An unmanned aerial vehicle abnormal behavior recognition method based on an improved residual network

Abstract

This invention discloses a method for identifying abnormal behavior of unmanned aerial vehicles (UAVs) based on an improved residual network, belonging to the field of UAV anomaly recognition technology. The method includes: hierarchical sampling and preprocessing of a multimodal image set based on the UAV's physical, functional, and environmental attributes to obtain a training dataset; improving the residual network model architecture and adding a double-constraint center loss mechanism; performing phased incremental training of the improved residual network model using the training dataset; extracting deep, medium, and shallow features from real-time images based on the trained residual network model, and outputting image recognition results after initial screening and classification; calculating the degree of anomalous behavior of the UAV based on the image recognition results of N consecutive frames, and triggering alarms and real-time tracking modes when the anomalousness exceeds a threshold. This method solves the problems of model degradation, gradient vanishing, and poor local anomaly detection performance in traditional methods.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) anomaly recognition technology, and in particular to a method for recognizing abnormal UAV behavior based on an improved residual network. Background Technology

[0002] In today's era of rapid technological development, drone technology has made remarkable progress, and its application areas are constantly expanding.

[0003] However, behind the rapid development of drones, operational safety issues have gradually become prominent. Drones mainly perform flight missions in low-altitude airspace, which is an extremely complex environment with various obstacles, unstable airflow, and electromagnetic interference. At the same time, drones themselves also have certain limitations, such as insufficient endurance and weak anti-interference capabilities. All of these factors make drones prone to abnormal behavior when performing missions.

[0004] To address these issues, researchers have proposed various methods for identifying abnormal drone behavior. Zhang Xuejun et al. proposed a method based on convolutional neural networks (CNNs), which imports historical drone flight data into a CNN for training and then uses new data to test and predict abnormal behavior, achieving automated identification. However, the traditional CNNs used in this method have significant drawbacks. Drone flight is a dynamic process, and the actual captured images vary in state. To obtain accurate classification results, precise image features need to be extracted, which requires deep network training on a massive dataset. However, traditional CNNs are slow to train on large datasets, consuming a lot of time. Furthermore, some drones have similar shapes, requiring deeper learning to extract more features to ensure image classification accuracy. However, traditional CNNs encounter problems such as model degradation and gradient vanishing during deep network training, leading to a decrease in the accuracy of the output results, making it difficult to meet the actual needs of drone image classification. Tang Li et al. proposed an intelligent identification method for abnormal behavior of logistics drones based on the isolated forest method. This method identifies anomalies by calculating outliers in drone flight data, constructing an isolated forest model to calculate anomaly scores, classifying abnormal data, and evaluating accuracy. However, this method has limitations in local anomaly detection and performs poorly when dealing with datasets with complex distributions.

[0005] Given the numerous problems existing in the current technology for identifying abnormal behavior of drones, there is an urgent need to develop a more efficient and accurate identification method to meet the pressing needs of ensuring airspace safety and promoting the healthy development of the drone industry. Summary of the Invention

[0006] This invention provides a method for identifying abnormal behavior of unmanned aerial vehicles (UAVs) based on an improved residual network, comprising:

[0007] Step 1: Based on the physical, functional, and environmental attributes of the UAV, perform stratified sampling and preprocessing of the multimodal image set to obtain the training dataset;

[0008] Step 2: Improve the residual network model architecture and add a double-constraint center loss mechanism to it;

[0009] Step 3: Use the training dataset to perform phased incremental training on the improved residual network model;

[0010] Step 4: Extract deep, medium and shallow features of the real-time image based on the trained residual network model, and output the image recognition results after initial screening and classification;

[0011] Step 5: Calculate the degree of abnormality of the drone's behavior based on the image recognition results of N consecutive frames, and trigger an alarm and real-time tracking mode when the degree of abnormality exceeds the threshold.

