WO2019176877A1 - オブジェクト識別システム、自動車、車両用灯具、オブジェクトの種類の識別方法 - Google Patents
オブジェクト識別システム、自動車、車両用灯具、オブジェクトの種類の識別方法 Download PDFInfo
- Publication number
- WO2019176877A1 WO2019176877A1 PCT/JP2019/009781 JP2019009781W WO2019176877A1 WO 2019176877 A1 WO2019176877 A1 WO 2019176877A1 JP 2019009781 W JP2019009781 W JP 2019009781W WO 2019176877 A1 WO2019176877 A1 WO 2019176877A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- intermediate data
- neural network
- identification system
- line data
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/88—Lidar systems specially adapted for specific applications
- G01S17/89—Lidar systems specially adapted for specific applications for mapping or imaging
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/02—Systems using the reflection of electromagnetic waves other than radio waves
- G01S17/06—Systems determining position data of a target
- G01S17/42—Simultaneous measurement of distance and other co-ordinates
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/88—Lidar systems specially adapted for specific applications
- G01S17/89—Lidar systems specially adapted for specific applications for mapping or imaging
- G01S17/894—3D imaging with simultaneous measurement of time-of-flight at a 2D array of receiver pixels, e.g. time-of-flight cameras or flash lidar
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/88—Lidar systems specially adapted for specific applications
- G01S17/93—Lidar systems specially adapted for specific applications for anti-collision purposes
- G01S17/931—Lidar systems specially adapted for specific applications for anti-collision purposes of land vehicles
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/48—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S17/00
- G01S7/4802—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S17/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/64—Three-dimensional objects
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/16—Anti-collision systems
- G08G1/165—Anti-collision systems for passive traffic, e.g. including static obstacles, trees
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/16—Anti-collision systems
- G08G1/166—Anti-collision systems for active traffic, e.g. moving vehicles, pedestrians, bikes
Definitions
- the present invention relates to an object identification system.
- LiDAR Light Detection and Ranging, Laser Imaging and Detection
- cameras millimeter wave radar, and ultrasonic sonar.
- LiDAR is capable of (i) object recognition based on point cloud data, and (ii) high-precision detection even in bad weather due to active sensing, compared to other sensors.
- Iii) has an advantage that a wide range of measurement is possible, and is expected to become the mainstream in the sensing system of automobiles in the future.
- the object identification based on the point cloud data generated by LiDAR is more accurate as the resolution of the point cloud data is higher, but the cost of arithmetic processing increases explosively. In consideration of mounting on a vehicle, it may be necessary to use a low-end low-end arithmetic processing unit, and it is necessary to reduce the number of scan lines.
- the present invention has been made in such a situation, and one of exemplary purposes of an aspect thereof is to provide a system, an apparatus, and a method capable of identifying an object with a small number of horizontal lines.
- An aspect of the present invention relates to an object identification system.
- An object identification system includes: a three-dimensional sensor that generates a plurality of line data for a plurality of horizontal lines having different heights; an arithmetic processing device that identifies an object type (category or class) based on the plurality of line data; Is provided.
- Each of the arithmetic processing units generates first intermediate data related to one corresponding one of the plurality of line data, and the first intermediate data has a probability that the corresponding line data corresponds to each of a plurality of parts of a plurality of types (
- a plurality of first neural networks and a plurality of first intermediate data corresponding to a plurality of line data, and the plurality of first intermediate data are combined to form at least one second intermediate data.
- a second neural network that receives at least one second intermediate data and generates final data indicating the probability that the object corresponds to each of a plurality of types.
- the automobile may include the object identification system described above.
- the 3D sensor may be built in the headlamp.
- the vehicular lamp may include the above-described object identification system.
- an object can be identified with a small number of horizontal lines.
- FIGS. 3A to 3D are diagrams showing a plurality of line data when a pedestrian, a bicycle, a car, and a utility pole are photographed with a three-dimensional sensor. It is a block diagram which shows the structural example of a 1st neural network. It is a block diagram which shows the structural example of a 2nd neural network. 6A to 6C are diagrams for explaining object extraction. 7A to 7C are diagrams for explaining the first learning method.
- FIGS. 8A and 8B are diagrams for explaining shooting of a pedestrian.
- FIGS. 13A and 13B are diagrams showing the relationship between the height of LiDAR and the objects.
- FIG. 13C shows the relationship between the object learning system learned by the first learning method and FIG. It is a figure which shows the last data obtained in the condition of (). It is a figure explaining the learning process of the 2nd arithmetic unit in the 2nd learning method. It is a figure explaining the effect of the 2nd learning method.
- An object identification system includes: a three-dimensional sensor that generates a plurality of line data for a plurality of horizontal lines having different heights; an arithmetic processing device that identifies an object type (category or class) based on the plurality of line data; Is provided.
- Each of the arithmetic processing units generates first intermediate data related to one corresponding one of the plurality of line data, and the first intermediate data has a probability that the corresponding line data corresponds to each of a plurality of parts of a plurality of types (
- a plurality of first neural networks and a plurality of first intermediate data corresponding to a plurality of line data, and the plurality of first intermediate data are combined to form at least one second intermediate data.
