Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N20/00—Machine learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F18/00—Pattern recognition
  • G06F18/20—Analysing
  • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
  • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
  • G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F9/00—Arrangements for program control, e.g. control units
  • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
  • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
  • G06F9/38—Concurrent instruction execution, e.g. pipeline or look ahead
  • G06F9/3885—Concurrent instruction execution, e.g. pipeline or look ahead using a plurality of independent parallel functional units
  • G06F9/3887—Concurrent instruction execution, e.g. pipeline or look ahead using a plurality of independent parallel functional units controlled by a single instruction for multiple data lanes [SIMD]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N5/00—Computing arrangements using knowledge-based models
  • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N5/00—Computing arrangements using knowledge-based models
  • G06N5/02—Knowledge representation; Symbolic representation
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V10/00—Arrangements for image or video recognition or understanding
  • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
  • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
  • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
  • G06V10/7747—Organisation of the process, e.g. bagging or boosting
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V10/00—Arrangements for image or video recognition or understanding
  • G06V10/94—Hardware or software architectures specially adapted for image or video understanding

Definitions

• AdaBoost adaptive boosting
• AdaBoost short for "adaptive boosting” is an algorithm for constructing a strong classifier as a linear combination of weak classifiers, such as decision trees.
• AdaBoost the output of the weak classifiers is combined into a weighted sum that represents the final output of the boosted classifier.
• AdaBoost is adaptive by tweaking subsequent weak learners in favor of instances misclassified by previous classifiers.
• AdaBoost (in which decision trees are used as the weak learners) is often referred to as the best out-of-the-box classifier and is a popular classifier for vision and data analytics.
• AdaBoost is located in Y. Fruend and R. Schapire, "A Decision-Theoretic Generalization of On-line Learning and an Application to Boosting," Journal of Computer and System Sciences, Vol. 55, Issue 1, August 1997, pp.119-139.
• the method includes: receiving feature vectors extracted from N consecutive window positions in an image in a memory coupled to the SIMD processor, in which N is a vector width of the SFMD processor divided by a bit size of a feature, and in which a feature vector includes N feature values, one feature value for each of the N consecutive window positions; and evaluating the N consecutive window positions concurrently by the AdaBoost classifier using the feature vectors and vector instructions of the SIMD processor, in which the AdaBoost classifier concurrently traverses decision trees for the N consecutive window positions until classification is complete for the N consecutive window positions, in which a decision tree includes nodes, a threshold value for each node, and leaves, each leaf including a partial score.
• AdaBoost adaptive boosting
• a digital system includes a single-instruction multiple-data (SIMD) processor, a memory component coupled to the SIMD processor, the memory component configured to store features extracted from an image, decision trees stored in the memory component, in which each decision tree includes nodes, a threshold value for each node, and leaves, each leaf including a partial score, and a decision tree based adaptive boosting (AdaBoost) classifier trained for object classification stored in the memory component, the AdaBoost classifier executable on the SIMD processor, in which the AdaBoost classifier uses the decision trees for object classification, the AdaBoost classifier configured to evaluate N consecutive window positions concurrently using the features and vector instructions of the SIMD processor, in which the AdaBoost classifier concurrently traverses decision trees for the N consecutive window positions until classification is complete for the N consecutive window positions and in which N is a vector width of the SIMD processor divided by a bit size of a feature.
• SIMD single-instruction multiple-data
• a non-transitory computer readable medium stores software instructions.
• the software instructions when executed on a single-instruction multiple-data (SIMD) processor, cause a method for object classification in a decision tree based adaptive boosting (AdaBoost) classifier to be executed.
• SIMD single-instruction multiple-data
• AdaBoost adaptive boosting
• the method includes receiving feature vectors extracted from N consecutive window positions in an image in a memory coupled to the SIMD processor, in which N is a vector width of the SIMD processor divided by a bit size of a feature, and in which a feature vector includes N feature values, one feature value for each of the N consecutive window positions, and evaluating the N consecutive window positions concurrently by the AdaBoost classifier using the feature vectors and vector instructions of the SIMD processor, in which the AdaBoost classifier concurrently traverses decision trees for the N consecutive window positions until classification is complete for the N consecutive window positions, in which a decision tree includes nodes, a threshold value for each node, and leaves, each leaf including a partial score.
