A target recognition method based on a single-pixel time-resolved detector

Abstract

The present invention provides a target recognition method based on a single-pixel time-resolved detector. The method includes: emitting pulsed light through a pulsed laser, and illuminating a target object after emission by a transmitter; through a receiver and an optical fiber unit, the single-pixel time-resolved detector receives echo signals returned from each surface of the target object and records the arrival time of the echo signals to obtain distance echo characteristic data; preprocessing the distance echo characteristic data, and processing the preprocessed distance echo characteristic data by using a trained target recognition neural network to obtain the recognition result of the target object. The present invention also provides a training method for a target recognition neural network, an electronic device, and a storage medium.

Description

Technical Field

[0001] The present invention relates to the field of lidar, and particularly to a target recognition method based on a single-pixel time-resolved detector, a training method for a target recognition neural network, an electronic device, and a storage medium. Background Art

[0002] Target recognition technology uses a computer to analyze certain features of a target, and based on a discrimination function determined by known training samples, makes a decision classification in a classifier, ultimately realizing the recognition and judgment of the type, posture, and other properties of the target. Target recognition technology has a wide range of applications in many fields such as autonomous driving, remote monitoring, and national defense security.

[0003] Existing target recognition methods include methods for recognizing targets based on two-dimensional images, methods for target recognition based on target three-dimensional point clouds, and methods for target recognition based on radar spectra, etc. However, the above target recognition methods all require the detector to detect the target in an array or scanning manner, and the array or scanning form will occupy more resources. Therefore, the above target recognition methods have technical problems such as complex processing procedures, long data acquisition time, and relatively slow recognition speed. Summary of the Invention

[0004] In view of the above problems, the present invention provides a target recognition method based on a single-pixel time-resolved detector, a training method for a target recognition neural network, an electronic device, and a storage medium, in order to at least solve one of the above problems.

[0005] According to a first aspect of the present invention, there is provided a target recognition method based on a single-pixel time-resolved detector, including:

[0006] Emitting pulsed light through a pulsed laser, and illuminating the entire target object after being emitted by a transmitter;

[0007] Through a receiver and an optical fiber unit, a single-pixel time distribution detector receives the echo signals returned from the respective surfaces of the target object and records the time when the echo signals arrive, obtaining distance echo characteristic data, wherein the field of view of the single-pixel time distribution detector covers the target object;

[0008] Preprocessing the distance echo characteristic data, and processing the preprocessed distance echo characteristic data by using a trained target recognition neural network to obtain the recognition result of the target object, wherein the recognition result of the target object includes the type and posture information of the target object.

[0009] According to an embodiment of the present invention, the preprocessing of the distance echo characteristic data includes:

[0010] Obtain the data intensity information and peak position information of the distance echo characteristic data, and perform normalization preprocessing on the distance echo characteristic data based on the data intensity information;

[0011] Based on the data intensity information and peak position information, perform a cyclic shift operation on the distance echo characteristic data obtained after normalization preprocessing to obtain the preprocessed distance echo characteristic data.

[0012] According to an embodiment of the present invention, the above-mentioned pulsed laser is a laser capable of emitting picosecond to sub-nanosecond pulsed light;

[0013] Wherein, the divergence angles of the transmitter and the receiver are determined according to the size and distance of the target object, and the transmitter and the receiver include a collimator head, a lens, a scattering sheet, and a bare fiber head;

[0014] Wherein, the single-pixel time-resolved detector uses an indium gallium arsenide single-photon avalanche diode;

[0015] Wherein, the signal synchronization is achieved by using a signal generator to send trigger signals to the pulsed laser and the single-pixel time-resolved detector simultaneously through a synchronization device, and using the single-pixel time-resolved detector to record the arrival time of target echo photons.

[0016] According to a second aspect of the present invention, there is provided a training method for a target recognition neural network, which is applied to a target recognition method based on a single-pixel time-resolved detector, including:

[0017] Generate a distance echo characteristic training data set for training the target recognition neural network based on the attribute information of the sample object;

[0018] Based on the intensity invariance and translational invariance of the distance echo characteristic data, preprocess the distance echo characteristic training data set to obtain a preprocessed training data set;

[0019] According to a predefined loss function and a predefined optimizer, use the preprocessed training data set to iteratively train the target recognition neural network until a preset training condition is satisfied, and obtain a trained target recognition neural network.