- a second neural network that receives at least one second intermediate data and generates final data indicating the probability that the object corresponds to each of a plurality of types.
- the type of object can be determined with a small number of horizontal lines. Further, by combining a plurality of first intermediate data, the dependency in the height direction can be reduced. Thereby, the restrictions of installation of a three-dimensional sensor can be eased. It should be noted that the height information is not completely lost by the combining process, and the site continues to hold the height information.
- the at least one second intermediate data is one, and the second intermediate data may be obtained based on all of the plurality of first intermediate data.
- each second intermediate data may be obtained based on some consecutive ones of the plurality of first intermediate data.
- Each of the at least one second intermediate data may be an average or a sum of some corresponding first intermediate data.
- the average may be a simple average or a weighted average.
- Each of the at least one second intermediate data may be a maximum value of some corresponding first intermediate data.
- a step of causing the first neural network to learn using a plurality of line data obtained by measuring each of a plurality of types of parts, and a combined processing unit for outputting the outputs of the plurality of learned first neural networks A step of allowing the second neural network to learn in a state of being coupled to the second neural network via.
- the arithmetic processing unit may perform normalization that divides a value included in each line data by a predetermined value as preprocessing.
- the types of objects may include at least pedestrians, bicycle riders, and cars.
- FIG. 1 is a block diagram of an object identification system 10 according to the first exemplary embodiment.
- the object identification system 10 is mounted on a vehicle such as an automobile or a motorcycle, and determines the type (also referred to as category or class) of an object OBJ existing around the vehicle.
- the object identification system 10 mainly includes a three-dimensional sensor 20 and an arithmetic processing device 40.
- the three-dimensional sensor 20 generates a plurality of line data LD 1 to LD N for a plurality of horizontal lines L 1 to L N having different heights.
- the number N of horizontal lines is not particularly limited, but 20 or less and about 4 to 12 are preferable.
- Each line data LD includes distance information to a plurality of sampling points P along the corresponding horizontal line L.
- a set of a plurality of line data LD 1 to LD N is referred to as distance measurement data.
- the three-dimensional sensor 20 is not particularly limited, it is preferable to use LiDAR when it is desired to accurately identify an object with small unevenness such as a pedestrian.
- the number N of horizontal lines is the resolution in the vertical direction.
- the configuration of LiDAR is not particularly limited, and may be a scanning type or a non-scanning type.
- the arithmetic processing unit 40 identifies the type (category) of the object based on the distance measurement data including the plurality of line data LD 1 to LD N.
- the arithmetic processing unit 40 is configured to process data including one object, and when one piece of distance measurement data includes a plurality of objects, the arithmetic processing unit 40 divides the data into subframes including one object by preprocessing.
- the arithmetic processing unit 40 uses subframes as processing units.
- the arithmetic processing unit 40 can be implemented by a combination of a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a processor (hardware) such as a microcomputer, and a software program executed by the processor (hardware).
- the arithmetic processing unit 40 may be a combination of a plurality of processors.
- examples of object types include pedestrians, bicycles, cars, and utility poles.
- the object OBJ is defined with a plurality of parts (referred to as categories or subcategories) for different heights.
- FIG. 2 is a diagram illustrating an example of a plurality of parts defined for a pedestrian.
- M sites H 0 to H M-1 are defined.
- H 0 is the knee
- H 1 is above the knee
- H 2 is the thigh
- H 3 is the waist
- H 4 is the abdomen
- H 5 is the chest
- H 6 is the shoulder
- H 7 is the face.
- a plurality of parts B 0 to B 7 having different heights are also defined for the bicycle.
- a plurality of portions C 0 to C 7 having different heights are also defined for the automobile.
- a plurality of parts P 0 to P 7 having different heights can be defined, but it is not necessary to distinguish them because the profile is substantially invariant regardless of the height, and therefore one output P 0 To summarize.
- FIGS. 3A to 3D are diagrams showing a plurality of line data when a pedestrian, a bicycle, a car, and a utility pole are photographed by the three-dimensional sensor 20.
- the plurality of line data indicate the shapes of a plurality of predefined parts.
- the arithmetic processing unit 40 generates intermediate data MD related to the type of the object OBJ and its part for each line data LD.
- the intermediate data MD i may statistically indicate which type and which part the corresponding line data LD i (horizontal line L i ) is.
- the processor 40 integrates the plurality of intermediate data MD 1 ⁇ MD N corresponding to a plurality of line data LD 1 ⁇ LD N, to produce a final data FD indicating the type of object OBJ.
- the final data FD may indicate statistically which type the object OBJ is.
- the arithmetic processing unit 40 functionally includes a plurality of first arithmetic units 42_1 to 42_N and a second arithmetic unit 44.
- the blocks indicated by the arithmetic units 42 and 44 do not necessarily mean that they are independent in hardware.
- the arithmetic processing unit 40 is configured with a single core, the plurality of arithmetic units 42 and 44 can correspond to a single core.
