Target recognition method and system for micro UAV

Abstract

Title: Unmanned aerial vehicle (UAV) target recognition utilising a Recurrent Spiking Neural Network (RSNN) A target recognition method for an unmanned aerial vehicle (UAV), comprising: obtaining an i

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of target recognition and information processing, and in particular, to a target recognition method and system for a micro unmanned aerial vehicle (UAV). BACKGROUND

[0002] Micro UAVs can be used for monitoring and patrolling equipment in narrow or dangerous areas, security-related remote reconnaissance and target tracking, as well as search and exploration in disaster areas. Perception and understanding of specific targets by UAVs have always been important tasks of UAV systems and have received widespread attention and research. Traditional target recognition methods mainly rely on image feature extraction and target classification, usually employing a combination of image processing methods and machine learning algorithms. For example, common methods include using Histogram of Oriented Gradients (HOG) for feature extraction and Support Vector Machine (SVM) for target recognition. With the rapid development of deep learning technology, target detection and recognition methods based on deep learning have become a research hotspot, e.g., Regional Convolutional Neural Networks (R-CNN). You Only Look Once (YOLO), Single Shot MultiBox Detector (SSD), and other algorithms. These algorithms can simultaneously perform tasks such as target detection and classification.

[0003] Most existing UAV target recognition technologies mainly use deep learning in the field of artificial intelligence as the primary recognition method. In this approach, deep convolutional neural networks are often used as the network for feature extraction and target recognition. However, deep convolutional neural networks employ rate-based coding, and the hardware implementation consumes more energy, which is not conducive to the endurance of micro UAVs. SUMMARY

[0004] An objective of the present disclosure is to provide a target recognition method and system for a micro UAV, to achieve target recognition using a trained target recognition model in a neuromorphic chip on the micro UAV. A trained recognizing network in the trained target recognition model is a Recurrent Spiking Neural Network (RSNN). Target recognition using the RSNN significantly reduces hardware energy consumption, addressing the endurance issues of micro UAVs.

[0005] To achieve the above objective, the present disclosure provides the following technical solutions:

[0006] According to a first aspect, the present disclosure provides a target recognition method for a micro UAV. The target recognition method for a micro UAV includes:

[0007] obtaining a target image of a to-be-recognized target, where the target image is an image of the to-be-recognized target captured by an image acquisition device of a micro UAV;

[0008] inputting the target image to a trained target recognition model to obtain a target recognition signal, where the trained target recognition model is deployed on a neuromorphic chip of the micro UAV; the trained target recognition model includes a trained feature extraction model and a trained recognizing network; the trained recognizing network is a Recurrent Spiking Neural Network (RSNN); a spiking neuron model in the RSNN is constructed from a Leaky Integrate-and-Fire (LIF) neuron model and an Adaptive Leaky Integrate-and-Fire (ALIF) neuron model; and

[0009] transmitting the target recognition signal to a ground device.

[0010] According to a second aspect, the present disclosure provides a target recognition system for a micro UAV, for implementing the target recognition method for a micro UAV as described in the first aspect. The target recognition system for a micro UAV includes an acquisition module, a recognition module, and a wireless transmission module.

[0011] The acquisition module is configured to acquire a target image of a to-be-recognized target.

[0012] The recognition module is a neuromorphic chip; a trained target recognition model is deployed on the neuromorphic chip; the trained target recognition model is configured to identify the target image to obtain a target recognition signal; the trained target recognition model includes a trained feature extraction model and a trained recognizing network; the trained recognizing network is an RSNN; a spiking neuron model in the RSNN is composed of a LIF neuron model and an adaptive LIF neuron model.

[0013] The wireless transmission module is configured to transmit the target recognition signal to a ground device.

[0014] According to specific embodiments provided in the present disclosure, the present disclosure discloses the following technical effects:

[0015] The present disclosure provides a target recognition method and system for a micro UAV. After a target image of a to-be-recognized target is obtained, target recognition is performed using a trained target recognition model in a neuromorphic chip on the micro UAV. A trained recognizing network in the trained target recognition model is an RSNN. Target recognition using the RSNN significantly reduces hardware energy consumption, thereby addressing the endurance issues of micro UAVs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To describe the technical solutions in the embodiments of the present disclosure or in the prior art more clearly, the following briefly describes the accompanying drawings required for the embodiments. Apparently, the accompanying drawings in the following description show merely some embodiments of the present disclosure, and a person of ordinary skill in the art may still derive other accompanying drawings from these accompanying drawings without creative efforts.

