 can be trained end-to-end based on labeled radar data. This end-to-end training can allow automatic learning of a complete object detection model without manually designing features, modules, or encoding.

[0044] While conventional assumptions are to estimate only a single point scatterer, the present method according to various embodiments uses the entire energy spectrum from antenna measurements without directly extracting any angular information, with the insight that the energy concentrated around the spectral peak contains useful information and can be a learnable feature for a neural network. In other words: according to various embodiments, the direction of arrival (DOA) is not limited to a single scattering point, but takes into account the spread of reflected energy from the object of interest, which eliminates the necessity to associate a single point (distance, angle) with the object. Each object or part of an object "seen" by the radar may be a complex pattern of radar energy, which may be a result of the shape and properties of the object. Therefore, the properties of the object can be reconstructed (or approximated) based on the pattern of energy received by the radar. Conventional radar processing systems extract the object position from the pattern in a pre-processing step of these radar processing systems, while discarding the pattern itself and can only detect the position of the object. According to various embodiments, the pattern (or a derivative generated during the compression and pyramid stages) can always be passed to the anchor-based detectors, allowing these detectors to reconstruct the properties of the object.

[0045] Figure 2A and Figure 2B 200 and 226 are shown of respective detectors with respective corresponding regions in which the detectors perform object detection at different levels of the distance of the scale pyramid according to various embodiments. It will be appreciated that Figure 2A and Figure 2B Only one level of the distance pyramid is shown in each, illustrating how different pyramid levels cover objects of various sizes.

[0046] Figure 2A The illustration 200 shows an illustration of a detector at a fine scale. For example, the overall region of interest for object detection is divided by several circular segments (in other words: arcs) 201, 203, 205, 207, 209, and detectors are arranged along the circular segments 201, 203, 205, 207, 209. At various distances (e.g., along various lines of the circular segments 201, 203, 205, 207, 209), multiple object detectors are trained to detect objects within the vicinity of these object detectors. Each arc 201, 203, 205, 207, 209 refers to a different distance in a single (in other words: in the same) pyramid level.

[0047] For the sake of illustration, only Figure 2AThe line of one of the circular segments 201, 203, 205, 207, 209 (along the circular segment 205) in FIG. 1 illustrates a detector. Each of the plurality of detectors may perform detection in a corresponding region (e.g., regions 204, 206, 208, 210, 212, 214, 216, and 218). It will be appreciated that although in FIG. Figure 2A The region is illustrated with a circle in , but the shape of the region (in other words: anchor) is not limited to a circular shape, but the region may have any shape other than a circle, for example, an ellipse, a rectangle, a square, or a general polygon.

[0048] There may be multiple objects to be detected in the region of interest (e.g., large object 220 and small object 222). It will be appreciated that although Figure 2A In FIG. 2 , objects 220 and 222 are illustrated by rectangular boxes, but the shape of the object is not limited to a rectangular shape, but the object may have any shape other than a rectangle. The shape of the object 220 and 222 may be unknown and may be one of the properties of the object that the detector should detect.

[0049] A feature vector 224 may be provided to each detector operating at a distance corresponding to arc 205, and the feature vector 224 may include data of the spatial scale pyramid 106. The detector operating at a distance corresponding to arc 205 may receive as input a feature vector 224, and the feature vector 224 may include data for a scale in the scale pyramid (corresponding to Figure 2A 201, 203, 205, 207, 209 shown in the figure) and a distance (corresponding to the distance of arc 205). Each detector can perform object detection based on feature vector 224 (in other words: based on at least one of the scale data in the spatial scale pyramid 106).

[0050] like Figure 2A As shown by the thick solid circle in , small object 222 may be detected by a detector performing object detection in region 212. Large object 220 may not be detected by any detector performing detection in small regions 204, 206, 208, 210, 212, 214, 216, and 218.

[0051] Figure 2B Example 226 shows that Figure 2A An illustration of detectors on a coarser scale than a fine scale is shown. The total region of interest for object detection is divided by several circular segments (in other words: arcs) 227, 229, 231, and detectors are arranged along the circular segments 227, 229, 231, wherein the distance between the circular segments 227, 229, 231 for the coarser scale is greater than that for the finer scale. Figure 2A201, 203, 205, 207, 209 shown in the figure. Each arc 227, 229, 231 refers to a different distance in a single (in other words: in the same) pyramid level. It will be understood that Figure 2B Compared with arcs 227, 229, and 231, Figure 2A The arcs 201, 203, 205, 207, 209 of are related to finer scales in the scale pyramid; in other words, Figure 2A The pyramid level ratio shown Figure 2B The pyramid levels shown are more refined; in other words, Figure 2B The pyramid level ratio shown Figure 2A The pyramid levels shown are coarser.At various distances (eg, along various lines of circular segments 227, 229, 231), multiple object detectors are trained to detect objects within the vicinity of these object detectors.

