A method and system for constructing a classification model of flying birds and drones based on centroid motion characteristics

Abstract

The present invention discloses a method and system for constructing a classification model for birds and drones based on centroid motion characteristics. The method comprises the following steps: shooting video data of birds and drones flying; pre-processing the video data and cutting it into several fixed-duration trajectories to form a video data set; marking the centroid motion trajectory using a computer vision algorithm and performing wavelet denoising pre-processing; calculating the speed and acceleration value motion characteristics of the trajectory segments in the centroid motion trajectory, and extracting the maximum fluctuation amplitude of the trajectory segments in different frequency intervals; using machine learning to establish a speed classifier, an acceleration classifier, and a trajectory fluctuation classifier respectively for the speed, acceleration, and trajectory fluctuation characteristics of birds and drones; weightedly integrating the prediction results of the three classifiers, obtaining the optimal weight using a five-fold cross-validation method, and establishing a classification model for birds and drones; the classification model for birds and drones provided by the present invention is fast and accurate, does not require manual judgment, and reduces labor costs.

Description

Technical Field

[0001] The present invention relates to the field of target classification technology, and in particular to a method and system for constructing a classification model of flying birds and unmanned aerial vehicles based on centroid motion characteristics, and establishing a real-time target classifier. Background Art

[0002] Birds and drones are both typical "low, small, and slow" aircraft, meaning they fly below 1,000 meters, at speeds slower than 200 kilometers per hour, and have a radar cross-section smaller than two square meters. They pose a major threat to low-altitude airspace safety. A bird strike can damage an aircraft's structure, and in severe cases, can result in fatalities. Bird strikes are a leading cause of accidents in aircraft operations. Furthermore, with the advancement of drone technology, drones are becoming a new threat. Their small size, portability, and ease of operation make them vulnerable to illegal activities such as ground mapping and interference with normal civil aviation operations. Different threats from birds and drones require different measures, and appropriate measures require accurate identification and classification of the targets. At long distances, birds and drones are small in size and have similar appearances, making accurate distinction between them difficult.

[0003] Since 2017, the SafeShore project, funded by the EU's Horizon 2020 program, has launched a bird and drone detection challenge. This competition, held every two years, has been held three times to date. Most algorithms in this competition rely on static image information, building convolutional neural network models based on the differences in the appearance of birds and drones. However, these algorithms are less effective at identifying small targets at a distance, with high false positive and false negative rates when the target image size is less than 32 pixels.

[0004] Some researchers have proposed classification based on the motion information of flying targets. Compared to image information, motion information has the advantages of being highly robust to distance and unaffected by object deformation, ambient brightness, and background interference. Srigrarom et al. calculated five motion features for each trajectory segment, including mean velocity, mean acceleration, turning angle, periodicity, and curvature radius. They used principal component analysis to reduce the dimensionality and constructed a support vector machine based on the first two key features, achieving an accuracy exceeding 80%. However, this algorithm has the disadvantage of being based on statistical indicators of motion information, which only reflect partial information and are susceptible to the length of trajectory segments, making it unsuitable for real-time classification tasks. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present invention provides a method and system for constructing a classification model for birds and drones based on center of mass motion characteristics. It is based on the differences in center of mass motion characteristics of birds and drones, and draws on the strategies adopted by the human visual system when classifying moving targets to establish a real-time flying target classifier.

[0006] The present invention adopts the following technical solutions:

[0007] In one aspect, the present invention provides a method for constructing a classification model for birds and drones based on centroid motion characteristics, characterized in that the construction method comprises the following steps:

[0008] Step 1: film the bird flying and control the drone to simulate the bird's trajectory to obtain video data of the bird and the drone respectively;

[0009] Step 2: Preprocess the video data and cut it into several fixed-length trajectories to form a video dataset;

[0010] Step 3: Use computer vision algorithms to mark the center of mass position of the flying target in each frame of the video dataset, form the center of mass motion trajectory, and perform wavelet denoising preprocessing;

[0011] Step 4: Divide each trajectory into several fixed-length segments without overlap. Calculate the velocity, acceleration, and trajectory fluctuation characteristics of the segment. Obtain all velocity and acceleration values, and extract the maximum fluctuation amplitude of the segment in different frequency intervals.

[0012] Step 5: Use machine learning to establish a speed classifier, an acceleration classifier, and a trajectory fluctuation classifier based on the speed, acceleration, and trajectory fluctuation characteristics of birds and drones, respectively.

[0013] In step 6, the prediction results of the three classifiers obtained in step 5 are weighted and integrated using the weighted integration method, and the optimal weight is obtained using the cross-validation method. A motion trajectory classification model is established and the flying bird and / or drone video dataset in the test set is identified.