[0017] FIG. 1 is a schematic flowchart of a target recognition method for a micro UAV according to an embodiment of the present disclosure;

[0018] FIG. 2 is a schematic flowchart of deploying a target recognition model on a neuromorphic chip according to an embodiment of the present disclosure;

[0019] FIG. 3 is a schematic structural diagram of a feature extraction model and inner and outer loop recognition models according to an embodiment of the present disclosure;

[0020] FIG. 4 is a schematic diagram of a meta-learning stage according to an embodiment of the present disclosure; and

[0021] FIG. 5 is a schematic diagram of functional modules of a target recognition system for a micro UAV according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The technical solutions in the embodiments of the present disclosure are clearly and completely described below with reference to the drawings in the embodiments of the present disclosure. Apparently, the described embodiments are only some rather than all of the embodiments of the present disclosure. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0023] To make the above objectives, features, and advantages of the present disclosure more obvious and easy to understand, the present disclosure will be further described in detail with reference to the accompanying drawings and specific implementations.

[0024] Target recognition of UAVs faces the requirements of real-time performance and accuracy. Real-time performance requires recognition algorithms to complete target detection and classification within a limited time and to update recognition information promptly. Accuracy demands that recognition algorithms accurately identify and classify targets, with a certain robustness to environmental factors such as lighting and weather. Traditional target recognition methods lack integration and cannot identify targets quickly and accurately. Deep learning methods require a large amount of labeled data for training, leading to poor model stability, weak adaptability, high hardware energy consumption, and inefficiency in completing recognition tasks.

[0025] The application of deep neural network models to micro UAVs often encounters the following issues: training deep neural network models requires a large number of samples to overcome overfitting. However, in reality, some recognition tasks lack sufficient training samples, such as facial recognition in law enforcement, recognition of equipment safety hazards, and exploration and recognition of the causes of major disasters.

[0026] Therefore, the present disclosure introduces a micro UAV based on neuromorphic chips, aiming to provide an efficient and accurate target recognition method. Neuromorphic chips are a novel type of neural computing hardware inspired by the working principle of the human brain, utilizing spiking neural network models for computation. Unlike traditional digital computers, neuromorphic chips leverage features like spiking neurons and spike encoding to simulate information transmission and processing between neurons, offering higher computational efficiency and energy advantages. Neuromorphic chips have broad application potential in pattern recognition, data processing, and robot control, providing computational support for target recognition. Neuromorphic chips have efficient computing capabilities, reasonable biological interpretability, low energy consumption, and can accelerate the computation process, enhancing the efficiency of target recognition. Moreover, the training model adopts a few-shot learning method based on meta-learning, enhancing recognition accuracy and robustness through coordination between inner and outer loops. This comprehensive application approach effectively addresses the limitations of traditional methods and deep learning methods, achieving efficient and accurate target recognition tasks.

[0027] In an exemplary embodiment, as shown in FIG. 1, a target recognition method for a micro UAV is provided. The method is executed by a computer device. Specifically, the method can be executed independently by a computer device such as a terminal or a server, or executed jointly by a terminal and a server. In this embodiment of the present disclosure, the target recognition method for a micro UAV includes step 101 to step 103 as follows:

[0028] Step 101: Obtain a target image of a to-be-recognized target, where the target image is an image of the to-be-recognized target captured by an image acquisition device of a micro UAV.

[0029] Step 102: Input the target image to a trained target recognition model to obtain a target recognition signal, where the trained target recognition model is deployed on a neuromorphic chip of the micro UAV; the trained target recognition model includes a trained feature extraction model and a trained recognizing network; the trained recognizing network is an RSNN; a spiking neuron model in the RSNN is constructed from a LIF neuron model and an ALIF neuron model.

[0030] The trained feature extraction model is a Convolutional Neural Network (CNN) model.

[0031] The step of inputting the target image to the trained target recognition model to obtain the target recognition signal specifically includes: inputting the target image to the trained feature extraction model to obtain a target feature sequence; and

[0032] inputting the target feature sequence to the trained recognizing network to obtain the target recognition signal.

[0033] Step 103: Transmit the target recognition signal to a ground device.