[0012] The UAV abnormal behavior recognition method based on improved residual networks, as described above, involves hierarchical sampling and preprocessing of a multimodal image set based on the UAV's physical, functional, and environmental attributes. Specifically, this involves the following sub-steps:

[0013] The system acquires UAV images in three modes: visible light, infrared, and thermal imaging. It also records the three-dimensional attribute data of the UAV during flight, quantizes it, and associates it with the attribute labels of the UAV images to form a multimodal image set.

[0014] The multimodal image set is mapped to a three-dimensional hierarchical space for initial layering;

[0015] A predetermined number of samples are extracted from the initial strata based on the differences in model training value among different strata.

[0016] The extracted samples are standardized and preprocessed to form a training dataset.

[0017] The above-described method for identifying abnormal UAV behavior based on improved residual networks involves the following sub-steps:

[0018] Replace each feature extraction convolution in the original residual network model with a deformable residual fusion unit;

[0019] A dual-channel attention module is inserted after each deformable residual fusion unit to generate feature weights;

[0020] The distribution of features learned by the model in space is optimized by adding a double-constraint center loss mechanism.

[0021] The above-described method for identifying abnormal drone behavior based on improved residual networks involves staged incremental training of the improved residual network model using a training dataset, specifically comprising the following sub-steps:

[0022] A stochastic gradient descent optimizer with momentum is used to perform basic training on the residual network model, enabling it to learn general and basic UAV features;

[0023] Once the loss of the double-constraint center stops decreasing, switch to the adaptive moment estimation optimizer for fine-tuning training, enabling it to learn to distinguish fine-grained categories.

[0024] The validation set accuracy is monitored in real time. If the validation set accuracy does not improve within 10 consecutive training rounds, the early stop mechanism is triggered to automatically terminate the training.

[0025] The above-described method for identifying abnormal UAV behavior based on improved residual networks involves extracting deep, medium, and shallow features from real-time images using a trained residual network model. After initial screening and classification, the image recognition result is output. The method comprises the following sub-steps:

[0026] The real-time acquired drone images are processed into a format that the residual network model can recognize and then input into the model. Feature maps are extracted from the first, third, and last deformable residual fusion units of the network, corresponding to shallow, medium, and deep features, respectively.

[0027] The extracted shallow features are used for initial screening. If the initial screening passes, the three-layer features are stitched together. If the initial screening fails, the system returns "No drone target identified" and terminates the subsequent process.

[0028] The concatenated three-layer features are input into the classifier, which outputs the classification label and its corresponding confidence score.

[0029] If the confidence level is higher than the threshold, the classification label is used as the image recognition result; if it is lower than the threshold, an arbitration mechanism is initiated to determine the final image recognition result.

[0030] The present invention also provides an abnormal behavior recognition system for unmanned aerial vehicles based on an improved residual network, comprising: a training dataset construction module, a residual network improvement module, a residual network training module, a real-time image recognition module, and an anomaly determination module;

[0031] The training dataset construction module is used to perform hierarchical sampling and preprocessing of multimodal image sets based on the physical, functional, and environmental attributes of the UAV to obtain the training dataset;

[0032] The residual network improvement module is used to improve the architecture of the residual network model and add a double-constraint center loss mechanism to it.

[0033] The residual network training module is used to perform phased incremental training on the improved residual network model using the training dataset.

[0034] The real-time image recognition module is used to extract deep, medium and shallow features of real-time images based on a trained residual network model, and output image recognition results after initial screening and classification.

[0035] The anomaly detection module is used to calculate the degree of anomaly in the drone's behavior based on the image recognition results of N consecutive frames, and to trigger an alarm and real-time tracking mode when the degree of anomaly exceeds the threshold.