- each core can function as a plurality of arithmetic units 42 and 44.
- the i-th (1 ⁇ i ⁇ N) arithmetic unit 42 — i processes the corresponding line data LD i to generate intermediate data MD i .
- the second arithmetic unit 44 integrates the intermediate data MD 1 ⁇ MD N of the plurality of first arithmetic units 42_1 ⁇ 42_N generated, to produce the final data FD.
- the above is the basic configuration of the object identification system 10.
- the implementation of the arithmetic processing unit 40 is not particularly limited, but can be configured using, for example, a neural network.
- the configuration verified by the inventor will be described below.
- the neural network corresponding to the first arithmetic unit 42 is referred to as a first neural network NN 1
- the neural network corresponding to the second arithmetic unit 44 is referred to as a second neural network NN 2 .
- FIG. 4 is a block diagram showing a first configuration example of the neural network NN 1.
- the first neural network NN 1 includes an input layer 50, three intermediate layers (hidden layers) 52, and an output layer 54.
- the number of units of the input layer 50 is determined according to the number of sample points in one line and is 5200. There are three intermediate layers, and the number of units is 200, 100, 50.
- affine transformation and transformation using a sigmoid function are performed in the output layer 54.
- the affine transformation and the calculation of the affiliation probability using the softmax function are performed.
- the intermediate data MD i includes a plurality of data Human-0th to Human-7th, Car-0th to Car-7th, Bicycle--0th to Bicycle7th, Pole-all, and parts H 0 to H 7 related to pedestrians, about cars
- the probabilities corresponding to the part C 0 to C 7 , the part B 0 to B 7 related to the bicycle, and the part P 0 related to the utility pole are shown.
- FIG. 5 is a block diagram showing a second configuration example of the neural network NN 2.
- the intermediate layer 62 is one layer, and the number of units is 50.
- the output layer 64 four categories of a pedestrian (Human), a car (Car), a bicycle (Bicycle), and a utility pole (Pole) are set. That is, the final data FD includes four data Human, Car, Bicycle, and Pole that indicate the possibility that the object OBJ corresponds to each of a pedestrian, a car, a bicycle, and a utility pole.
- the parameter update method is Adam
- the learning coefficient is 0.01
- the number of iterations is 20,000.
- Extraction is a process of removing the background and extracting the object OBJ.
- 6A to 6C are diagrams for explaining object extraction.
- FIG. 6A is a diagram illustrating an automobile that is an object.
- FIG. 6B shows a plurality of line data LD 1 to LD 8 when the object of FIG. 6A is photographed with LiDAR.
- FIG. 6C shows line data LD 1 to LD 8 extracted so as to include the object.
- Shift is a process of shifting data so that the object is positioned at the center.
- Normalization is a process of dividing distance data by a predetermined value.
- the predetermined value may be a distance (reference distance) between the three-dimensional sensor 20 and a predetermined portion of the object OBJ at the time of learning. Thereby, the line data is normalized to a value near 1.
- FIGS. 7A to 7C are diagrams illustrating the first learning method.
- data learning data or teacher data
- Learning data is acquired by measuring a plurality of objects with LiDAR. Specifically, you can measure candidate objects (pedestrians, cars, utility poles, people on bicycles, etc.) you want to identify under different conditions (for example, from different distances, from different directions)
- Data FD 1 , FD 2 ,... are prepared.
- a plurality of valid line data LD ij included in the plurality of frame data is individually input to the first arithmetic unit 42 together with the teacher data TD ij .
- the learning result obtained for one first arithmetic unit 42 is used in all the first arithmetic units 42.
- the second arithmetic unit 44 is learned. Specifically, as shown in FIG. 7B, a plurality of first arithmetic units 42 and a second arithmetic unit (second neural network) 44 are connected. In this state, a plurality of frame data FD 1 , FD 2 ... Are individually input to the arithmetic processing unit 40. For each frame FD i , a plurality of sets of intermediate data MD 1 to MD 8 are generated by the plurality of first arithmetic units 42 and supplied to the second arithmetic unit 44 at the subsequent stage.
- the number of horizontal lines of LiDAR used for verification is 8.
- the irradiation angle of the horizontal line (vertical angle resolution) is -18.25 °, -15.42 °, -12.49 °, -9.46 °, -6.36 °, -3.19 ° from the bottom. , 0 °, 3.2 °.
- FIGS. 8A and 8B are diagrams for explaining shooting of a pedestrian.
- the distance (reference distance) of the object OBJ from the center of LiDAR was 3 m.
- the pedestrian sample is an adult male with a height of 166 cm and has nine directions (0 °, 22.5 °, 45 °, 67.5 °, 90 °, 112.5 °, 135 °, 157.5 °, (0 °, 180 °)
- the direction in which the front (face, headlamp) of the object OBJ can be seen is 0 °.
- the LiDAR is tilted by 7 ° (elevation angle) in the vertical direction so that the eight horizontal lines coincide with the eight portions H 0 to H 7 shown in FIG.
- the photograph was taken with a person straddling and standing still.