[0020] According to an embodiment of the present invention, the above-mentioned target recognition neural network is constructed based on a one-dimensional Unet;

[0021] Wherein, the predefined loss function includes a cross-entropy loss function;

[0022] Wherein, the predefined optimizers include an adaptive moment estimation optimizer and a stochastic gradient descent optimizer.

[0023] According to an embodiment of the present invention, generating a distance echo characteristic training dataset for training a target recognition neural network based on the attribute information of a sample object includes:

[0024] When the three-dimensional model of the sample object is known, a simulation means is used to generate the distance echo characteristic training dataset;

[0025] When the three-dimensional model of the sample object is unknown, a method of expanding measured data is used to generate the distance echo characteristic training dataset;

[0026] Among them, using a simulation means to generate the distance echo characteristic training dataset includes:

[0027] Project the three-dimensional model of the sample object to obtain a two-dimensional reflectivity map and a two-dimensional depth map of the sample object;

[0028] Based on predefined Poisson sampling, process the two-dimensional reflectivity map and the two-dimensional depth map of the sample object to generate the distance echo characteristic training dataset;

[0029] Among them, using a method of expanding measured data to generate the distance echo characteristic training dataset includes:

[0030] Convert the measured data of arrival time-photon count obtained from the data acquisition system into data of a single photon arrival time series, and randomly shuffle the data of the single photon arrival time series;

[0031] Perform a data augmentation operation on the shuffled data of the single photon arrival time series to generate the distance echo characteristic training dataset.

[0032] According to an embodiment of the present invention, the above-mentioned projecting the three-dimensional model of the sample object to obtain a two-dimensional reflectivity map and a two-dimensional depth map of the sample object includes:

[0033] Determine a projection plane for displaying the projection of the sample object, and discretize the projection plane to determine the triangles on the surface of the three-dimensional model that the grid points pass through before projection;

[0034] Calculate the depth value corresponding to the grid points by means of plane interpolation, and after traversing all the grid points, the two-dimensional depth map of the sample object can be obtained;

[0035] Assign values to the corresponding positions in the two-dimensional reflectivity map of the special grid points, and remove the special grid points during the simulation process, where the special grid points refer to the grid points that do not pass through any triangles on the surface of the three-dimensional model.

[0036] According to an embodiment of the present invention, the above-mentioned performing a data augmentation operation on the shuffled data of the single photon arrival time series to generate the distance echo characteristic training dataset includes:

[0037] Randomly extract a part of the data in the scrambled single-photon arrival time series according to a preset ratio to complete the operation of randomly weakening the photon count;

[0038] Randomly add the extracted data to photons whose arrival times follow a uniform distribution to complete the operation of randomly adding noise counts;

[0039] Restatistical the data with added noise into data in the form of arrival time - photon count, and perform Poisson sampling to complete the operation of simulating the Poisson detection process, thereby generating a training data set for distance echo characteristics.

[0040] According to the third aspect of the present invention, there is provided an electronic device, including:

[0041] One or more processors;

[0042] A storage device for storing one or more programs,

[0043] wherein, when the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the target recognition method based on a single-pixel time-resolved detector and the training method of the target recognition neural network.

[0044] According to the third aspect of the present invention, there is provided a computer-readable storage medium, on which executable instructions are stored, and when the instructions are executed by a processor, the processor is caused to execute the target recognition method based on a single-pixel time-resolved detector and the training method of the target recognition neural network.

[0045] The target recognition method based on a single-pixel time-resolved detector provided by the present invention utilizes the distance echo characteristic data of the target to be recognized collected by the single-pixel time-resolved detector, and processes the distance echo characteristic data of the target to be recognized by the trained target recognition neural network to complete the recognition task; it is used to determine the type information and attitude information of the target to be recognized; compared with various technical solutions in the prior art, the above recognition method provided by the present invention has technical advantages such as simple hardware structure, long operating distance, and fast recognition speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of the target recognition method based on a single-pixel time-resolved detector according to an embodiment of the present invention;

[0047] Figure 2 is a schematic structural diagram of a data acquisition system based on a single-pixel time-resolved detector according to an embodiment of the present invention;