[0036] The beneficial effects achieved by this invention are as follows: it solves the problems of model degradation, gradient vanishing, and poor local anomaly detection in traditional methods; it improves the residual network to adapt to UAV deformation, optimizes feature distribution through dual-constraint loss, and enhances robustness through phased training; it improves the accuracy of UAV anomaly identification, reduces false alarm rate, adapts to complex environments, and ensures airspace safety. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0038] Figure 1 This is a flowchart of an abnormal behavior recognition method for unmanned aerial vehicles based on an improved residual network, provided in Embodiment 1 of this application. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] like Figure 1 As shown, Embodiment 1 of this application provides a method for identifying abnormal behavior of unmanned aerial vehicles (UAVs) based on an improved residual network, including:

[0042] Step S110: Based on the physical, functional, and environmental attributes of the UAV, perform hierarchical sampling and preprocessing of the multimodal image set to obtain the training dataset;

[0043] This step focuses on "drone attribute-driven data optimization," overcoming the limitations of existing technologies that "use image data indiscriminately." Through stratified sampling and preprocessing, it achieves a triple improvement in training data's "sample balance, scene adaptability, and feature effectiveness," providing a high-quality data foundation for subsequent improvements to the residual network model. Specifically, it consists of the following sub-steps:

[0044] Step S111: Collect UAV images in three modes: visible light, infrared, and thermal imaging. Simultaneously record the three-dimensional attribute data of the UAV during flight. After quantization, associate the data with the attribute labels of the UAV images to form a multimodal image set.

[0045] A joint UAV test site was established, constructing a data acquisition network of "fixed ground acquisition stations + mobile aerial acquisition drones" to simultaneously record the three-dimensional attribute data of the UAVs during flight. The physical attribute P describes the static appearance characteristics of the UAV, quantified into a discrete set of values, such as 0: ultra-small / toy-sized, 1: small, 2: medium, 3: large. The functional attribute F describes the dynamic behavioral intent exhibited by the UAV when captured, also quantified into a discrete set of values, such as 0: hovering, 1: smooth cruise, 2: acceleration / sprint, 3: irregular maneuvering, 4: landing. The environmental adaptation attribute describes the external environmental conditions during image acquisition, directly affecting image quality and feature visibility, and is quantified into a continuous value. This is calculated using an environmental complexity scoring function S(E): S(E) = ω1×(1-I_Normalized) + ω2×B_Entropy + ω3×W_score + ω4×O_Ratio

[0046] Where ω1, ω2, ω3, and ω4 are adjustable environmental weight coefficients, and I_Normalized, B_Entropy, W_score, and O_Ratio are the light intensity, background clutter, weather score, and occlusion degree at the time of image capture, respectively.

[0047] For a set of multimodal images, the associated attribute labels are represented as (P,F,S(E)).

[0048] Step S112: Map the multimodal image set to a three-dimensional hierarchical space for initial layering;

[0049] Using the three attributes mentioned above as coordinate axes in a three-dimensional space (P-axis, F-axis, E-axis), the entire multimodal image set is mapped onto this space. Subsequently, each dimension is divided into intervals according to business requirements, thus dividing the entire space into N. p ×N f ×N e There are several levels, of which N p N f N eThe number of intervals divided for the P-axis, F-axis, and E-axis respectively; each image is assigned to a unique level based on its associated attribute label.

[0050] Step S113: Extract a predetermined number of samples from the initial strata based on the differences in the training value of different strata for the model;

[0051] The training value of the i-th level for the model is W. i The quantification formula is expressed as:

[0052]

[0053] Where N i N is the number of images in the i-th level. total S(E) represents the total number of images in the entire three-dimensional space. i It is the average environmental complexity of the i-th level, α and β are adjustable parameters, and cosθ jk It is the angle between the feature vectors of the j-th sample and the k-th sample, where j takes values ​​from 1 to N. i k takes values ​​from j+1 to N i .

[0054] Let the preset number of samples be N. extract Then the number of samples to be drawn from the i-th level is

[0055] Step S114: Standardize and preprocess the extracted samples to form a training dataset;

[0056] Specifically targeting images with high environmental complexity, we employ data augmentation strategies such as more significant brightness / contrast adjustments, simulated rain and fog noise, and random occlusion to further enhance the model's robustness under harsh conditions through adversarial enhancement. We adjust all images to the network input size and normalize pixel values ​​to [0,1] or perform mean-variance standardization.