[0034] Through step 101 to step 103, target recognition is implemented using the trained target recognition model in the neuromorphic chip on the micro UAV. The trained recognizing network in the trained target recognition model is an RSNN. Target recognition using the RSNN significantly reduces hardware energy consumption, thereby addressing the endurance issues of micro UAVs. Additionally, the present disclosure can train the target recognition model using a few-shot learning method based on meta-learning, and the entire learning process is divided into two stages. Inspired by human ability to abstract general concepts from a small number of samples, the idea is to leverage past knowledge and experience to guide learning for new tasks. In the first stage of learning, a large amount of data of the same class as the target in the current specific task is required. For example, identifying device safety hazards requires accumulated images of safety hazards to allow the model to learn a general pattern for such tasks. In the second stage of learning, only a small amount of data specific to the target of the current task is needed. For example, facial recognition during law enforcement would require one or a few images of a fugitive to specialize the general pattern, significantly improving the accuracy and adaptability of the target recognition model.

[0035] The process of target recognition using the neuromorphic chip in the present disclosure is illustrated in FIG. 2. By embedding the neuromorphic chip into the micro UAV, a series of tasks including target acquisition, target recognition, and transmission of recognition results are completed. The method for target recognition involves the following steps: constructing a neural network model with inner and outer loop coordination; pre-training the neural network model using a few-shot learning method based on meta-learning, and importing the trained neural network model into the neuromorphic chip; capturing a target image by a camera of the micro UAV in real time, and transmitting the target image to the neuromoiphic chip through internal circuits for processing by the neural network model to complete recognition; and outputting a corresponding recognition signal by the neuromorphic chip, and transmitting the signal to the ground in a wireless manner, to reveal a recognition result. This method, inspired by the learning mechanism of the human brain, utilizes a training method based on meta-learning to enhance recognition accuracy for specific targets in special scenarios and reduces power consumption through neuromorphic computing.

[0036] In another exemplary embodiment of the present disclosure, to deploy the trained target recognition model on the neuromorphic chip for target recognition functionality, as shown in FIG. 2, the following step 301 to step 304 are included:

[0037] Step 301: Construct a neural network model with inner and outer loop coordination.

[0038] Step 302: Pre-train the neural network model using a few-shot learning method based on meta-learning, and import the trained neural network model into the neuromorphic chip.

[0039] Step 303: A camera of the micro UAV acquires a target image in real time, and transmits the target image to the neuromorphic chip for processing by the neural network model to complete recognition.

[0040] Step 304: The neuromorphic chip outputs a target recognition signal, and transmits the target recognition signal to the ground in a wireless manner to reveal a recognition result.

[0041] The neural network model with inner and outer loop coordination in step 301 includes a feature extraction model and inner and outer loop recognition models. In the present disclosure, a Convolutional Neural Network (CNN) and a Spiking Neural Network (SNN) are used to construct the neural network model with inner and outer loop coordination, which consists of one CNN model and two SNN models. The CNN model, as shown in Table 1, is used to extract features of the target in the image. Structurally, the present disclosure utilizes binary neuron models to build the CNN model, similar to the McCulloch-Pitts model (MP) but with a threshold changed to 1.2. The output of a neuron is as shown in formula (1) below:

[0042] s = 0(c — 1.2) (1); rl x >0

[0043] s is the output of the neuron; 0 represents a step function, that is, 0(x) = ; the variable c is a total weight of input values, which varies for each neuron. Structurally, in the present disclosure, the CNN model is divided into three layers with 16, 32, and 64 sets of filters, where the convolution kernel has a size of 3x3; valid padding is employed, with a stride of 1; and an average pooling layer has a size of 2. The input is an image, and the output is a binary feature sequence of the image.

[0044] It is should be noted that while ensuring the input and output dimensions to be consistent, the feature extraction model can use other structures to extract target features from the image.

[0045] Table 1: CNN Model Structure First layer Convolutional layer, with 16 sets of filters Batch normalization Second layer Convolutional layer, with 32 sets of filters Average pooling layer Batch normalization Third layer Convolutional layer, with 64 sets of filters Average pooling layer Batch normalization Fully connected layer

[0046] The two SNN models are a recognizing network (RN) and an optimizer network (ON): the recognizing network learns target features extracted by the feature extraction model and outputs corresponding results during recognition, where the recognizing network belongs to the inner loop of the model; the optimizer network provides optimization signals to the recognizing network, enabling the recognizing network to learn target features quickly, where the optimizer network belongs to the outer loop of the model. The entire model structure is depicted in FIG. 3.