[0048] Figure 3 is a flowchart of the training method of the target recognition neural network according to an embodiment of the present invention;

[0049] Figure 4 It is a schematic diagram of the projection process of a three-dimensional target to be recognized during the process of simulating and generating target distance echo characteristic data according to an embodiment of the present invention;

[0050] Figure 5 It is a schematic diagram of generating distance echo characteristic data of a target to be recognized by expanding measured data according to an embodiment of the present invention;

[0051] Figure 6 It schematically shows a block diagram of an electronic device suitable for implementing a target recognition method based on a single-pixel time-resolved detector and a training method of a target recognition neural network according to an embodiment of the present invention. Detailed implementation manners

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings.

[0053] Currently, the most commonly used method for target recognition is the target recognition method based on two-dimensional images, that is, first use a camera to image the target, and then recognize the target according to the image. Among them, the target recognition method based on two-dimensional images includes traditional target recognition and classification methods such as those based on ISODATA (Iterative Self-Organizing Data Analysis Technique A), K-means, minimum distance, maximum likelihood, etc., and also includes target recognition methods based on deep learning born with the rapid development of hardware such as GPUs and the rapid growth of data sets.

[0054] In addition to the above-mentioned methods for recognizing targets based on two-dimensional images, there are currently methods for recognizing targets based on features such as target three-dimensional point clouds and radar spectra. In the target recognition method based on target three-dimensional point clouds, usually through devices such as lidar, using array or scanning detection methods, obtain the three-dimensional point cloud data of the target, then extract features from it, and use traditional or deep learning methods to obtain the type and pose of the target, and further downstream tasks such as tracking and segmentation can be performed; while in the target recognition method based on radar spectra, the spectral data obtained after radar scanning is used to classify different objects.

[0055] However, existing target recognition methods all require detectors (such as CCD cameras, lidar) to detect targets in an array or scanning manner, and the array or scanning form will occupy more resources, and there are also technical problems such as slow recognition speed, long data acquisition time, and long required data acquisition time.

[0056] In view of various problems existing in the prior art, the present invention provides a target recognition method based on a single-pixel time-resolution detector. By only using an acquisition system with a time-resolution detector based on a single pixel to obtain one-dimensional feature data of a target from the distance dimension, the recognition of the target type and posture can be achieved. The above method provided by the present invention has the advantages of a long action distance, a simple structure of the data acquisition system, a short required data acquisition time, and a fast recognition speed.

[0057] It should be specifically noted that all the data of the target to be recognized involved in the present invention have obtained the consent and authorization of the relevant parties, and with the permission of the relevant parties, the data of the target to be recognized are processed, stored, and applied. The relevant processes comply with the requirements of laws and regulations and necessary confidentiality measures are taken, meeting the requirements of public order and good customs.

[0058] Figure 1 It is a flowchart of a target recognition method based on a single-pixel time-resolution detector according to an embodiment of the present invention.

[0059] As Figure 1 shown, the above target recognition method based on a single-pixel time-resolution detector includes operations S110 to S130.

[0060] In operation S110, a pulsed laser emits pulsed light, and after being emitted by a transmitter, it illuminates the entire target object.

[0061] According to an embodiment of the present invention, the above pulsed laser is a laser capable of emitting picosecond to sub-nanosecond pulsed light, such as a laser that can emit narrow pulses; wherein, the divergence angles of the transmitter and the receiver are determined according to the size and distance of the target object, and the transmitter and the receiver include a collimator head, a lens, a diffuser, and a bare fiber head; wherein, the single-pixel time-resolution detector uses an indium gallium arsenide single-photon avalanche diode; wherein, the signal synchronization is achieved by using a signal generator in a synchronization device to send trigger signals to the pulsed laser and the time-to-digital converter simultaneously, and using the time-to-digital converter to record the arrival time of the pulsed light.

[0062] In operation S120, through the receiver and the fiber unit, the single-pixel time-resolution detector receives the echo signals returned from the respective surfaces of the target object and records the arrival time of the echo signals to obtain distance echo characteristic data, wherein the field of view of the single-pixel time distribution detector covers the target object.

[0063] In operation S130, the distance echo characteristic data is preprocessed, and the preprocessed distance echo characteristic data is processed by using a trained target recognition neural network to obtain the recognition result of the target object, wherein the recognition result of the target object includes the type and posture information of the target object.