• FIG. 1 is an example illustrating the sliding window approach for pedestrian detection in the scene of an image.
• FIG. 2 is an example illustrating the feature extraction for an image and the arrangement of the resulting features in memory for object detection.
• FIG. 3 is an example binary decision tree for an AdaBoost classifier.
• FIG. 4 is an example illustrating the general concept of a feature tuple in an AdaBoost classifier.
• FIG. 5 is an example illustrating the general concept of partial scores in the leaves of each decision tree of an AdaBoost classifier.
• FIG. 6 is an example illustrating feature vectors.
• FIG. 7 is a flow diagram of a method for executing an AdaBoost classifier on a single-instruction multiple-data (SIMD) processor.
• SIMD single-instruction multiple-data
• FIGS. 8-18 are examples.
• FIG. 19 is a simplified block diagram of an example digital system configured to execute an embodiment of the method of FIG.7.
• FIG. 20 is a block diagram of an example SIMD digital signal processor.
• an AdaBoost classifier may be constructed as a linear combination of weak classifiers such as decision trees.
• Embodiments of the disclosure are directed to decision tree based AdaBoost classifiers.
• embodiments are directed to implementing decision tree based AdaBoost classification on wide single-instruction multiple-data (SIMD) processors, also known as vector processors.
• SIMD wide single-instruction multiple-data
• Vector processors implement instructions that process multiple data points, i.e., vectors of data points, simultaneously. More specifically, multiple data points can be packed into one data word and a vector instruction can perform an operation on each data point simultaneously. For example, in a 128-bit vector processor, eight 16 bit pixels of an image can be packed into one 128-bit word and the eight pixels can be processed simultaneously.
• FIG. 1 is an example illustrating the sliding window approach for pedestrian detection in the scene of an image.
• a window also referred to as a box or an object model or an object patch
• features computed based on the pixel values in the window at each position
• the size of the window is based on the expected size of a pedestrian and is assumed to be 36x68.
• a window is examined at every fourth pixel vertically and horizontally. To identify different sizes of pedestrians, i.e., pedestrians at differing distances in a scene, the pedestrian detection is performed on multiple different scales of the image.
• example features may include gradient magnitude, gradient orientation, block sums, intensity, and color information.
• a conventional classifier may be used.
• computation of features for an image at multiple scales may be performed in a conventional manner.
• the example assumes the use of ten features computed using a cell size of 8x8: the components of the color space, i.e., Y (luma component), Cb (blue difference chroma component) and Cr (red difference chroma component), the gradient magnitude, and a histogram of gradients (HOG) for six bins between 0-180 degrees.
• 8* 16* 10 1280 features exist per window.
• FIG. 2 is an example illustrating the feature extraction for an image and the arrangement of the resulting features in memory for object detection.
• features for pedestrian detection are computed for the original image (base resolution) and for multiple scales of the image.
• a sliding window approach (as described hereinabove) is used to compute the features, resulting in set of ten feature channels for each window for each scale.
• a feature channel contains the computed values of a particular type of feature for a window. For this example, feature channels 0-5 are the bins of the HOG, channel 6 is the gradient magnitudes, and channels 7-9 are the respective color components.
• the feature channels of two sequential windows in a row of the image overlap in all but one position. For example, consider a window A at position x,y and the subsequent window B at position x+4, y. Each feature channel of window B is offset by one from the corresponding feature channel of window A. For example, if feature channel 0 of window A contains eight values, vl, v2, v8, then feature channel 0 of window B contains eight values v2, v8, v9. Similarly, feature channel 0 of window C at position x+8, y contains 8 values v3, ... , v8, v9, vlO.
• the feature data for each scale of the image is arranged in memory as shown in FIG. 2. For every fourth row of the image beginning with row 0, ten rows of feature data corresponding to the ten feature channels are "stacked up" in memory such that contiguous values in a row of feature data corresponding to a row of a scale can be loaded by a vector load instruction. Thus, ten rows of feature data for row 0 of a scale are stored, followed by ten rows of feature data for row 4, followed by ten rows of feature data for row 8, etc.
• the classifier is made up of 1280 binary two-level decision trees, each tree evaluating a feature at each node.
• a binary decision tree includes nodes and leaves with a binary split at each node as shown in the example of FIG. 3.
• the nodes are decision points in the tree and the leaves hold partial scores.
• a collective decision is formed based on partial scores resulting from traversals of multiple decision trees in a window.
• a feature value is compared against a threshold. The result of the comparison determines whether the left branch or the right branch out of a node is selected.
• the feature value input to the decision tree at each node will result in the selection of one of the four leaves.
• FIG. 4 is an example illustrating the general concept of a feature tuple.