- the shooting direction was set to 9 as with the pedestrian.
- the object identification system 10 it is possible to determine the type of an object with a very high probability with the number of only eight horizontal lines.
- the distance between the object and LiDAR is fixed at 3 m, but the distance is actually variable. Therefore, the distance may be divided into a plurality of ranges and the neural network may be learned for each range.
- FIG. 11 is a block diagram of an automobile provided with the object identification system 10.
- the automobile 100 includes headlamps 102L and 102R.
- the object identification system 10 at least the three-dimensional sensor 20 is built in at least one of the headlamps 102L and 102R.
- the headlamp 102 is located at the foremost end of the vehicle body, and is most advantageous as an installation location of the three-dimensional sensor 20 in detecting surrounding objects.
- the arithmetic processing unit 40 may be built in the headlamp 102 or provided on the vehicle side. For example, in the arithmetic processing unit 40, intermediate data may be generated inside the headlamp 102, and final data generation may be left to the vehicle side.
- FIG. 12 is a block diagram showing a vehicular lamp 200 including the object identification system 10.
- the vehicular lamp 200 includes a light source 202, a lighting circuit 204, and an optical system 206. Further, the vehicular lamp 200 is provided with a three-dimensional sensor 20 and an arithmetic processing unit 40. Information regarding the object OBJ detected by the arithmetic processing unit 40 is transmitted to the vehicle ECU 104. The vehicle ECU may perform automatic driving based on this information.
- the information regarding the object OBJ detected by the arithmetic processing device 40 may be used for light distribution control of the vehicular lamp 200.
- the lamp ECU 208 generates an appropriate light distribution pattern based on information regarding the type and position of the object OBJ generated by the arithmetic processing device 40.
- the lighting circuit 204 and the optical system 206 operate so that the light distribution pattern generated by the lamp ECU 208 is obtained.
- FIGS. 13A and 13B are diagrams illustrating the relationship between the height of LiDAR and the objects.
- FIG. 13A shows a case where the installation height of LiDAR is 145 cm during learning. At this time, the lower three line data are invalid, and learning is performed using the fourth to eighth line data LD 4 to LD 8 from the bottom.
- the installation height of LiDAR is 70 cm lower than that during learning. It is also assumed that the distance between the pedestrian and LiDAR is closer than at the time of learning.
- the line data LD 1 to LD 3 has no corresponding part (no subcategory)
- FIG. 13C shows the final data obtained in the situation of FIG. 13B by the object identification system 10 that has been learned by the first learning method. Where it should be recognized as a pedestrian, it is misrecognized as having a higher probability of being in another category. This is presumably due to the fact that the classification in the second arithmetic unit 44 strongly depends on the order and combination of subcategories in the first learning method. That is, when the first learning method is adopted, the height of the three-dimensional sensor in the actual use stage may be restricted by the height used in the learning stage.
- the second learning method incorporates a device for reducing such restrictions.
- FIG. 14 is a diagram illustrating the learning process of the second arithmetic unit 44 in the second learning method. Specifically, in the second learning method, while changing the correspondence 46 of the plurality of learned first computation units 42 and the plurality of input nodes I 1 to I 8 of the second computation unit 44, The second arithmetic unit 44 is learned.
- the correspondence relationship 46 may be randomly changed for each frame data FD i .
- the learning may be performed by switching the correspondence relationship with a plurality of patterns for one frame data FD i .
- N 8
- ⁇ 7 56 combinations of input and output. Therefore, learning may be performed in all combinations for each frame data.
- FIG. 15 is a diagram for explaining the effect of the second learning method.
- FIG. 15 shows final data FD when a pedestrian is assumed in the situation of FIG. 13B using the object identification system 10 learned by the second learning method. Compared to the final data of FIG. 13C obtained as a result of the first learning method, the probability of recognizing a pedestrian can be increased.
- the learning process of the second arithmetic unit 44 by changing the correspondence relationship between the plurality of first arithmetic units 42 and the plurality of inputs of the second arithmetic unit 44, the installation of a three-dimensional sensor such as LiDAR can be performed.
- the degree of freedom can be increased.
- FIG. 16 is a block diagram of an object identification system 10A according to a modification.
- the first arithmetic unit 42A is implemented as a convolutional neural network.
- a convolutional neural network targets an M ⁇ N pixel two-dimensional image, but this embodiment is new in that it uses one-dimensional line data as a target.
- a convolutional neural network is a combination of a convolutional layer and a pooling layer. By using the convolutional neural network, it is possible to improve the robustness against the positional deviation of the object in the horizontal direction.
- the discriminating power is improved by devising a learning method.
- improvement in discrimination power is realized by devising the configuration of the arithmetic processing unit.
- FIG. 17 is a block diagram of an object identification system 10B according to the second embodiment.
- the object identification system 10B includes a three-dimensional sensor 20 and an arithmetic processing device 70.
- the arithmetic processing unit 70 includes a plurality of first neural networks 72_1 to 72_N, a joint processing unit 74, and a second neural network 76.