[0064] In operation S130, preprocessing the range echo characteristic data includes: obtaining data strength information and peak position information of the range echo characteristic data, and performing normalization preprocessing on the range echo characteristic data based on the data strength information; and performing a cyclic shift operation on the range echo characteristic data obtained after the normalization preprocessing based on the data strength information and the peak position information to obtain the preprocessed range echo characteristic data.

[0065] The target recognition method based on a single-pixel time-resolved detector provided by the present invention collects distance echo characteristic data of the target to be identified based on the single-pixel time-resolved detector, and processes the distance echo characteristic data of the target to be identified with a trained target recognition neural network to complete the recognition task, so as to determine the type information and posture information of the target to be identified; compared with various technical solutions in the prior art, the above-mentioned recognition method provided by the present invention has the technical advantages of simple hardware structure, long effective distance, fast recognition speed, etc.

[0066] The following is combined with Figure 2 A brief description is given of how the above-mentioned target recognition method based on a single-pixel time-resolved detector performs the data collection process of the target to be identified.

[0067] Figure 2 4 is a schematic diagram of the structure of a data acquisition system based on single-pixel time-resolved detectors according to an embodiment of the present invention.

[0068] like Figure 2 As shown in the figure, first, the pulse light source emits pulse light, which illuminates the target after passing through the transmitter; the echo signal is received by the single-pixel time-resolved detector after passing through the receiver and optical fiber and other devices, and the detector's field of view covers the entire target. The time-resolved detector can receive and record the arrival time of the light signal returned from various locations on the target surface, forming a one-dimensional distance echo characteristic data, which contains information such as the type and posture of the object.

[0069] Figure 3 4 is a flow chart of a method for training a target recognition neural network according to an embodiment of the present invention.

[0070] like Figure 3 As shown, the above-mentioned training method of the target recognition neural network is applied to the target recognition method based on the single-pixel time-resolved detector, including operations S310 to S330.

[0071] In operation S310, a range echo characteristic training data set for training a target recognition neural network is generated based on attribute information of a sample object.

[0072] In operation S320, based on the intensity invariance and translation invariance of the range echo characteristic data, the range echo characteristic training data set is preprocessed to obtain a preprocessed training data set.

[0073] In operation S330, according to a predefined loss function and a predefined optimizer, the target recognition neural network is iteratively trained using the preprocessed training data set until a preset training condition is met, and the trained target recognition neural network is obtained.

[0074] According to an embodiment of the present invention, the above-mentioned target recognition neural network is constructed based on a one-dimensional Unet; wherein, the predefined loss function includes a cross-entropy loss function; wherein, the predefined optimizer includes an adaptive moment estimation optimizer and a stochastic gradient descent optimizer.

[0075] The above operations S310 to S330 are intended to obtain a target recognition neural network for target recognition. By generating targeted training data, the efficiency of the subsequent training process can be improved, and the recognition effect of the model can be improved and the generalization of the model can be enhanced.

[0076] According to an embodiment of the present invention, the generation of the range echo characteristic training data set for training the target recognition neural network based on the attribute information of the sample object includes: when the three-dimensional model of the sample object is known, a simulation method is used to generate the range echo characteristic training data set; when the three-dimensional model of the sample object is unknown, a method of expanding the measured data is used to generate the range echo characteristic training data set.

[0077] Among them, the use of the simulation method to generate the range echo characteristic training data set includes: projecting the three-dimensional model of the sample object to obtain a two-dimensional reflectivity map and a two-dimensional depth map of the sample object; based on the predefined Poisson sampling, processing the two-dimensional reflectivity map and the two-dimensional depth map of the sample object to generate the range echo characteristic training data set.

[0078] Among them, the use of the method of expanding the measured data to generate the range echo characteristic training data set includes: converting the measured data of arrival time-photon count obtained from the data acquisition system into data of a single photon arrival time series, and randomly shuffling the data of the single photon arrival time series; performing a data augmentation operation on the shuffled data of the single photon arrival time series to generate the range echo characteristic training data set.