[0047] In terms of the network structure, both the recognizing network and the optimizer network employ a recursive form known as the Recurrent Spiking Neural Network (RSNN). It has the following characteristics: the input is connected to each neuron, and the neurons are connected to each other without self-connections, creating a tightly connected RSNN. In terms of composition units, the recognizing network and the optimizer network both use a Leaky Integrate-and-Fire (LIF) neuron model and an Adaptive Leaky Integrate-and-Fire (ALIF) neuron model. In the SNN model, information is transmitted in the form of a sequence of spikes. For example, neuron j generates a spike at time t, denoted as 5- = 1. Further, the LIF neuron model has a hidden variable, namely, a membrane potential vf of neuron j at time t. The membrane potential is influenced not only by presynaptic neuron spikes but also by stimulus of input information, and decays exponentially with a membrane time constant Tm. When the membrane potential reaches a trigger threshold voltage, the neuron fires a spike and enters a refractory period. Considering the time for transmitting presynaptic spike sequences, a time constant d is used to represent the transmission delay, and 8t is a time step. Therefore, the LIF neuron model can be expressed as follows:

[0048] vf = Xvf1 + Si WA" xf + SiWycc sf d - 1 Vt / l (2);

[0049] s? = 0(vf - Vth) (3);

[0050] Wj1" represents a weight between input x- and neuron j, and W^ec represents a synaptic A weight from neuron i to neuron j. A = e Tm represents the influence of the membrane time constant on membrane potential decay, 0 still represents a step function, and Vth represents the threshold voltage. The ALIF neuron model has two hidden variables: an adaptive threshold a) in addition to the membrane potential v) of neuron j at time t. Introducing the adaptive threshold as a hidden variable provides the neuron with Spike-Frequency-Adaptation (SFA), preventing prolonged firing of the neuron: the adaptive threshold is influenced by the firing activity of the neuron. The adaptive threshold increases after the neuron generates a spike, to raise the threshold for subsequent spikes, and also decays with an adaptive time constant Ta. At each time point, the neuron generates a new value of a-, and this updated value, together with the default threshold voltage, forms an adapted trigger threshold voltage Aj. The ALIF neuron model can be expressed as follows:

[0051] s? = 0(v? - Aj) (4);

[0052] Aj = [Ba) + Vth (5);

[0053] a- = pa'1 + sf1 (6);

[0054] The coefficient p represents the degree of adaptation and has no specific physical meaning. 7 St p = e Ta represents the influence of the adaptive time constant on the decay of the adaptive threshold, s^1 denotes the output of neuron j at time t-1, and a--1 represents the adaptive threshold at time t-1.

[0055] In the human brain, the synaptic plasticity of certain learning regions is regulated by specific learning signals, such as dopamine modulation. Substances like dopamine are often produced in specific brain regions, such as the Ventral Tegmental Area (VTA) where dopamine is generated. These specific regions have been optimized during evolution to getter regulate learning, making rapid learning possible. Inspired by this working mechanism of the human brain, a collaborative relationship between the recognizing network and the optimizer network is established as follows: the recognizing network on the inner loop uses the sample feature sequence and label of the current image as input, outputs a recognition result, and feeds neural activity information of network neurons into the optimizer network; the optimizer network at the outer loop provides learning signals to the recognizing network based on the label, sample feature sequence, and neural activity information of the network neurons, to adjust the synaptic plasticity of the recognizing network.

[0056] Inspired by the synaptic plasticity mechanism in the human brain, an RSNN based on the LIF neuron model and the ALIF neuron model is established as the primary recognizing network, and an overall structure with inner and outer loop coordination is adopted. Additionally, the synaptic weight update method of the recognizing network has good biological interpretability and is a natural heuristic approach. The shallow CNN (trained feature extraction model) is used solely for feature extraction, to generate the target feature sequence. The spiking neural network employs spike-based encoding. In a short period of time when an image is recognized, numerous neurons in the network emit spikes at only a few moments to transmit information, significantly reducing hardware energy consumption and addressing the endurance issue of micro UAVs.