- the first intermediate data MD 1_I the corresponding line data LD i is illustrates a probability corresponding to each of the plurality of sites of a plurality of kinds (categories) (subcategories).
- the combination processing unit 74 receives a plurality of first intermediate data MD 1_1 to MD 1_N corresponding to the plurality of line data LD 1 to LD N , combines them, and generates at least one second intermediate data MD 2 .
- the second intermediate data MD 2 is one. That is, all of the first intermediate data MD 1_1 ⁇ MD 1_N is coupled to the second intermediate data MD 2 one. Similar to the first intermediate data MD 1 , the second intermediate data MD 2 indicates the probability that the corresponding line data LD i corresponds to each of a plurality of parts (subcategories) of a plurality of types (categories).
- FIG. 18 is a diagram illustrating an example of processing of the combination processing unit 74.
- the i-th first intermediate data MD 1_i includes K elements a 1i to a Ki .
- Coupling processor 74 an average of the plurality of first intermediate data MD 1_1 ⁇ MD 1_N, the second intermediate data MD 2.
- the j-th element b j of the second intermediate data MD 2 is given by the following equation.
- Second neural network 76 the second receiving the intermediate data MD 2, object OBJ to generate the final data FD indicating the probability corresponding to each of the plurality of types (categories).
- the second neural network 76 can be configured similarly to the neural network of FIG.
- the object identification system 10B is learned using the first learning method described above. That is, the first neural network 72 is trained using a plurality of line data obtained by measuring each of a plurality of types of parts. A common learning result is applied to all the first neural networks 72_1 to 72_N.
- the second neural network 76 is trained in a state where the outputs of the plurality of learned first neural networks 72_1 to 72_N are coupled to the second neural network 76 via the coupling processing unit 74.
- FIG. 19 is a diagram showing final data obtained in the situation of FIG. 13B by the object identification system in the second embodiment. According to the second embodiment, the probability of recognizing a pedestrian can be increased as compared to the final data of FIG. 13C obtained by a combination of the first embodiment and the first learning method. .
- the sum may be taken in the combination processing unit 74.
- the combination processing unit 74 may take the maximum value.
- b j max (a j1 , a i2 ,... a iK )
- Second intermediate data MD 2 to one and may be plural.
- N pieces of first intermediate data MD 1_1 to MD 1_N may be combined with two pieces of second intermediate data MD 2_1 and MD 2_2 .
- the plurality of first intermediate data MD 1_1 to MD 1_N are divided into two groups, the second intermediate data MD 2_1 is generated from one group, and the second intermediate data MD 2_2 is generated from the other group. May be.
- FIG. 20 is a block diagram of an object identification system 10C according to the third modification.
- a convolutional neural network is applied to the object identification system 10B of FIG.
- the convolutional neural network it is possible to improve the robustness against the positional deviation of the object in the horizontal direction.
- N of the plurality of line data is set to 8
- N may be about 4 to 12 in consideration of the calculation capability of the calculation processing device 40 and the required identification capability of the object OBJ.