[0079] According to an embodiment of the present invention, the above-mentioned projection of the three-dimensional model of the sample object to obtain the two-dimensional reflectivity map and the two-dimensional depth map of the sample object includes: determining a projection plane for displaying the projection of the sample object, and discretizing the projection plane to determine the grid points passing through the triangles on the surface of the three-dimensional model before projection; calculating the depth values corresponding to the grid points by means of plane interpolation, and obtaining the two-dimensional depth map of the sample object after traversing all the grid points; assigning values to the corresponding positions in the two-dimensional reflectivity map of the special grid points, and removing the special grid points during the simulation process, where the special grid points refer to the grid points that do not pass through any triangles on the surface of the three-dimensional model.

[0080] According to an embodiment of the present invention, the above-mentioned data augmentation operation on the data of the scrambled single-photon arrival time sequence to generate a training data set of distance echo characteristics includes: randomly extracting a part of the data of the scrambled single-photon arrival time sequence according to a preset ratio to complete the operation of randomly weakening the photon count; randomly adding photons with arrival times following a uniform distribution to the extracted data to complete the operation of randomly adding noise counts; re-statistical the data with added noise into data in the form of arrival time-photon count, and performing Poisson sampling to complete the operation of simulating the Poisson detection process, thereby generating a training data set of distance echo characteristics.

[0081] The following combines specific embodiments and attached Figure 4 and 5 to further elaborate on the generation process of the above-mentioned distance echo characteristic data set in detail.

[0082] Figure 4 FIG. is a schematic diagram of the projection process of a three-dimensional target to be recognized according to an embodiment of the present invention.

[0083] Figure 5 FIG. is a schematic diagram of generating distance echo characteristic data of a target to be recognized by augmenting measured data according to an embodiment of the present invention.

[0084] For different tasks, corresponding data sets need to be generated as training sets: for the task of identifying target types, distance echo characteristic data corresponding to different types of targets are required as training sets; for the task of identifying target postures, distance echo characteristic data corresponding to the target in different postures are required as training sets.

[0085] When the three-dimensional model of the target to be recognized is known, simulation means are used to generate a target distance echo characteristic data set.

[0086] If the three-dimensional model of the target is known, first project it into a two-dimensional reflectivity map and a two-dimensional depth map. During the projection process, first determine the direction of the projection plane and then discretize the projection area. Find which small triangle on the surface of the target three-dimensional model each grid point passes through before projection, and then calculate the depth value corresponding to the grid point by plane interpolation. After traversing all grid points, the two-dimensional depth map of the target can be obtained; for some grid points that do not pass through any small triangles, assign 0 (assign 1 to others) to the corresponding positions in the two-dimensional reflectivity map of these grid points to eliminate these grid points in the subsequent simulation process. As shown in the appendix Figure 4 As shown, taking the target "rabbit" as an example to illustrate this projection process.

[0087] After completing the projection operation, continue to directly simulate and generate the target range echo characteristic data according to the target two-dimensional reflectivity map and two-dimensional depth map. The finally obtained target range echo characteristic data can be represented by a one-dimensional histogram to represent, The th element has the form shown in formula (1):

[0088] (1),

[0089] where, represents the operation of Poisson sampling; is the number of vertical and horizontal pixels of the two-dimensional reflectivity map and intensity map after projection; is a coefficient used to control the overall intensity of the target range echo characteristic data obtained by simulation; represents the value of the corresponding grid point in the target two-dimensional reflectivity map; is the overall impulse response function of the data acquisition system, and its full width at half maximum (FWHM) matches the overall time jitter characteristic of the data acquisition system; is the two-dimensional depth map of the target obtained by projection; is the speed of light, and its value is ; is the time resolution of the data acquisition system; is a coefficient used to control the noise level of the target range echo characteristic data obtained by simulation. During the simulation process, according to the characteristics of the target range echo characteristic data in practice (including signal level, noise level, etc.), randomly take values for and within a reasonable range.