[0057] The trained target recognition model is trained using a few-shot learning method based on meta-learning. Before the target image is inputted into the trained target recognition model, the target recognition method for a micro UAV further includes a model training process, where the training process, i.e., step 302, can be replaced by the following step 201 to step 202:

[0058] Step 201: Train the feature extraction model and inner and outer loop recognition models by using a first sample set to obtain the trained feature extraction model and a trained optimizer network, where the first sample set includes a plurality of first sample images and labels corresponding to each first sample image; the inner and outer loop recognition models include a recognizing network and an optimizer network; the optimizer network provides learning signals to the recognizing network during training; both the recognizing network and the optimizer network are RSNNs; and a spiking neuron model in the RSNN is constructed from a LIF neuron model and an ALIF neuron model.

[0059] Step 202: Test the models obtained in step 201 by using a second sample set to obtain the trained recognizing network, where the trained recognizing network and the trained feature extraction model form the trained target recognition model; the models obtained in step 201 include the recognizing network, the trained feature extraction model, and the trained optimizer network; the second sample set includes a plurality of second sample images and labels corresponding to each second sample image; the number of second sample images is less than the number of first sample images in the first sample set.

[0060] In the training method, still inspired by the above human brain mechanisms, the synaptic plasticity of the recognizing network is regulated by the learning signal generated by the optimizer network as follows:

[0061] A14< = (7);

[0062] Lj represents the learning signal to neuron j at time t, and represents an eligibility trace of the synapse from neuron i to neuron j at time t. The eligibility trace reflects the influence of the weight Wjt on the firing of neuronj at time t, but only considers the dependency between neurons i and j, simplifying the calculation of information ttansmission between neurons, hj represents a hidden variable of neuron j at time t, and the eligibility trace is related to the partial derivative as? t However, since this partial derivative may not exist in the RSNN, a pseudo-derivative = 0.3 max 0,1 — Vth~ Vth is used instead. The eligibility trace of the LIF neuron model is shown as follows:

[0063] = (8);

[0064] sf = (9):

[0065] s- represents a low-pass filtered presynaptic spike sequence of neuron i. For input weights,sf d is replaced by xf. ALIF has two hidden variables, and the calculation of the eligibility ttace is more complex, as shown below:

[0066] = / ¾) (io);

[0067] = — + (11);

[0068] ji is a recursively calculated eligibility vector.

[0069] The entire training process adopts a few-shot learning method based on meta-leaming, divided into two stages: meta-learning and few-shot learning.

[0070] In the first stage, the meta-learning stage, the parameters learned are outer loop parameters. The training dataset uses a large number of first sample images relevant to the current recognition target, where each first sample image has a label ln. For a facial recognition task, multiple facial images of different individuals are used for training; for an equipment inspection task, images of relevant equipment in different states are used for training. These first sample images can be accumulated historically or from different regions, as long as they are relevant to the current recognition task.

[0071] Each experiment can be divided into two periods. In period 1, one first sample image is inputted and held for a duration of tc=20ms. During this period, the optimizer network generates learning signals based on input features and the neural activity of the recognizing network. At the end of tc, the recognizing network immediately updates the weight from the initial weight Winit according to formula (7). In period 2, the optimizer network no longer generates learning signals, and the recognizing network needs to make real-time judgments on the first sample images that appear, i.e., determining whether the first sample images belong to the same class as the images in period I. Five first sample images are sequentially inputted, with each first sample image still lasting tc=20ms. At the end of the input duration of each first sample image input, the recognizing network outputs an analog signal yout, which is the sample recognition signal. When the sample recognition signal yollt is greater than 0, the current first sample image belongs to the same class as the images in period 1; if the sample recognition signal y0llt is less than 0, it belongs to a different class. For a new experiment, the recognizing network starts again from the initial weight and goes through two periods sequentially. By setting an appropriate number of experiment batches, the learning effect is evaluated based on loss, thereby optimizing the outer loop parameters of the model. The loss function Ec is as follows:

[0072] Ec = i -lnlogcy(yout) -(1- In)log (1 - (j^y011^ (12);

[0073] ln is the label of the first sample image, which is a binary value where 1 indicates that the image belongs to the same class as the images in period 1. a represents a sigmoid activation function; m represents the number of batches, and n=l, 2, 3...m; each time the experiments form a batch, a BackPropagation Through Time (BPTT) algorithm is used to update the outer loop parameters, including the weight parameters fl of the CNN model and the optimizer network, as well as the initial weight Winit of the recognizing network. Once all training samples have been learned, one iteration is completed. Through continuous iterations, the CNN model can effectively extract target features, and the optimizer network can generate learning signals relevant to the recognition target, completing the learning of the first stage. The training method for the first stage is shown in FIG. 4.