[0057] Step S120: Improve the residual network model architecture and add a double-constraint center loss mechanism to it;

[0058] Based on ResNet-50 as the framework, this embodiment makes the following improvements to address the difficulty of feature extraction caused by deformation and occlusion during high-speed maneuvers of UAVs:

[0059] Step S121: Replace each feature extraction convolution in the original residual network model with a deformable residual fusion unit;

[0060] ResNet-50 uses a large number of 3x3 convolutions for feature extraction, but the sampling positions of traditional convolution kernels on the input image are fixed and regular (a 3x3 convolution kernel samples at 9 fixed grid points), which cannot adapt to the drastic geometric deformations generated by drones in flight. Therefore, this embodiment replaces it with a deformable residual fusion unit. This unit has two parallel network branches and an adaptive fusion module. The two parallel network branches are a deformable convolution branch and a separable convolution branch. The deformable convolution branch uses a 3x3 deformable convolution. For each input feature map, this branch learns the offset Δp of each sampling point through a parallel light quantum network. The convolution kernel samples at the new position of the original position + Δp, thus actively "deforming" the convolution kernel to fit the actual shape of the drone in the current image. The separable convolution branch decomposes the original standard 3x3 convolution into two consecutive steps:

[0061] Depthwise convolution: uses a 1*3 convolution kernel, with each channel convolved individually, responsible for processing spatial information (vertical relationships);

[0062] Pointwise convolution: uses a 3*1 convolution kernel to process channel information and fuse features from different channels.

[0063] The adaptive fusion module is used to fuse the outputs of multiple network branches using learnable adaptive weights.

[0064] Step S122: Insert a dual-channel attention module after each deformable residual fusion unit to generate feature weights;

[0065] The dual-channel attention module consists of a feature enhancement channel and a spatial enhancement channel connected in series.

[0066] The workflow of the feature enhancement channel is as follows: global average pooling and global max pooling are performed on the feature map output by the deformable residual fusion unit to obtain two 1*1*C vectors, where C is the number of channels in the corresponding stage; the two vectors are input into the shared fully connected network to generate two weight vectors M_avg and M_max; the weight vector M_channel of the feature enhancement channel is then output using the formula sigmoid(M_avg+M_max); finally, M_channel is multiplied with the input feature map channel by channel to obtain the feature map enhanced by the feature enhancement channel.

[0067] The workflow of the spatial augmentation channel is as follows: Channel average pooling and channel max pooling are performed on the feature maps augmented by the feature augmentation channel to obtain two H*W*1 feature maps, where H*W is the resolution of the original feature maps. These two feature maps are then concatenated along the channel dimension to form an H*W*2 feature map. A 7*7 large-size convolution is then used to extract long-range spatial correlation features from the H*W*2 feature map. This output feature map is then activated by a sigmoid function to obtain the spatial augmentation channel weight map. Finally, the spatial augmentation channel weight map is multiplied element-wise with the feature maps augmented by the feature augmentation channel to obtain the spatially augmented feature map, which is the final output of the dual-channel attention module.

[0068] Step S123: Optimize the distribution of features learned by the model in space by adding a double-constraint center loss mechanism;

[0069] The dual-constraint center loss mechanism allows the features learned by the model to be both discriminative and robust. Its core function is expressed as:

[0070]

[0071] Where L total This represents the total loss, where γ1, γ2, and γ3 are adjustable weighting coefficients, and L... ce It is the cross-entropy loss, where B is the total number of samples in a training batch, and x b y is the feature vector of the b-th sample. b It is the true class label of the b-th sample. It is category y b The center vector, b takes values ​​from 1 to B, and K is the total number of categories. c j c v These are the center vectors of the j-th and v-th classes, respectively, cos(c j ,c v ) is a vector c j and c v The cosine similarity is given by j and v, where j and v take values ​​from 1 to K, and v ≠ j.

[0072] when When updating category y b center vector for To ensure the accuracy of the classification centers, η is the learning rate.