[0090] For the case where the three-dimensional model of the target to be recognized is unknown, the present invention adopts the method of expanding the measured data to generate the target range echo characteristic data set, as shown in Figure 5The distance echo characteristic data of the target to be recognized is schematically shown as being augmented from measured data. The present invention uses a data acquisition system to obtain a certain number of measured target distance echo characteristic data, and generates a large-scale target distance echo characteristic data set by means of randomly weakening photon counts, randomly adding noise counts, and simulating the Poisson detection process. First, the data in the form of "arrival time - photon count" obtained by the data acquisition system is converted into the form of "single photon arrival time sequence", and this sequence is randomly shuffled. In the operation of randomly weakening photon counts, the present invention randomly selects a part of this sequence in proportion; in the operation of randomly adding noise counts, the present invention randomly adds photons with arrival times subject to a uniform distribution to this sequence; in the operation of simulating the Poisson detection process, the present invention re-statistics the sequence after the first two operations into the data form of "arrival time - photon count", and then adds the Poisson sampling process.

[0091] Before using the above training data set for training the target recognition neural network, it is necessary to preprocess the above training data, and on the premise of maintaining the characteristics of the training data set, it is better used for model training. And for the tasks to be performed (i.e., target type recognition or target pose recognition), the preprocessed training data set is used to train the target recognition neural network.

[0092] First, considering the intensity invariance of the target distance echo characteristic data, that is, the absolute intensity of the data does not affect the final recognition result, the data is preprocessed by normalization, as shown in formula (2):

[0093] (2),

[0094] where, is the length of the target distance echo characteristic data. After the above normalization operation, the intensity invariance of the target distance echo characteristic data can be guaranteed.

[0095] Secondly, considering the translational invariance of the target distance echo characteristic data, that is, the absolute distance of the target relative to the data acquisition system, corresponding to the front and back positions of the signal peak in the target distance echo characteristic data, does not affect the final recognition result, the data is subjected to a cyclic shift operation, as shown in formula (3)

[0096] (3),

[0097] where satisfies formula (4):

[0098] (4),

[0099] The above cyclic shift operation can shift the "peak value" (i.e., the subscript is (position) is moved to the center of the echo characteristic data, and the obtained satisfies formula (5):

[0100] (5).

[0101] Through the above cyclic shift operation, the translational invariance of the target echo characteristic data can be guaranteed.

[0102] In summary, the above two-step preprocessing can improve the versatility of the deep learning algorithm. The preprocessed data can be directly input into the network for training, and at the same time, the GPU is used to accelerate the training process.

[0103] The above recognition method provided by the present invention will be further described in detail below in combination with specific experiments.

[0104] When building a data acquisition system based on a single-pixel time-resolved detector, the pulsed light source can use a picosecond pulsed laser with a central wavelength of 1550 nm (narrow pulse width, sub-nanosecond, picosecond pulsed light sources in other bands can be used as alternatives); the emitter and receiver can use a collimator with a divergence angle of 2.25 mrad (depending on the actual distance of the target, devices with other different divergence angles can be used as alternatives, such as lenses, diffusers, bare fiber heads, collimators with other different focal lengths, etc.); the detector can use an indium gallium arsenide single-photon avalanche diode (other detectors with time-resolving capabilities and matching the laser band can be used as alternatives); the synchronization device can use a signal generator to send trigger signals to the laser and the TDC (time-to-digital converter) that cooperates with the detector at the same time, and use the TDC to record the arrival time of the detector optical signal to achieve (it can also be replaced by the method of triggering the TDC after the laser optical signal passes through the photoelectric converter, and at the same time the TDC records the arrival time of the detector optical signal).

[0105] In the recognition algorithm, first according to the final usage requirements (target type recognition or pose recognition), the three-dimensional model of the target can be obtained from a public database (such as ShapeNet), and the target distance echo characteristic dataset can be generated according to the foregoing simulation means, or first use the above-mentioned data acquisition system to collect the target distance echo characteristic data, and then obtain the target echo characteristic dataset according to the foregoing data augmentation method; after generating about 1000 target echo characteristic data samples for each category, they are preprocessed according to the foregoing preprocessing method and then input into the neural network for training. The specific structure of the neural network can use 1D-UNet (other network structures that can be used to process one-dimensional data can be used as alternatives). The loss function used during training can use cross-entropy (other loss functions that can describe the classification accuracy can be used as alternatives), and the optimizer can use ADAM (other optimizers such as SGD, etc. can be used as alternatives). The parameters of the optimizer are = 0.9, = 0.999, and the learning rate is 0.001 (other reasonable parameter settings can be used as alternatives).