[0014] Furthermore, the method further includes step 7, wherein the motion trajectory classification model calculates the trajectory segments in the trajectory to obtain the category probabilities of the trajectory segments; and uses a stability assessment module based on the working memory mechanism and decision-making mechanism of the human brain to perform stability assessment on the category probabilities of the trajectory segments and output a classification result.

[0015] Furthermore, in step 7, the motion features in the trajectory segment are input into the motion trajectory classification model constructed in step 6 for calculation to obtain the category probability of the trajectory segment; the motion trajectory classification model continuously inputs category probability information to the stability assessment module; when the same category probability is input, at least 7 of 10 consecutive category probabilities are greater than 0.8, and the category judgments of the velocity classifier, acceleration classifier, and trajectory fluctuation classifier are consistent, the classification result is output.

[0016] Preferably, the constructed motion trajectory classification model calculates the category probability of a trajectory segment containing 24 frames with a duration of 200 ms.

[0017] Preferably, in step 2, when pre-processing the original video data, the motion features of the birds and drones during takeoff, landing and hovering in the original video data are removed to form trajectories of at least 2 seconds each.

[0018] Preferably, in step 3, the ECO target tracking algorithm is used to mark the center of mass position of the flying target in each frame of the collected pigeon video data set, and the DIMP target tracking algorithm is used to mark the center of mass position of the flying target in each frame of the collected topaz, pearl bird and drone video data set.

[0019] Preferably, in step 4, each trajectory is divided into several trajectory segments with a duration of 200ms without overlap, and the method for calculating the trajectory fluctuation motion characteristics in the trajectory segments is: performing moving mean smoothing processing on the center of mass motion trajectory in step 3 to obtain a smooth trajectory; subtracting the horizontal coordinates and vertical coordinates of the smooth trajectory from the horizontal coordinates and vertical coordinates of the center of mass motion trajectory respectively to obtain the trajectory fluctuations in the horizontal and vertical directions; using fast Fourier transform to perform spectral analysis on the trajectory fluctuations to obtain the maximum fluctuation amplitudes in the frequency ranges of 0-10Hz, 10-20Hz and 20-30Hz respectively.

[0020] Preferably, in step 5, a naive Bayes algorithm is used to establish a velocity classifier and an acceleration classifier respectively, and a support vector machine algorithm is used to establish a trajectory fluctuation classifier.

[0021] On the other hand, the present invention also provides a classification model system for birds and drones based on centroid motion characteristics, the system comprising: a camera for capturing video data of birds and drones in flight;

[0022] The preprocessing module is used to preprocess the video data and cut it into several fixed-length trajectories to form a video dataset;

[0023] The visual algorithm annotation module is used to annotate the center of mass position of the flying target in each frame of the pre-processed bird and drone trajectory to form the center of mass motion trajectory of the flying target;

[0024] The wavelet denoising module is used to perform wavelet denoising preprocessing on the centroid motion trajectory to remove high-frequency noise and labeling errors in the centroid motion trajectory;

[0025] The data processing module divides each trajectory into several fixed-length segments without overlap, calculates the velocity, acceleration, and trajectory fluctuation characteristics of each segment, obtains all velocity and acceleration values, and extracts the maximum fluctuation amplitude of the segment in different frequency ranges;

[0026] A classifier construction module extracts the velocity value, acceleration value and trajectory fluctuation amplitude from the data processing module and uses machine learning to construct a velocity classifier, an acceleration classifier and a trajectory fluctuation classifier respectively;

[0027] The classification model construction module uses the weighted integration method to weight the prediction results of the velocity classifier, acceleration classifier and trajectory fluctuation classifier, and uses the cross-validation method to obtain the optimal weight to obtain the motion trajectory classification model.

[0028] Furthermore, the system further includes a stability evaluation module. The motion trajectory classification model calculates the category probability of the trajectory segments in each trajectory, and outputs the classification result after the stability evaluation module performs probability evaluation.

[0029] The technical solution of the present invention has the following advantages:

[0030] A. The bird and drone classification model provided by this invention is fast and highly accurate. Based on a visual system, this invention establishes a motion trajectory classification model based on short-term trajectory segments. This model can automatically and in real time determine the category of moving targets, eliminating the need for manual judgment and reducing labor costs. The established motion trajectory classification model has a high accuracy rate of over 90% for bird and drone trajectory classification, and its simple framework can be transferred to other tasks that classify targets based on motion features.