[0073] Step S130: Perform phased incremental training on the improved residual network model using the training dataset;

[0074] The core of segmented incremental training lies in first allowing the model to learn general, basic drone features, and then guiding it to finely distinguish special functions and difficult samples, thereby avoiding training instability, overfitting, or model degradation caused by learning all concepts at once. The specific implementation process is as follows:

[0075] Step S131: Use the stochastic gradient descent (SGD) optimizer with momentum to perform basic training on the residual network model, so that it learns general and basic UAV features;

[0076] Samples containing basic types of UAVs (such as fixed-wing, multi-rotor, helicopter, etc.) are selected from the training dataset, and data augmentation is performed using standard random horizontal flipping, random cropping, and color jittering. The augmented samples are then divided into training and validation sets in an 8:2 ratio. Training begins using the training set, employing a step-by-step decay strategy. Every 20 epochs (rounds), the learning rate is decayed to 1 / 10 of the previous stage, for a total of 100 training rounds. After each epoch, the double-constraint center loss value is calculated and recorded on the validation set to monitor the training process.

[0077] Step S132: After the double-constraint center loss stops decreasing, switch to the adaptive moment estimation optimizer (Adam) for fine-tuning training so that it learns to distinguish fine-grained categories;

[0078] All training data, including samples from special-purpose drones (such as agricultural drones, surveying drones, and manned drones), are incorporated and divided into training and validation sets in an 8:2 ratio. Building upon the previous stage, a more effective regularization strategy is added:

[0079] Label Smoothing: Using a formula The label distribution is redistributed, where q'(j|x) is the corrected probability of the input sample x being predicted as the j-th class label after label smoothing, q(j|x) is the original probability, ε is the smoothing factor, and K is the total number of classes. This can reduce overfitting and improve the model's generalization ability.

[0080] Dropout: A Dropout layer is enabled before the fully connected layers of the network with a random dropout rate of 0.3 to suppress complex co-adaptation of neurons.

[0081] Switch to the Adam optimizer to begin training, using a cosine annealing strategy. Over 50 training epochs, the learning rate is slowly decayed from its initial value to near 0, allowing the model to converge to a better local optimum.

[0082] Step S133: Monitor the validation set accuracy in real time. If the validation set accuracy does not improve within 10 consecutive training rounds, trigger the early stop mechanism and automatically terminate the training.

[0083] Throughout the training process, a copy of the model parameters that perform best on the validation set is always maintained in memory. When training is stopped early or ends normally, this copy is saved as the final model, ensuring that the delivered model is the best performing model, not the model from the last potentially overfitted epoch.

[0084] Step S140: Extract deep, medium and shallow features of the real-time image based on the trained residual network model, and output the image recognition result after initial screening and classification;

[0085] The real-time acquired drone images are processed into a format recognizable by the residual network model and then input into the model. Feature maps are extracted from the first, third, and last deformable residual fusion units of the network, corresponding to shallow, medium, and deep features, respectively. The initial screening and secondary classification are specifically divided into the following sub-steps:

[0086] Step S141: Perform preliminary screening based on the extracted shallow features. If the preliminary screening passes, perform three-layer feature splicing. If it fails, return "No drone target identified" and terminate the subsequent process.

[0087] Calculate the cosine similarity between the shallow texture feature vector of the real-time image and the center vector of the shallow features of all known categories in the sample library; if the maximum value of all cosine similarities is less than 0.85 (threshold adjustable), it is determined that "no drone target was identified" in the image and the process terminates; if it is greater than 0.85, three-layer feature stitching is performed.

[0088] Step S142: Input the concatenated three-layer features into the classifier and output the classification label and its corresponding confidence score;

[0089] Step S143: If the confidence level is higher than the threshold, the classification label is used as the image recognition result; if it is lower than the threshold, the arbitration mechanism is activated to determine the final image recognition result.

[0090] The arbitration mechanism's decision-making process is as follows: calculate the Euclidean distance between the three layers of features of the image and the corresponding layer center vectors of each category in the sample library; assign weights to the distances calculated for the deep, middle, and shallow features, and sum them to obtain the total distance for each candidate category; select the category with the smallest total distance as the final decision result.