• a feature tuple is a set of features mapped to a decision tree in the classifier. This mapping is formed during the training of the AdaBoost classifier. More specifically, the location in a window of each feature in a tuple is established. Different decision trees may have different tuples as inputs. The number of features in a tuple is equal to the number of nodes in a decision tree.
• FIG. 5 is an example illustrating the general concept of partial scores in the leaves of each decision tree.
• a feature tuple When a feature tuple is traversed through its respective decision tree, one of the leaves is selected, which contains a partial score for the particular feature tuple.
• the partial scores are summed and compared to a threshold, also referred to as a minima, exit threshold, or soft cascade threshold. If the minima is observed, the classification process may be terminated (exited) at that point or traversal of the remaining trees may continue to observe additional local minima, if any.
• the decision regarding exiting at any minima may be application specific. In the example classifier, tree traversal in a given window is terminated when the threshold is crossed during evaluation of the window.
• FIG. 6 is an example illustrating feature vectors assuming 32 sequential candidate object patches.
• the first ten rows (labeled 0-9) correspond respectively to the ten feature channels described hereinabove.
• the feature values for each node for each object patch are contiguous in memory, which enables a vector load of the 32 values for each node.
• FIG. 7 is a flow diagram of a method for executing a decision tree based AdaBoost classifier on a SIMD processor. For example, the method evaluates N consecutive windows (object patches) concurrently, where N is the SIMD processor vector width divided by the feature size. For purposes of explanation, the feature size is assumed to be 16 bits and N is assumed to be 32. Further, the classifier is assumed to be trained for pedestrian classification. For clarity of explanation, the method is described in reference to examples in FIGS. 8-18.
• the AdaBoost classifier is assumed to be constructed as a linear combination of two level binary decision trees. As described in reference to FIG. 3, and as shown in the example of FIG. 8, a two level binary decision tree has three nodes and four leaves. As shown in the example of FIG. 9, each node logically includes a threshold value T, offset value O, and a feature F. The value of a feature F is fetched from memory at an offset O to be compared against the threshold T. The threshold value, offset, and particular feature for each node of each tree are identified when the classifier is trained, as are the leaf values L for each tree.
• the tree parameters, i.e., the threshold values, the offsets, and the leaf values, of each tree of the AdaBoost classifier are stored linearly in memory as illustrated in FIG. 10.
• the extraction of features of an image that the AdaBoost classifier is to evaluate for the presence of pedestrians and how these features are stored in memory is described hereinabove.
• an image is searched for pedestrians using a sliding window approach in which the window size is based on the expected size of a pedestrian.
• the window positions searched are at an offset of one both vertically and horizontally. Further, the window positions are searched in raster scan order.
• a vector of features is loaded 600 for each node of the first decision tree of the classifier. Accordingly, three vectors of features are loaded, one for each node of the tree. The memory address of the vector for each feature is determined by the offset O for the feature in the tree. Due of the way the features are striped in memory, the offset of one between window positions, and the vector load capability of the SIMD processor, features for the 32 consecutive window positions are automatically loaded.
• FIG. 12 is an example illustrating this step.
• a threshold vector is then generated 702 for each node of the tree, i.e., three threshold vectors are generated.
• the threshold vector for a node is generated by replicating the threshold value for the node 32 times.
• Replication in a SIMD processor is an operation of reading one scalar value from memory and populating all elements of a vector with this scalar value.
• FIG. 13 is an example illustrating the threshold vectors and the comparison.
• the naming convention used in this and other figures is: F ⁇ node> ⁇ position>, where node is the node number of the tree and position is the relative window position, e.g. , F23 is the feature evaluated at node 2 of the tree for window position 3.
• FIG. 14 is an example illustrating the mask vectors resulting from the vector compare operations. In this example, the mask value corresponding to the comparison of each feature to the corresponding threshold is 0x0000 if the comparison is false and OxFFFF if the comparison is true. Also, other mask values are useful to indicate the results of the compare operations.
• the three mask vectors are then used to select a partial score value, i.e. , leaf value, for the tree traversal for each of the 32 object patches and to generate 706 a vector of partial scores in which each entry is the resulting partial score (leaf value) from the tree traversal for a corresponding object patch.
• the generation of the vector of partial score values is performed as illustrated in the examples of FIGS. 15-17.
• the three mask vectors Ml, M2, and M3, are used to generate four leaf selection mask vectors Kl, K2, K3, K4, one for each of the four leaf values of the tree. These leaf selection mask vectors are then used to select leaf values from four leaf vectors, LVl, LV2, LV3, LV4.