[0052] For the sake of illustration, only Figure 2B The detectors are illustrated by the circular segments 227 and 229 in FIG. A detector operating at a distance corresponding to arc 229 may receive as input a feature vector 242, which may include a feature vector for a scale in the scale pyramid (corresponding to Figure 2B The detector operating on the distance corresponding to arc 227 may receive as input a feature vector 244, which may include data for one scale in the scale pyramid (corresponding to Figure 2B The data of the roughness of the arcs 227, 229, 231 shown in FIG) and a distance (corresponding to the distance of the arc 227). Each of the plurality of detectors may perform detection in a corresponding region (e.g., regions 230, 232, 234, 236, 238, and 240). It will be understood that although Figure 2B The region is illustrated with a circle in , but the shape of the region (in other words: anchor) is not limited to a circular shape, but the region may have any shape other than a circle, for example, an ellipse, a rectangle, a square, or a general polygon.

[0053] like Figure 2B As shown by the thick solid circle in , large object 220 can be detected by the detector performing object detection in area 232, and large object 220 can also be detected by the detector performing object detection in area 238. Small object 222 may not be detected by any detector performing detection in large areas 230, 232, 234, 236, 238, and 240.

[0054] As Figure 2A and Figure 2BAs shown, according to various embodiments, sensor data is processed at multiple scales and using detectors on regions of different sizes, so that larger objects are detected at coarser scales. For example, Figure 2A Compared to Figure 2B The hierarchical scales shown correspond to the levels of the scale pyramid that are finer in scale. For example, Figure 2A The levels of the scale pyramid corresponding to the coarseness (in other words: fineness) of the arcs 201 , 203 , 205 , 207 , 209 are illustrated. Figure 2B The levels of the scale pyramid corresponding to the coarseness (in other words: fineness) of the arcs 227 , 229 , 231 are illustrated.

[0055] As described above, the feature pyramid according to various embodiments (in other words: corresponding multiple data for multiple spatial scales) may include distinctive features for each distance in each pyramid level. For example, each of the three exemplified feature vectors 224, 242, and 244 may be distinctive. For each pyramid level, all detectors (which may be referred to as anchors) in the same distance may receive the same feature vector describing the distance in the pyramid level. Detectors at the same pyramid level but at different distances may receive different feature vectors. Detectors in different pyramid levels may receive different feature vectors.

[0056] Figure 3 A flowchart 300 illustrating a computer-implemented method for object detection according to various embodiments is shown. At 302, signal representation data may be provided. The signal representation data may include distance information, velocity information, and angle information. At 304, for each of a plurality of spatial scales, corresponding scale data for the corresponding spatial scale may be determined based on the signal representation data to obtain a plurality of scale data. At 306, the plurality of scale data may be provided to a plurality of detectors. At 308, each detector may perform object detection based on at least one of the plurality of scale data.

[0057] According to various embodiments, the signal representative data may be or may be based on at least one of a radar signal or an ultrasonic signal. According to various embodiments, the signal representative data may include or may be frequency domain radar data. According to various embodiments, the signal representative data may be or may be based on a signal received by an antenna array.

[0058] According to various implementations, the signal representation data may be determined as dense data based on the sparse input data.

[0059] According to various embodiments, multiple spatial scales may be provided in a hierarchy from fine spatial resolution to coarse spatial resolution.

[0060] According to various embodiments, a plurality of spatial scales may be associated with distance information of the signal representation data.

[0061] According to various embodiments, each detector may perform object detection in a respective predetermined spatial region. According to various embodiments, the respective predetermined spatial regions of the plurality of detectors may be provided over a predetermined range in polar coordinates.

[0062] According to various embodiments, individual ones of the detectors may provide respective confidence levels for object detections performed.

[0063] According to various embodiments, each detector may predict the properties of the detected object when detecting the object. According to various embodiments, the properties of the detected object may include or may be the position of the detected object, and / or the size of the detected object, and / or the orientation of the detected object, and / or the speed of the detected object, and / or the shape of the detected object and / or the category of the detected object.

[0064] According to various embodiments, at least one of steps 302, 304, 306, 308 or any other step performed by the method according to various embodiments may be trained based on labeled data. The labeled data may include reference object information and data based on which reference signal representation data can be obtained.

[0065] Each of the above steps 302, 304, 306, 308 and other steps may be performed by computer hardware components.