[0091] Step S150: Calculate the degree of abnormality of the drone behavior based on the image recognition results of N consecutive frames, and trigger an alarm and real-time tracking mode when the degree of abnormality exceeds the threshold;

[0092] First, based on the image recognition results of N consecutive frames (set as needed), the proportion P of abnormal frames is statistically determined. eb Average confidence level C of abnormal frames ea and inter-frame behavior abrupt change coefficient T s Then, substitute the statistically obtained quantities into the formula:

[0093] S e =μ1×P eb +μ2×C ea +μ3×T s In the process, the degree of abnormality S of the drone's behavior is obtained. e Where μ1, μ2, and μ3 are adjustable weighting coefficients; if S e If the threshold is exceeded, an alarm and real-time tracking mode will be triggered immediately. Real-time tracking mode refers to using Kalman filtering to predict the coordinates of the next frame based on the drone's position in the current frame, and continuously locking onto the target using visible light and infrared images for subsequent target processing; if S e If the threshold is not exceeded, continue monitoring the next set of N consecutive frames of data.

[0094] Example 2

[0095] Embodiment 2 of this application provides an abnormal behavior recognition system for unmanned aerial vehicles based on an improved residual network, including: a training dataset construction module, a residual network improvement module, a residual network training module, a real-time image recognition module, and an anomaly determination module;

[0096] (1) Training dataset construction module, used to perform hierarchical sampling and preprocessing of multimodal image sets based on UAV physical attributes, functional attributes and environmental attributes to obtain training dataset;

[0097] (2) Residual network improvement module, which is used to improve the architecture of the residual network model and add a double-constraint center loss mechanism to it;

[0098] (3) Residual network training module, used to perform phased incremental training on the improved residual network model using the training dataset;

[0099] (4) Real-time image recognition module, which is used to extract deep, medium and shallow features of real-time images based on the trained residual network model, and output image recognition results after initial screening and classification;

[0100] (5) Anomaly detection module, used to calculate the degree of anomaly of the UAV behavior based on the image recognition results of N consecutive frames, and to trigger an alarm and real-time tracking mode when the degree of anomaly exceeds the threshold.

[0101] Corresponding to the above embodiments, the present invention provides a computer storage medium, including: at least one memory and at least one processor;

[0102] The memory is used to store one or more program instructions;

[0103] A processor is used to run one or more program instructions to execute a method for identifying anomalous behavior of unmanned aerial vehicles based on an improved residual network.

[0104] Corresponding to the above embodiments, this embodiment of the invention provides a computer-readable storage medium containing one or more program instructions, which are executed by a processor to provide a method for identifying abnormal behavior of unmanned aerial vehicles based on an improved residual network.

[0105] The embodiments disclosed in this invention provide a computer-readable storage medium storing computer program instructions. When the computer program instructions are executed on a computer, the computer executes the above-described method for identifying abnormal behavior of unmanned aerial vehicles based on an improved residual network.