[0074] The second stage is the few-shot learning stage, where the parameters learned are inner loop parameters. A small number of second sample images related to the current recognition target are used as training data, without any annotations. For example, in an equipment inspection task, images of fault conditions of the equipment are needed, but only a few or even just one second sample image is sufficient. During the training process, the weights fl of the CNN model and the optimizer io network, as well as the initial weight Winit of the recognizing network, remain unchanged, that is, the outer loop parameters remain unchanged. The current model includes the recognizing network, the trained feature extraction model, and the trained optimizer network. The training images are inputted and held for 20ms, during which the CNN model extracts target features from the images and inputs the sample feature sequence to the recognizing network. While the recognizing network receives the sample feature sequence, the internal LIF neurons and AL1F neurons are active. The optimizer network provides learning signals based on the spiking activity of the neurons in the recognizing network. At the end of the duration, the recognizing network immediately updates the network weight based on the learning signal and eligibility trace, and freezes the weight, to obtain a trained recognizing network. The trained recognizing network, combined with the trained feature extraction model, forms a trained target recognition model. At this point, the recognizing network has learned the features of the recognition target and can be used for target recognition and detection.

[0075] The entire training is divided into the two stages mentioned above, with the following special effects: the meta-learning stage enables the outer-loop network of the model to "learn how to learn," that is, learn a general rule or experience from many samples of the same class as the current recognition target to better guide the learning of specific recognition targets. The few-shot learning stage helps the inner-loop network of the model better fit the current recognition task. The general rule learned by the outer-loop network is used to guide the inner-loop network in learning a specific rule needed for the current task, which facilitates precise and efficient completion of the current task.

[0076] After the trained target recognition model is obtained, the trained target recognition model is imported into a neuromorphic chip, and the neuromorphic chip is embedded into the micro UAV. A target image of the to-be-recognized target is then collected, and the trained target recognition model in the neuromorphic chip identifies the target image to obtain a target recognition signal. After the identification is completed, the neuromorphic chip outputs the target recognition signal, and transmits the target recognition signal to the ground device.

[0077] In another exemplary embodiment of the present disclosure, the target recognition method for a micro UAV includes:

[0078] decoding and decompressing the target recognition signal by the ground device to obtain a decoded result; and

[0079] displaying the decoded result.

[0080] The ground receiver (ground device) decodes and decompresses the received digital signal (target recognition signal) and displays the decoded result on the ground device.

[0081] The present disclosure also provides an application scenario where the target recognition method for a micro UAV is utilized. Specifically, the target recognition method for a micro UAV in this embodiment can be applied in a scenario of criminal facial recognition for arrests. The scenario of criminal facial recognition for arrests includes image acquisition, image recognition, and arrest phases; the target image goes from image acquisition to image recognition, where the trained target recognition model in the neuromorphic chip on the micro UAV performs facial recognition to obtain a target recognition signal, and transmits the target recognition signal to ground personnel for arrest. The target recognition method for a micro UAV provided in this embodiment falls under the image recognition phase. During the target recognition process for the target image, the trained target recognition model in the neuromorphic chip on the micro UAV is used for target recognition, to obtain the target recognition signal corresponding to the target image.

[0082] Based on the same inventive conception, an embodiment of the present disclosure further provides a target recognition system for a micro UAV, for implementing the aforementioned target recognition method for a micro UAV. The implementation solution provided by the target recognition system for a micro UAV addresses the same issues as the implementation solution described in the above method. Therefore, for the specific limitations in one or more embodiments of the target recognition system for a micro UAV provided below, reference can be made to the limitations for the target recognition method for a micro UAV described above, and details are not described herein again.

[0083] In an exemplary embodiment, as shown in FIG. 5, a target recognition system for a micro UAV is provided. The target recognition system for a micro UAV includes an acquisition module, a recognition module, and a wireless transmission module.

[0084] The acquisition module is configured to acquire a target image of a to-be-recognized target.

[0085] The recognition module is a neuromorphic chip; a trained target recognition model is deployed on the neuromorphic chip; the trained target recognition model is configured to identify the target image to obtain a target recognition signal; the trained target recognition model includes a trained feature extraction model and a trained recognizing network; the trained recognizing network is an RSNN; a spiking neuron model in the RSNN is constructed from a LIF neuron model and an adaptive LIF neuron model.