• the leaf vectors for the leaves of the tree which are shown in FIG. 16, are generated by replicating each leaf value LI, L2, L3, and L4 in the corresponding vector 32 times.
• the mask vectors Ml, M2, M3 are logically combined as illustrated in FIG. 15 to generate the leaf selection mask vectors Kl, K2, K3, K4.
• the logic is as follows: 1) when corresponding locations in Ml and M2 are true, then set the corresponding location in leaf selection mask vector Kl to select the corresponding location in leaf vector LV1; 2) when a location in Ml is true and the corresponding location in M2 is false, then set the corresponding location in leaf selection mask vector K2 to select the corresponding location in leaf vector LV2; 3) when a location in Ml is false and the corresponding location in M3 is false, then set the corresponding location in leaf selection mask vector K3 to select the corresponding location in leaf vector LV3; and 4) when corresponding locations in Ml and M3 are false, then set the corresponding location in leaf selection mask vector K4 to select the corresponding location in leaf vector LV4.
• leaf selection mask vectors not set to select a leaf value by the logical mask combinations are set to not select a leaf value.
• a selection value in a leaf selection mask vector is 0x0000 if the corresponding leaf value in the corresponding leaf vector is not to be selected and to OxFFFF if the corresponding leaf value is to be selected.
• other mask values are useful.
• a leaf selection mask vector is a logical combination of the mask vectors for the nodes in the traversal path of the decision tree that reaches the corresponding leaf.
• the resulting leaf selection mask vectors Kl, K2, K3, K4 are illustrated in FIG. 16.
• a logical and operation is performed between each leaf vector and the corresponding leaf selection mask vector to select leaf values (partial scores) from each leaf vector.
• the result of these four logical and operations is four vectors of leaf values as illustrated in FIG. 17.
• the four vectors are combined with logical or operations to generate the partial score vector in which each entry is the result of evaluating the decision tree for corresponding ones of the 32 window positions.
• the partial score vector is accumulated 708 into an accumulated score vector having one entry for each of the 32 windows.
• the partial score vector may be added to the accumulated score vector, which stores the sum of any previous partial scores from previous tree traversals.
• an exit mask vector may be applied to the partial score vector before accumulating the partial scores to mask out partial scores for window positions that have met the criteria to terminate classification.
• the accumulated score vector is compared 710 to an exit threshold vector and any accumulated partial score values below the exit threshold are saved as final scores. If an accumulated partial score value for a given window position is below the exit threshold, then tree evaluation, i.e., classification, for that window is complete. If the classification process for all 32 windows is complete 712, i.e., all accumulated scores are below the exit threshold, or all trees have been traversed 714, then the classification process for the 32 windows is terminated and the final accumulated score vector is returned 716. Otherwise, the classification continues 700 with the next tree in the classifier. The order in which the trees are traversed is determined during the training process.
• an exit mask vector is maintained that indicates which of the window positions has completed the classification process and which have not.
• the generation and use of the exit mask vector is illustrated in the example of FIG. 18.
• the vector comparison operation of the accumulated score vector to the exit threshold vector results in an exit mask vector that indicates which of the accumulated scores meets the exit criteria and which do not.
• a logical and operation of the inverse of the exit mask vector (" ⁇ " is bit invert) and the partial score vector is performed to mask out partial scores for any window positions that have previously exited.
• the resulting partial score vector is then added to the accumulated score vector, the comparison to the exit threshold vector is performed, and an updated exit mask is generated.
• FIG. 19 is a simplified block diagram of an example digital system 1900 configured to execute an embodiment of the method of FIG. 7.
• the digital system may be an integrated circuit, i.e., a system-on-a-chip.
• the digital system 1900 includes a master processor 1902, a camera 1904, an image signal processor (ISP) 1906, a feature extraction component 1908, a SEVID instruction set digital signal processor (DSP) 1910, and a shared memory 1912.
• the master processor 1902 controls the operation of the other components to perform operations needed for pedestrian classification in scenes captured by the camera 1904.
• the master processor 1902 may be any suitable processor, such as central processing units available from ARM Ltd.
• the camera 1904 captures images of a scene and provides those images to the ISP 1906.
• the ISP 1906 performs image processing on each image to prepare the image for feature extraction.
• the ISP 1906 may perform operations on the images, such as white balancing, black level adjustment, noise filtering, conversion from RGB to YCbCr, and edge enhancement.
• the feature extraction component 1908 performs feature extraction on images from the ISP 1906. Feature extraction is described hereinabove. The extracted features are stored in shared memory 1912 for use in the method.
• the shared memory component 1912 may be on-chip memory, external memory, or a combination thereof. Any suitable memory design may be used.