[0106] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0107] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0108] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0109] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0110] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0111] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0112] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0113] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for recognizing abnormal behavior of a UAV based on an improved residual network, characterized in that, include: Step 1: Based on the physical, functional, and environmental attributes of the UAV, perform stratified sampling and preprocessing of the multimodal image set to obtain the training dataset; Step 2: Improve the residual network model architecture and add a double-constraint center loss mechanism to it; Step 3: Use the training dataset to perform phased incremental training on the improved residual network model; Step 4: Extract deep, medium and shallow features of the real-time image based on the trained residual network model, and output the image recognition results after initial screening and classification; Step 5: Calculate the degree of abnormality of the drone's behavior based on the image recognition results of N consecutive frames, and trigger an alarm and real-time tracking mode when the degree of abnormality exceeds the threshold. The improvement of the residual network model architecture is divided into the following sub-steps: Replace each feature extraction convolution in the original residual network model with a deformable residual fusion unit; A dual-channel attention module is inserted after each deformable residual fusion unit to generate feature weights; The distribution of features learned by the model in space is optimized by adding a double-constraint center loss mechanism. 2.The UAV abnormal behavior recognition method based on improved residual network according to claim 1, characterized in that, Based on the physical, functional, and environmental attributes of the UAV, hierarchical sampling and preprocessing of the multimodal image set are performed, specifically in the following sub-steps: The system acquires UAV images in three modes: visible light, infrared, and thermal imaging. It also records the three-dimensional attribute data of the UAV during flight, quantizes it, and associates it with the attribute labels of the UAV images to form a multimodal image set. The multimodal image set is mapped to a three-dimensional hierarchical space for initial layering; A predetermined number of samples are extracted from the initial strata based on the differences in model training value among different strata. The extracted samples are standardized and preprocessed to form a training dataset. 3.The UAV abnormal behavior recognition method based on improved residual network according to claim 2, characterized in that, The improved residual network model is trained incrementally in stages using the training dataset, specifically in the following sub-steps: A stochastic gradient descent optimizer with momentum is used to perform basic training on the residual network model, enabling it to learn general and basic UAV features; Once the loss of the double-constraint center stops decreasing, switch to the adaptive moment estimation optimizer for fine-tuning training, enabling it to learn to distinguish fine-grained categories. The validation set accuracy is monitored in real time. If the validation set accuracy does not improve within 10 consecutive training rounds, the early stop mechanism is triggered to automatically terminate the training. 4.The UAV abnormal behavior recognition method based on improved residual network according to claim 3, characterized in that, Based on the trained residual network model, deep, medium, and shallow features of real-time images are extracted. After initial screening and classification, the image recognition results are output. The process is divided into the following sub-steps: The real-time acquired drone images are processed into a format that the residual network model can recognize and then input into the model. Feature maps are extracted from the first, third, and last deformable residual fusion units of the network, corresponding to shallow, medium, and deep features, respectively. The initial screening is performed based on the extracted shallow features. If the initial screening passes, the three-layer features are stitched together. If the initial screening fails, the system returns "No drone target identified" and terminates the subsequent process. The concatenated three-layer features are input into the classifier, which outputs the classification label and its corresponding confidence score. If the confidence level is higher than the threshold, the classification label is used as the image recognition result; if it is lower than the threshold, an arbitration mechanism is initiated to determine the final image recognition result. 5.The UAV abnormal behavior recognition method based on improved residual network according to claim 4, characterized in that, The arbitration process is as follows: Calculate the Euclidean distance between the three-layer features of the image and the corresponding layer center vectors of each category in the sample database; Weights are assigned to the distances calculated from deep, mid, and shallow features, and the sums are used to obtain the total distance for each candidate category. The category with the smallest total distance is selected as the final judgment result.

6. An improved residual network based UAV abnormal behavior recognition system, characterized in that, The method for performing the UAV abnormal behavior recognition method based on the improved residual network as described in any one of claims 1-5 includes: a training dataset construction module, a residual network improvement module, a residual network training module, a real-time image recognition module, and an anomaly determination module. The training dataset construction module is used to perform hierarchical sampling and preprocessing of multimodal image sets based on the physical, functional, and environmental attributes of the UAV to obtain the training dataset; The residual network improvement module is used to improve the architecture of the residual network model and add a double-constraint center loss mechanism to it. The residual network training module is used to perform phased incremental training on the improved residual network model using the training dataset. The real-time image recognition module is used to extract deep, medium and shallow features of real-time images based on a trained residual network model, and output image recognition results after initial screening and classification. The anomaly detection module is used to calculate the degree of anomaly in the drone's behavior based on the image recognition results of N consecutive frames, and to trigger an alarm and real-time tracking mode when the degree of anomaly exceeds the threshold.

7. A computer storage medium, comprising, include: At least one memory and at least one processor; Memory, used to store one or more program instructions; A processor for running one or more program instructions to perform the UAV abnormal behavior recognition method based on an improved residual network as described in any one of claims 1-5.

Images