[0086] The wireless transmission module is configured to transmit the target recognition signal to a ground device.

[0087] As an optional implementation, the target recognition system for a micro UAV further includes a flight control module configured to control flight operations of the micro UAV

[0088] The ground device is configured to decode and decompress the target recognition signal to obtain a decoded result.

[0089] The ground device includes a display unit configured to display the decoded result.

[0090] In this implementation, after the trained CNN model and recognizing network are imported into the neuromorphic chip, the neuromorphic chip is embedded inside the micro UAV. The target recognition system for a micro UAV includes an acquisition module, a recognition module, a wireless transmission module, and a flight control module. The acquisition module includes a camera and an image processing chip; the recognition module is the embedded neuromorphic chip; the wireless transmission module includes Wi-Fi wireless transmission and 2.4GHz wireless transmission; the flight control module includes a motor, a rotor, and other flight hardware.

[0091] When the neuromorphic chip is embedded, the image processing chip of the acquisition module needs to be connected to an input end of the neuromorphic chip so that the neuromorphic chip can identify the acquired target image. An output signal end of the neuromorphic chip is connected to the wireless transmission module so that the output signal can be transmitted back to the ground. This step completes the connection of the internal modules of the micro UAV.

[0092] The micro UAV performs recognition tasks for specific targets. Controlled by ground personnel, the micro UAV flies to a task area to perform the task of recognizing a specific target. A control signal from the ground personnel is transmitted in the 2.4GHz frequency band and modulated by a flight remote control device with a dedicated communication protocol before being sent to the wireless transmission module of the micro UAV. The signal is demodulated and used to control the flight behavior of the micro UAV. When the micro UAV captures images of a target in a scene or an area, it hovers briefly to ensure stable video streaming and clear input images. During this period, the camera captures images in real time, and the image data is sent to the neuromorphic chip via the image processing chip. The neuromorphic chip receives the input images, extracts target features in the images by using the CNN model, and the trained recognizing network identifies the target feature sequence. After processing for a period of time, the neuromorphic chip outputs an analog signal, namely, the target recognition signal. During the recognition process, if the analog signal (the target recognition signal) is greater than 0, it is considered that the currently recognized target belongs to the same class as or is similar to the desired target; if the analog signal is less than 0, features of the currently recognized target differ significantly from those of the desired target. The entire recognition process, from capturing image information to outputting the recognition signal, typically lasts 1 to 2 seconds.

[0093] It should be noted that the method of transmitting the target recognition signal back to the ground device can use a different channel from the control signal, as long as they do not interfere or conflict with each other.

[0094] The analog signal output by the neuromorphic chip is transmitted back to the ground in real time via the wireless transmission module. The ground personnel can take appropriate actions based on the recognition result of the micro UAV. After the neuromorphic chip outputs the corresponding recognition signal for an image, this signal is modulated by the wireless transmission module and transmitted back to the ground using the 2.4GHz frequency band. To ensure that the recognition signal does not interfere with the flight control signal, different channels, communication protocols, and modulation methods are used for the two signals. Compared to higher frequency bands like 5.8GHz, the 2.4GHz frequency band has better penetration capabilities and longer transmission distances, making it more suitable for remote signal transmission through walls and obstacles, which is advantageous for applications requiring signal transmission over a large area. Additionally, to assist ground personnel in making further judgments, the wireless transmission module in the micro UAV can also transmit real-time captured images back to the ground using Wi-Fi wireless transmission technology. Wi-Fi is a digital communication protocol primarily used for transmitting digital data such as images and videos. In this process, the acquisition module converts image data into digital signals and sends the digital signals to the wireless transmission module, which encodes and compresses the data to reduce the data amount and bandwidth requirements. The Wi-Fi module establishes a connection between the micro UAV and the ground by connecting to a wireless router or directly to the Wi-Fi receiver on the ground device. The ground receiver decodes and decompresses the received digital signal for display on the ground device.

[0095] The ground Wi-Fi receiver used to receive image data transmitted by the micro UAV can be a dedicated image transmission receiver or a Wi-Fi receiver connected to a computer, mobile phone, or tablet via Wi-Fi.

[0096] The technical characteristics of the above embodiments can be employed in arbitrary combinations. To provide a concise description of these embodiments, all possible combinations of all the technical characteristics of the above embodiments may not be described; however, these combinations of the technical characteristics should be construed as falling within the scope defined by the specification as long as no contradiction occurs.