[0106] Finally, by combining the time-related data acquisition system and the target recognition algorithm, this technology first detects high-speed dynamic targets and obtains their range echo characteristic data; then uses the target recognition algorithm to output the prediction results of the target type and attitude, realizing target recognition.

[0107] The target recognition method based on a single-pixel time-resolved detector provided by the present invention has the following positive effects compared with the existing mainstream target recognition technologies.

[0108] Simple hardware system structure: Compared with the existing mainstream technologies (such as lidar) that require array detectors or scanning means, this technology only needs a single-pixel time-resolved detector to collect the target range characteristic data and realize the recognition of the target, without adding scanning components or using array detectors.

[0109] Long operating range: Compared with the existing technical approach of "imaging first and then recognizing", this technology does not require imaging of the target and has no requirement for the lateral spatial resolution of the detection and recognition system. As the operating range increases, the lateral spatial resolution of the existing detection and recognition system will deteriorate, and the imaging quality will decrease significantly, thus affecting the system's recognition ability. However, the technology described in this patent does not rely on the lateral resolution, so the operating range is not limited. Combining with the gradually mature single-photon detection devices in recent years, the detection and recognition range of this technology can reach more than one hundred kilometers.

[0110] Fast recognition speed: This method only needs to collect one-dimensional target range echo characteristic data, which has less data volume and simpler data form compared with the existing technologies (mainly using two-dimensional images or three-dimensional point cloud data). At the same time, this technology uses a deep learning algorithm accelerated by GPU, so the algorithm runs faster (about 10 ms order of magnitude), which is beneficial to the related applications of high-speed dynamic target recognition.

[0111] Figure 6 A block diagram of an electronic device suitable for implementing the target recognition method based on a single-pixel time-resolved detector and the training method of the target recognition neural network according to an embodiment of the present invention is schematically shown.

[0112] As Figure 6As shown, the electronic device 600 according to an embodiment of the present invention includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage section 608 into a random access memory (RAM) 603. The processor 601 can include, for example, a general-purpose microprocessor (such as a CPU), an instruction set processor, and / or a related chipset, and / or a dedicated microprocessor (such as an application-specific integrated circuit (ASIC)), etc. The processor 601 can also include on-board memory for caching purposes. The processor 601 can include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0113] In the RAM 603, various programs and data required for the operation of the electronic device 600 are stored. The processor 601, ROM 602, and RAM 603 are connected to each other via a bus 604. The processor 601 performs various operations of the method flow according to an embodiment of the present invention by executing the programs in the ROM 602 and / or RAM 603. It should be noted that the program can also be stored in one or more memories other than the ROM 602 and RAM 603. The processor 601 can also perform various operations of the method flow according to an embodiment of the present invention by executing the programs stored in one or more memories.

[0114] According to an embodiment of the present invention, the electronic device 600 may further include an input / output (I / O) interface 605, and the input / output (I / O) interface 605 is also connected to the bus 604. The electronic device 600 may further include one or more of the following components connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including, for example, a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed so that a computer program read from it can be installed into the storage section 608 as needed.

[0115] The present invention also provides a computer-readable storage medium, which can be included in the device / device / system described in the above embodiment; or can exist separately without being assembled into the device / device / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to an embodiment of the present invention is implemented.

[0116] According to an embodiment of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present invention, the computer-readable storage medium may be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, the computer-readable storage medium may include one or more memories other than the above-described ROM 602 and / or RAM 603 and / or ROM 602 and RAM 603.

[0117] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the above module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram or flowchart, as well as the combination of blocks in the block diagram or flowchart, may be implemented by a dedicated hardware-based system that performs the specified functions or operations, or may be implemented by a combination of dedicated hardware and computer instructions.