- an object may be defined as a different type (category) for each direction in which the object is desired. That is, a certain object is identified as a different type depending on whether or not it is facing the host vehicle. This is because it is useful for estimating the moving direction of the object OBJ.
- the arithmetic processing unit 40 may be configured only by hardware using an FPGA or the like.
- the present invention relates to an object identification system.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Networks & Wireless Communication (AREA)
- Electromagnetism (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
- Optical Radar Systems And Details Thereof (AREA)
Abstract
Description
本明細書に開示される一実施の形態は、オブジェクト識別システムに関する。オブジェクト識別システムは、高さが異なる複数の水平ラインについて、複数のラインデータを生成する3次元センサと、複数のラインデータにもとづいてオブジェクトの種類(カテゴリーあるいはクラス)を識別する演算処理装置と、を備える。演算処理装置は、それぞれが、複数のラインデータの対応するひとつに関する第1中間データを生成し、第1中間データは、対応するラインデータが、複数の種類の複数の部位それぞれに該当する確率(所属確率)を示すものである、複数の第1ニューラルネットワークと、複数のラインデータに対応する複数の第1中間データを受け、複数の第1中間データを結合し、少なくともひとつの第2中間データを生成する結合処理部と、少なくともひとつの第2中間データを受け、オブジェクトが、複数の種類それぞれに該当する確率を示す最終データを生成する第2ニューラルネットワークと、を備える。
図1は、第1の実施の形態に係るオブジェクト識別システム10のブロック図である。このオブジェクト識別システム10は、自動車やバイクなどの車両に搭載される車載用であり、車両の周囲に存在するオブジェクトOBJの種類(カテゴリあるいはクラスともいう)を判定する。
図7(a)~(c)に示す第1の学習方法では、LiDARの設置(高さ、仰俯角、あるいはオブジェクトとの距離)に強く依存した学習が行われる場合がある。図13(a)、(b)は、LiDARの高さと、オブジェクトの関係を示す図である。図13(a)は、学習時にLiDARの設置高さを145cmとした場合を示す。このとき、下側の3本のラインデータは無効であり、下から4本目~8本目のラインデータLD4~LD8を用いた学習が行われる。
上の変形例は、学習方法を工夫することにより、識別力の改善を図るものであった。第2実施の形態では、演算処理装置の構成を工夫することにより、識別力の改善を実現する。
bj=Σi=1:Naji/N
図19は、第2の実施の形態におけるオブジェクト識別システムによって、図13(b)の状況で得られる最終データを示す図である。第2の実施の形態によれば、第1の実施の形態および第1の学習方法の組み合わせで得られる図13(c)の最終データに比べて、歩行者と認識する確率を高めることができる。
図18では、結合処理部74において単純平均値をとったが、重み付け平均をとってもよい。ciは、高さ(ライン)ごとの重み付けの係数である。
bj=Σi=1:Najici/N
bj=Σi=1:Naji
bj=max(aj1,ai2,…aiK)
図17では、第2中間データMD2を1個としたが、複数としてもよい。たとえばN個の第1中間データMD1_1~MD1_Nを、2個の第2中間データMD2_1,MD2_2に結合してもよい。この場合、たとえば、複数の第1中間データMD1_1~MD1_Nを2つの群に分けて、一方の群から第2中間データMD2_1を生成し、他方の群から第2中間データMD2_2を生成してもよい。
図20は、変形例3に係るオブジェクト識別システム10Cのブロック図である。この変形例は、図17のオブジェクト識別システム10Bに畳み込みニューラルネットワークを適用したものである。畳み込みニューラルネットワークの適用により、オブジェクトの横方向の位置ズレに対するロバスト性を高めることができる。
複数のラインデータの本数Nを8としたが、演算処理装置40の演算能力と、要求されるオブジェクトOBJの識別能力を考慮して、N=4~12程度としてもよい。
20 3次元センサ
40 演算処理装置
42 第1演算ユニット
44 第2演算ユニット
NN1 第1ニューラルネットワーク
NN2 第2ニューラルネットワーク
70 演算処理装置
72 第1ニューラルネットワーク
74 結合処理部
76 第2ニューラルネットワーク
50 入力層
52 中間層
54 出力層
60 入力層
62 中間層
64 出力層
100 自動車
102 前照灯
104 車両ECU
200 車両用灯具
202 光源
204 点灯回路
206 光学系
208 灯具ECU
Claims (11)
- 高さが異なる複数の水平ラインについて、複数のラインデータを生成する3次元センサと、
前記複数のラインデータにもとづいてオブジェクトの種類を識別する演算処理装置と、
を備え、
前記演算処理装置は、
それぞれが、前記複数のラインデータの対応するひとつに関する第1中間データを生成し、前記第1中間データは、対応するラインデータが、複数の種類の複数の部位それぞれに該当する確率を示すものである、複数の第1ニューラルネットワークと、
前記複数のラインデータに対応する複数の第1中間データを受け、前記複数の第1中間データを結合し、少なくともひとつの第2中間データを生成する結合処理部と、
前記少なくともひとつの第2中間データを受け、前記オブジェクトが、前記複数の種類それぞれに該当する確率を示す最終データを生成する第2ニューラルネットワークと、
を備えることを特徴とするオブジェクト識別システム。 - 前記少なくともひとつの第2中間データはひとつであり、前記第2中間データは、前記複数の第1中間データのすべてにもとづいて得られることを特徴とする請求項1に記載のオブジェクト識別システム。
- 前記少なくともひとつの第2中間データは複数であり、各第2中間データは、前記複数の第1中間データのうち連続するいくつかにもとづいて得られることを特徴とする請求項1に記載のオブジェクト識別システム。
- 前記少なくともひとつの第2中間データはそれぞれ、対応するいくつかの第1中間データの平均または総和であることを特徴とする請求項2または3に記載のオブジェクト識別システム。
- 複数の種類の複数の部位それぞれを測定した複数のラインデータを利用して、前記第1ニューラルネットワークに学習させるステップと、
学習済みの前記複数の第1ニューラルネットワークの出力を前記結合処理部を介して前記第2ニューラルネットワークと結合した状態で、前記第2ニューラルネットワークに学習させるステップと、
が実行されることを特徴とする請求項1から4のいずれかに記載のオブジェクト識別システム。 - 前記オブジェクトの種類は、少なくとも、歩行者、自転車に乗った人、自動車を含むことを特徴とする請求項1から5のいずれかに記載のオブジェクト識別システム。
- 請求項1から6のいずれかに記載のオブジェクト識別システムを備えることを特徴とする自動車。