• the memory component 1912 may include static random access memory (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), read-only memory (ROM), flash memory, or a combination thereof.
• SRAM static random access memory
• DRAM dynamic random access memory
• SDRAM synchronous DRAM
• ROM read-only memory
• flash memory or a combination thereof.
• the memory component 1912 stores software instructions for the AdaBoost classifier 1916 that include software instructions to perform an embodiment of the method of FIG. 6.
• the memory component also stores the features 1914 computed by the feature extraction component 1908, and the decision trees 1918 used by the classifier 1916.
• Some or all of the software instructions and decisions trees may be initially stored in a computer-readable medium such as a compact disc (CD), a diskette, a tape, a file, memory, or any other computer readable storage device and loaded and stored on the digital system 1900.
• the software instructions may also be sold in a computer program product, which includes the computer-readable medium and packaging materials for the computer-readable medium.
• the software instructions may be distributed to the digital system 1900 via removable computer readable media (e.g., floppy disk, optical disk, flash memory, USB key), via a transmission path from computer readable media on another computer system (e.g., a server).
• removable computer readable media e.g., floppy disk, optical disk, flash memory, USB key
• the DSP 1910 executes the software instructions of the classifier 1916 to perform pedestrian classification using the extracted features 1914.
• the DSP implements a SFMD instruction set providing at least vector load operations, vector compare operations, vector addition and subtraction operations, vector logical operations, and replication operations. Any suitable DSP with an appropriate SIMD instruction set may be used. One such DSP is described in reference to FIG. 20.
• the results of the classification are communicated to the master processor 1902 for further processing, such as pedestrian detection based on the classification results and decision making based on the results of the pedestrian detection.
• FIG. 20 is a high level block diagram of an example SIMD digital signal processor (DSP) 2000 suitable for executing one or more embodiments of the method of FIG. 7.
• DSP SIMD digital signal processor
• the illustrated DSP is the TMS32C66x DSP available from Texas Instruments, Inc.
• the C66x DSP 2000 includes eight functional units, two register files, and two data paths.
• the two general-purpose register files each contain thirty-two 32-bit registers for a total of 64 registers.
• the general-purpose registers are useful for data or can be data address pointers.
• the data types supported include packed 8-bit data, packed 16-bit data, 32-bit data, 40-bit data, and 64-bit data.
• the C66x DSP 2000 supports up to 4 way SEVID operations for 16 bit data and up to 8 way SEVID operations for 8 bit data.
• the SEVID width for each data path is 64 bits, other than for some multiply operations which can handle up to 128 bits of packed data.
• a detailed description of the C66x and instruction set is located in "TMS320C66x DSP CPU and Instruction Set Reference Guide," SPRUGH7, November 2010, which is incorporated by reference herein.
• the method of FIG. 7 evaluates N windows (object patches) concurrently, where N is the SIMD width divided by the feature size.
• N the SIMD width divided by the feature size.
• the decision trees are assumed to be two level binary decision trees.
• the decision trees may include more than two levels and/or are not required to be binary.
• Components may be referred to by different names and/or may be combined in ways not shown herein without departing from the described functionality. Also, for example, if a first device couples to a second device, that connection may be through a direct electrical connection, through an indirect electrical connection via other devices and connections, through an optical electrical connection, and/or through a wireless electrical connection.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Software Systems (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• Evolutionary Computation (AREA)
• Artificial Intelligence (AREA)
• Data Mining & Analysis (AREA)
• Computing Systems (AREA)
• Mathematical Physics (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Medical Informatics (AREA)
• Multimedia (AREA)
• Computational Linguistics (AREA)
• Health & Medical Sciences (AREA)
• Databases & Information Systems (AREA)
• General Health & Medical Sciences (AREA)
• Life Sciences & Earth Sciences (AREA)
• Bioinformatics & Cheminformatics (AREA)
• Bioinformatics & Computational Biology (AREA)
• Evolutionary Biology (AREA)
• Image Analysis (AREA)
• Image Processing (AREA)

Priority Applications (4)

Applications Claiming Priority (2)

Publications (1)

Family

ID=57686191

Family Applications (1)

Country Status (5)

Cited By (1)

Families Citing this family (20)

Citations (3)

Family Cites Families (10)

• 2015
  • 2015-07-06 US US14/792,596 patent/US10325204B2/en active Active
• 2016
  • 2016-07-06 WO PCT/US2016/041159 patent/WO2017007831A1/en not_active Ceased
  • 2016-07-06 CN CN201680039290.8A patent/CN107851197B/zh active Active
  • 2016-07-06 EP EP16821919.4A patent/EP3320488B1/en active Active
  • 2016-07-06 JP JP2018500704A patent/JP7242975B2/ja active Active
• 2019
  • 2019-04-22 US US16/390,082 patent/US10977560B2/en active Active
• 2022
  • 2022-11-02 JP JP2022176623A patent/JP2023022031A/ja active Pending

Patent Citations (3)

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events