[0066] Reference Numbers List

[0067] 100 shows an overview of the data flow 100

[0068] 102 Radar

[0069] 104 compressed into feature vector

[0070] 106 Space Scale Pyramid

[0071] 108 First Object Detector

[0072] 110 second object detector

[0073] 112 Other object detectors

[0074] 114 Nth Object Detector

[0075] 116 First Output

[0076] 118 Second output

[0077] 120 Other outputs

[0078] 122 Nth output

[0079] 124 Examples of Areas of Interest

[0080] 126 Examples of Areas of Interest

[0081] 128 Other exemplary points of the exemplary region of interest

[0082] 200 illustrative example of each detector having each corresponding small area

[0083] 201 Circular Segment

[0084] 203 circular segment

[0085] 204 small area

[0086] 205 circular segment

[0087] 206 small area

[0088] 207 Circular Segment

[0089] 208 small areas

[0090] 209 circular segment

[0091] 210 small area

[0092] 212 small area

[0093] 214 small area

[0094] 216 small areas

[0095] 218 small area

[0096] 220 large objects

[0097] 222 Small Objects

[0098] 224 eigenvectors

[0099] 226 Illustration of each detector having each corresponding large area

[0100] 227 Circular Segment

[0101] 229 circular segment

[0102] 230 major regions

[0103] 231 Circular Segment

[0104] 232 Large Areas

[0105] 234 large areas

[0106] 236 large areas

[0107] 238 large areas

[0108] 240 large areas

[0109] 242 eigenvectors

[0110] 244 eigenvectors

[0111] 300 Flowchart

[0112] 302 Methods and Steps

[0113] 304 Methods and Steps

[0114] 306 Methods and Steps

[0115] 308 method steps.

Claims

1. A computer-implemented method for object detection, The computer-implemented method comprises the following steps performed by computer hardware components: Providing signal representation data including distance information, speed information and angle information; For each spatial scale in the plurality of spatial scales, determining corresponding scale data of the corresponding spatial scale based on the signal representation data to obtain a plurality of scale data, wherein the plurality of scale data is in the form of a spatial scale pyramid; providing the plurality of scale data to a plurality of detectors, each detector being assigned a corresponding predetermined spatial region; as well as Each detector performs object detection in the corresponding predetermined spatial area based on at least one scale data among the plurality of scale data, wherein the respective predetermined spatial regions of the plurality of detectors overlap and have different sizes, Wherein each level in the spatial scale pyramid includes multiple feature vectors, and wherein, for each level, all detectors in the same distance receive the same feature vector describing the distance in the level, detectors at the same level but at different distances receive different feature vectors, and detectors at different levels receive different feature vectors.

2. The computer-implemented method according to claim 1, in, The signal representative data is based on at least one of a radar signal and an ultrasonic signal.

3. The computer-implemented method according to claim 1, in, The signal representative data includes frequency domain radar data.

4. A computer-implemented method according to any one of claims 1 to 3, in, The signal represents data based on signals received by the antenna array.

5. A computer-implemented method according to any one of claims 1 to 3, in, The signal representative data is determined as dense data based on the sparse input data.

6. The computer-implemented method of claim 5, in, The signal representation data is determined to be dense data based on a transformation that preserves the spatial order of the sparse input data.

7. A computer-implemented method according to any one of claims 1 to 3, in, The multiple spatial scales are provided in a hierarchy from fine spatial resolution to coarse spatial resolution.

8. A computer-implemented method according to any one of claims 1 to 3, in, The plurality of spatial scales are associated with the distance information of the signal representative data.

9. A computer-implemented method according to any one of claims 1 to 3, in, The respective predetermined spatial regions of the plurality of detectors are provided over a predetermined range in polar coordinates.

10. A computer-implemented method according to any one of claims 1 to 3, in, Each of the detectors provides a respective confidence level for the object detection performed.

11. A computer-implemented method according to any one of claims 1 to 3, in, Each detector, when detecting an object, predicts the properties of the detected object.

12. The computer-implemented method of claim 11, in, The property of the detected object includes at least one of a position of the detected object, a size of the detected object, an orientation of the detected object, a speed of the detected object, a shape of the detected object, or a category of the detected object.

13. A computer-implemented method according to any one of claims 1 to 3, in, At least one of the steps is performed using a neural network, the neural network being trained based on labeled data, the labeled data comprising reference object information and data based on which reference signal representative data can be obtained.

14. A computer system comprising a plurality of computer hardware components configured to perform the steps of the computer-implemented method according to any one of claims 1 to 13.

15. A non-transitory computer-readable medium comprising instructions for executing the computer-implemented method according to any one of claims 1 to 13.