- 前記3次元センサは、前照灯に内蔵されることを特徴とする請求項7に記載の自動車。
- 請求項1から6のいずれかに記載のオブジェクト識別システムを備えることを特徴とする車両用灯具。
- 3次元センサから得られる複数のラインデータにもとづきオブジェクトの種類を識別する方法であって、
前記ラインデータごとに、複数の種類の複数の部位それぞれに該当する確率を示す第1中間データを生成するステップと、
前記複数のラインデータについて得られる複数の第1中間データを結合し、少なくともひとつの第2中間データを生成するステップと、
前記少なくともひとつの第2中間データにもとづいて、前記オブジェクトが、前記複数の種類それぞれに該当する確率を示す最終データを生成するステップと、
を備えることを特徴とする方法。 - 3次元センサから得られる複数のラインデータにもとづきオブジェクトの種類を識別する演算処理装置の学習方法であって、
前記演算処理装置は、
それぞれが、前記複数のラインデータの対応するひとつに関する第1中間データを生成し、前記第1中間データは、対応するラインデータが、複数の種類の複数の部位それぞれに該当する確率を示すものである、複数の第1ニューラルネットワークと、
前記複数のラインデータに対応する複数の第1中間データを受け、前記複数の第1中間データを結合し、少なくともひとつの第2中間データを生成する結合処理部と、
前記少なくともひとつの第2中間データを受け、前記オブジェクトが、前記複数の種類それぞれに該当する確率を示す最終データを生成する第2ニューラルネットワークと、
を備え、
前記学習方法は、
複数の種類の複数の部位それぞれを測定した複数のラインデータを利用して、前記第1ニューラルネットワークに学習させるステップと、
学習済みの前記複数の第1ニューラルネットワークの出力を前記結合処理部を介して前記第2ニューラルネットワークと結合した状態で、前記第2ニューラルネットワークに学習させるステップと、
を備えることを特徴とする方法。
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2020506523A JP7174037B2 (ja) | 2018-03-15 | 2019-03-11 | オブジェクト識別システム、自動車、車両用灯具、オブジェクトの種類の識別方法 |
CN201980018990.2A CN111919139A (zh) | 2018-03-15 | 2019-03-11 | 对象识别系统、汽车、车辆用灯具、对象的种类的识别方法 |
EP19767179.5A EP3767338A4 (en) | 2018-03-15 | 2019-03-11 | OBJECT IDENTIFICATION SYSTEM, AUTOMOTIVE, VEHICLE LAMP ACCESSORY AND OBJECT TYPE IDENTIFICATION METHOD |
US17/021,410 US11841435B2 (en) | 2018-03-15 | 2020-09-15 | Object identification system |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2018047840 | 2018-03-15 | ||
JP2018-047840 | 2018-03-15 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/021,410 Continuation US11841435B2 (en) | 2018-03-15 | 2020-09-15 | Object identification system |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019176877A1 true WO2019176877A1 (ja) | 2019-09-19 |
Family
ID=67908260
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2019/009781 WO2019176877A1 (ja) | 2018-03-15 | 2019-03-11 | オブジェクト識別システム、自動車、車両用灯具、オブジェクトの種類の識別方法 |
Country Status (5)
Country | Link |
---|---|
US (1) | US11841435B2 (ja) |
EP (1) | EP3767338A4 (ja) |
JP (1) | JP7174037B2 (ja) |
CN (1) | CN111919139A (ja) |
WO (1) | WO2019176877A1 (ja) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH0829135A (ja) * | 1994-07-15 | 1996-02-02 | Sumitomo Osaka Cement Co Ltd | 物品形状測定装置およびそれを用いた物品認識装置 |
US20040036261A1 (en) * | 1995-06-07 | 2004-02-26 | Breed David S. | Method and apparatus for sensing a vehicle crash |
JP2009098023A (ja) | 2007-10-17 | 2009-05-07 | Toyota Motor Corp | 物体検出装置及び物体検出方法 |
JP2011186584A (ja) * | 2010-03-05 | 2011-09-22 | Daihatsu Motor Co Ltd | 物体認識装置 |
JP2015114261A (ja) * | 2013-12-13 | 2015-06-22 | 株式会社デンソーアイティーラボラトリ | 対象物検出装置、対象物検出方法およびプログラム |
JP2017056935A (ja) | 2015-09-14 | 2017-03-23 | トヨタ モーター エンジニアリング アンド マニュファクチャリング ノース アメリカ,インコーポレイティド | 3dセンサにより検出されたオブジェクトの分類 |
WO2019035363A1 (ja) * | 2017-08-18 | 2019-02-21 | 株式会社小糸製作所 | 認識センサおよびその制御方法、自動車、車両用灯具、オブジェクト識別システム、オブジェクトの識別方法 |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2005316888A (ja) * | 2004-04-30 | 2005-11-10 | Japan Science & Technology Agency | 顔認識システム |
EP2894600B1 (en) * | 2014-01-14 | 2018-03-14 | HENSOLDT Sensors GmbH | Method of processing 3D sensor data to provide terrain segmentation |
EP3295422B1 (en) * | 2015-05-10 | 2020-01-01 | Mobileye Vision Technologies Ltd. | Road profile along a predicted path |
CN106250812B (zh) * | 2016-07-15 | 2019-08-20 | 汤一平 | 一种基于快速r-cnn深度神经网络的车型识别方法 |
US10466714B2 (en) * | 2016-09-01 | 2019-11-05 | Ford Global Technologies, Llc | Depth map estimation with stereo images |
CN106599869B (zh) * | 2016-12-22 | 2019-12-03 | 安徽大学 | 一种基于多任务卷积神经网络的车辆属性识别方法 |