[0118] The above specific embodiments have further elaborated on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A target recognition method based on a single-pixel time-resolved detector, characterized in that, Including: Emitting pulsed light through a pulsed laser, and illuminating the entire target object after being emitted by a transmitter; Through a receiver and an optical fiber unit, the single-pixel time-resolved detector receives the echo signals returned from the respective surfaces of the target object and records the arrival time of the echo signals, obtaining range echo characteristic data, wherein the field of view of the single-pixel time-resolved detector covers the target object; Preprocessing the range echo characteristic data, and processing the preprocessed range echo characteristic data by using a trained target recognition neural network to obtain the recognition result of the target object, wherein the recognition result of the target object includes the type and pose information of the target object; Wherein, the trained target recognition neural network is trained by the following training method: When the three-dimensional model of the sample object is known, using simulation means to generate the range echo characteristic training data set; When the three-dimensional model of the sample object is unknown, using the method of expanding measured data to generate the range echo characteristic training data set; Wherein, the using simulation means to generate the range echo characteristic training data set includes: Projecting the three-dimensional model of the sample object to obtain the two-dimensional reflectivity map and two-dimensional depth map of the sample object; Based on predefined Poisson sampling, processing the two-dimensional reflectivity map and two-dimensional depth map of the sample object to generate the range echo characteristic training data set; Wherein, the using the method of expanding measured data to generate the range echo characteristic training data set includes: Converting the measured data of arrival time-photon count obtained from the data acquisition system into data of a single photon arrival time series, and randomly shuffling the data of the single photon arrival time series; Performing a data augmentation operation on the shuffled data of the single photon arrival time series to generate the range echo characteristic training data set; Based on the intensity invariance and translational invariance of the range echo characteristic data, preprocessing the range echo characteristic training data set to obtain a preprocessed training data set; According to a predefined loss function and a predefined optimizer, using the preprocessed training data set to iteratively train the target recognition neural network until a preset training condition is satisfied, obtaining a trained target recognition neural network.

2. The method according to claim 1, wherein Preprocessing the range echo characteristic data includes: Obtaining the data intensity information and peak position information of the range echo characteristic data, and performing normalization preprocessing on the range echo characteristic data based on the data intensity information; Based on the data intensity information and the peak position information, performing a cyclic shift operation on the range echo characteristic data obtained after normalization preprocessing to obtain the preprocessed range echo characteristic data.

3. According to the method of claim 1, characterized in that The pulsed laser is a laser capable of emitting picosecond to sub-nanosecond pulsed light; Wherein, the divergence angles of the transmitter and the receiver are determined according to the size and distance of the target object, and the transmitter and the receiver include a collimator head, a lens, a diffuser, and a bare fiber head; Among them, the single-pixel time-resolved detector uses an indium gallium arsenide single-photon avalanche diode; Among them, the signal synchronization is realized by using a signal generator in the synchronization device to send trigger signals to the pulsed laser and the single-pixel time-resolved detector at the same time, and using the single-pixel time-resolved detector to record the arrival time of the target echo photons.

4. The method according to claim 1, wherein The target recognition neural network is constructed based on one-dimensional Unet; Among them, the predefined loss function includes the cross-entropy loss function; Among them, the predefined optimizer includes the adaptive moment estimation optimizer and the stochastic gradient descent optimizer.

5. The method according to claim 1, characterized in that, Projecting the three-dimensional model of the sample object to obtain the two-dimensional reflectivity map and the two-dimensional depth map of the sample object includes: Determine the projection plane for displaying the projection of the sample object, and discretize the projection plane to determine the triangles on the surface of the three-dimensional model passed by the lattice points before projection; Calculate the depth value corresponding to the lattice points by the method of plane interpolation. After traversing all the lattice points, the two-dimensional depth map of the sample object can be obtained; Assign values to the corresponding positions in the two-dimensional reflectivity map of the special lattice points, and remove the special lattice points during the simulation process, where the special lattice points represent the lattice points that do not pass through any triangles on the surface of the three-dimensional model.

6. The method according to claim 1, characterized in that, Performing data augmentation operations on the data of the shuffled single-photon arrival time sequence to generate the distance echo characteristic training data set includes: Randomly extract part of the data of the shuffled single-photon arrival time sequence according to a preset ratio to complete the operation of randomly weakening the photon count; Randomly add photons with arrival times obeying a uniform distribution to the extracted data to complete the operation of randomly adding noise counts; Re-stat the data with added noise into data in the form of arrival time-photon count, and perform Poisson sampling to complete the operation of simulating the Poisson detection process, thereby generating the distance echo characteristic training data set.

7. An electronic device, comprising: One or more processors; A storage device for storing one or more programs, Among them, when the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, on which executable instructions are stored, and when the instructions are executed by a processor, the processor is caused to execute the method according to any one of claims 1 to 6.

Images