[0097] Several examples are used herein for illustration of the principles and implementations of the present disclosure. The description of the foregoing examples is used to help illustrate the method of the present disclosure and the core principles thereof. In addition, those of ordinary skill in the art can make various modifications in terms of specific implementations and scope of application in accordance with the teachings of the present disclosure. In conclusion, the content of the present specification shall not be construed as a limitation to the present disclosure.

Claims

1. A target recognition method for a micro unmanned aerial vehicle (UAV), comprising:obtaining a target image of a to-be-recognized target, wherein the target image is an image of the to-be-recognized target captured by an image acquisition device of a micro UAV;inputting the target image to a trained target recognition model to obtain a target recognition signal, wherein the trained target recognition model is deployed on a neuromorphic chip of the micro UAV; the trained target recognition model comprises a trained feature extraction model and a trained recognizing network; the trained recognizing network is a Recurrent Spiking Neural Network (RSNN); a spiking neuron model in the RSNN is constructed from a Leaky Integrate-and-Fire (LIF) neuron model and an Adaptive Leaky Integrate-and-Fire (ALIF) neuron model; andtransmitting the target recognition signal to a ground device.

2. The target recognition method for a micro UAV according to claim 1, wherein the trained target recognition model is trained using a few-shot learning method based on meta-learning.

3. The target recognition method for a micro UAV according to claim 2, wherein before inputting the target image into the trained target recognition model, the target recognition method for a micro UAV further comprises:step 201: training the feature extraction model and inner and outer loop recognition models by using a first sample set, to obtain the trained feature extraction model and a trained optimizer network, wherein the first sample set comprises a plurality of first sample images and labels corresponding to each first sample image; the inner and outer loop recognition models comprise a recognizing network and an optimizer network; the optimizer network provides learning signals to the recognizing network during training; both the recognizing network and the optimizer network are RSNNs; and a spiking neuron model in the RSNN is constructed from a LIF neuron model and an ALIF neuron model; andstep 202: testing the models obtained in step 201 by using a second sample set to obtain the trained recognizing network, wherein the trained recognizing network and the trained feature extraction model form the trained target recognition model; the models obtained in step 201 comprise the recognizing network, the trained feature extraction model, and the trained optimizer network; the second sample set comprises a plurality of second sample images and labels corresponding to each second sample image; the number of second sample images is less than the number of first sample images in the first sample set.

4. The target recognition method for a micro UAV according to claim I, wherein the trainedfeature extraction model is a Convolutional Neural Network (CNN) model.

5. The target recognition method for a micro UAV according to claim 1, wherein said inputting the target image to the trained target recognition model to obtain the target recognition signal specifically comprises:inputting the target image to the trained feature extraction model to obtain a target feature sequence; andinputting the target feature sequence to the trained recognizing network to obtain the target recognition signal.

6. The target recognition method for a micro UAV according to claim 1, further comprising:decoding and decompressing the target recognition signal by the ground device to obtain a decoded result; anddisplaying the decoded result.

7. A target recognition system for a micro unmanned aerial vehicle (UAV), for implementing the target recognition method for a micro UAV according to claim 1, wherein the target recognition system for a micro UAV comprises an acquisition module, a recognition module, and a wireless transmission module;the acquisition module is configured to acquire a target image of a to-be-recognized target;the recognition module is a neuromorphic chip; a trained target recognition model is deployed on the neuromorphic chip; the trained target recognition model is configured to identify the target image to obtain a target recognition signal; the trained target recognition model comprises a trained feature extraction model and a trained recognizing network; the trained recognizing network is a Recurrent Spiking Neural Network (RSNN); a spiking neuron model in the RSNN is constructed from a Leaky Integrate-and-Fire (LIF) neuron model and an Adaptive Leaky Integrate-and-Fire (ALIF) neuron model; andthe wireless transmission module is configured to transmit the target recognition signal to a ground device.

8. The target recognition system for a micro UAV according to claim 7, further comprising a flight control module configured to control flight operations of the micro UAV.

9. The target recognition system for a micro UAV according to claim 7, wherein the ground device is configured to decode and decompress the target recognition signal to obtain a decoded result.

10. The target recognition system for a micro UAV according to claim 9, wherein the ground device comprises a display unit configured to display the decoded result.

Images