CN107067020B (zh) * | 2016-12-30 | 2019-11-15 | 腾讯科技(上海)有限公司 | 图片识别方法及装置 |
CN107506740B (zh) * | 2017-09-04 | 2020-03-17 | 北京航空航天大学 | 一种基于三维卷积神经网络和迁移学习模型的人体行为识别方法 |
-
2019
- 2019-03-11 CN CN201980018990.2A patent/CN111919139A/zh active Pending
- 2019-03-11 WO PCT/JP2019/009781 patent/WO2019176877A1/ja active Application Filing
- 2019-03-11 JP JP2020506523A patent/JP7174037B2/ja active Active
- 2019-03-11 EP EP19767179.5A patent/EP3767338A4/en active Pending
-
2020
- 2020-09-15 US US17/021,410 patent/US11841435B2/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH0829135A (ja) * | 1994-07-15 | 1996-02-02 | Sumitomo Osaka Cement Co Ltd | 物品形状測定装置およびそれを用いた物品認識装置 |
US20040036261A1 (en) * | 1995-06-07 | 2004-02-26 | Breed David S. | Method and apparatus for sensing a vehicle crash |
JP2009098023A (ja) | 2007-10-17 | 2009-05-07 | Toyota Motor Corp | 物体検出装置及び物体検出方法 |
JP2011186584A (ja) * | 2010-03-05 | 2011-09-22 | Daihatsu Motor Co Ltd | 物体認識装置 |
JP2015114261A (ja) * | 2013-12-13 | 2015-06-22 | 株式会社デンソーアイティーラボラトリ | 対象物検出装置、対象物検出方法およびプログラム |
JP2017056935A (ja) | 2015-09-14 | 2017-03-23 | トヨタ モーター エンジニアリング アンド マニュファクチャリング ノース アメリカ,インコーポレイティド | 3dセンサにより検出されたオブジェクトの分類 |
WO2019035363A1 (ja) * | 2017-08-18 | 2019-02-21 | 株式会社小糸製作所 | 認識センサおよびその制御方法、自動車、車両用灯具、オブジェクト識別システム、オブジェクトの識別方法 |
Non-Patent Citations (1)
Title |
---|
See also references of EP3767338A4 |
Also Published As
Publication number | Publication date |
---|---|
EP3767338A1 (en) | 2021-01-20 |
CN111919139A (zh) | 2020-11-10 |
US20200410262A1 (en) | 2020-12-31 |
JP7174037B2 (ja) | 2022-11-17 |
US11841435B2 (en) | 2023-12-12 |
EP3767338A4 (en) | 2021-12-01 |
JPWO2019176877A1 (ja) | 2021-03-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6682833B2 (ja) | 物体認識アルゴリズムの機械学習のためのデータベース構築システム | |
JP6398347B2 (ja) | 画像処理装置、認識対象物検出方法、認識対象物検出プログラム、および、移動体制御システム | |
JP6184877B2 (ja) | 車両用外界認識装置 | |
US7263209B2 (en) | Vehicular vision system | |
Ming et al. | Vehicle detection using tail light segmentation | |
CN105335955A (zh) | 对象检测方法和对象检测装置 | |
US11341604B2 (en) | Processing device for object identification | |
JP4937844B2 (ja) | 歩行者検出装置 | |
CN112037249A (zh) | 一种对摄像装置影像中物体进行跟踪的方法及装置 | |
CN116229408A (zh) | 一种图像信息与激光雷达点云信息融合的目标识别方法 | |
JP6483360B2 (ja) | 対象物認識装置 | |
CN111033316B (zh) | 识别传感器及其控制方法、汽车、车辆用灯具、对象识别系统、对象的识别方法 | |
US20190180122A1 (en) | Detection and Tracking of Objects from Images of a Camera | |
WO2019176877A1 (ja) | オブジェクト識別システム、自動車、車両用灯具、オブジェクトの種類の識別方法 | |
CN116740657A (zh) | 一种基于相似三角形的目标检测以及测距方法 | |
Kurnianggoro et al. | Camera and laser range finder fusion for real-time car detection | |
KR20230009598A (ko) | 차량 번호판 인식 및 위조 판단 방법과 그 방법이 수행되는 시스템 | |
JP7288460B2 (ja) | 車載用物体識別システム、自動車、車両用灯具、分類器の学習方法、演算処理装置 | |
Gorobetz et al. | Simulation of Computer Vision Based Sensor System for Autonomous Transport | |
CN113569803A (zh) | 一种基于多尺度卷积的多模态数据融合车道目标检测的方法及系统 | |
CN115713656A (zh) | 基于Transformer的多传感器融合目标检测方法 | |
Zhang et al. | Automatic recognition of traffic signs in natural scene image based on central projection transformation | |
Kwon et al. | Design and Implementation of Traffic-Scene Understanding System for Autonomous Vehicle |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19767179 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2020506523 Country of ref document: JP Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2019767179 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref document number: 2019767179 Country of ref document: EP Effective date: 20201015 |