[0031] B. The present invention requires a relatively small number of training samples. The image features of existing flying targets are affected by factors such as target category, viewing angle, deformation, and occlusion, and are diverse. Building a classifier based on these image features requires a large number of training samples. Motion features, however, are constrained by flight dynamics. The motion features of birds and drones differ fundamentally, and these differences are unaffected by bird category or drone model. Therefore, the present invention can achieve stable results using small sample training, making it applicable to the identification of different types of birds and drones. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the specific embodiments of the present invention, the following will briefly introduce the drawings required for use in the specific embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 Flowchart 1 of the method for constructing a motion trajectory classification model provided by the present invention;

[0034] Figure 2 Flowchart II of the method for constructing a motion trajectory classification model provided by the present invention;

[0035] Figure 3 This is a flow chart of the application of the motion trajectory classification model provided by the present invention;

[0036] Figure 4 This is an example diagram of the flight trajectory provided by the present invention;

[0037] Figure 5 Flowchart of the bird and drone classifier provided by the present invention;

[0038] Figure 6 This is a diagram showing the composition of the motion trajectory classification model construction system provided by the present invention;

[0039] Figure 7 This is the bird test trajectory classification example provided by the present invention - the pearl bird;

[0040] Figure 8 This is the drone test trajectory classification example provided by the present invention - Tello. DETAILED DESCRIPTION

[0041] The technical solutions of the present invention will be described clearly and completely below with reference to the accompanying drawings. It is obvious that the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0042] like Figure 1 、 Figure 3 and Figure 5 As shown, the present invention provides a method for constructing a classification model of flying birds and drones based on centroid motion characteristics. The method adopted is as follows:

[0043]

S01

[0044] To expand the model's applicability, we selected three bird and drone sizes: the bird species, topaz, and pigeon, in descending order; and the drones, Tello, DJI Mavic Pro, and DJI Phantom 4, in descending order. We collected flight video footage of the three bird species indoors, with a minimum duration of two seconds. Then, on a clear, windless day, we had experienced pilots fly the three drones outdoors in an open parking lot, simulating the bird's motion trajectory. This captured the drone flight video footage.

[0045] [S02] The original video data is pre-processed and cut into several trajectories of fixed length. All trajectories form a video dataset.

[0046] The preprocessing here includes using video software to edit all captured video data into 2-second segments. During the preprocessing, the typical motion features of birds and drones, such as takeoff, landing, and hovering, are removed. For example, each type of flying object contains 30 segments of flight video data, each lasting 2 seconds, to form a video dataset. There is no specific limit on the length and number of segments of the flight video dataset.

[0047]

S03

[0048] Computer vision algorithms (including ECO and DIMP) are used to mark the center of mass position of the flying target in each frame of the video dataset. The specific method is: draw a bounding box around the flying target and define the center of the bounding box as the center of mass, such as Figure 4 As shown in the figure, the position of the center of mass target in each frame of the video constitutes the trajectory of the center of mass. After all annotations were completed, the accuracy of the video annotations was manually checked. The ECO target tracking algorithm achieved higher accuracy when annotating pigeons, while the DIMP target tracking algorithm achieved higher accuracy when annotating the other five flying target types (topaz, pearl bird, Tello, Mavic Pro, and Phantom 4).

[0049] Then, the centroid motion trajectory formed after annotation is preprocessed with wavelet denoising to remove high-frequency noise and possible annotation errors in the trajectory.

[0050] The drones used in this study are all quadrotors. The high-frequency rotation of the rotors can cause high-frequency fluctuations in the trajectory. Furthermore, jitter (random noise) and labeling errors can also cause abnormal fluctuations in the trajectory. To remove this noise and obtain a relatively accurate center of mass position, the raw trajectory data must be preprocessed. In this study, wavelet denoising is used to reduce the noise of the raw signal. This wavelet denoising method removes high-frequency noise while preserving signal characteristics.

[0051] [S04] The present invention divides the video dataset into a training set and a test set in a 4:1 ratio. The 2-second trajectories in the training set are divided into short-duration (200ms) trajectory segments without overlap. Three motion features of the trajectory segments are calculated, including velocity, acceleration, and trajectory fluctuation motion features. All velocity and acceleration values ​​are obtained, and the maximum fluctuation amplitude of each trajectory segment in different frequency ranges is extracted.

[0052] Trajectory fluctuation refers to the fluctuations in the original center of mass trajectory compared to the smoothed trajectory. The smoothed trajectory is the trajectory obtained by applying a moving mean smoothing operation to the original center of mass trajectory, with a smoothing window of 15. Trajectory fluctuation is calculated by subtracting the horizontal and vertical coordinates of the smoothed trajectory from the horizontal and vertical coordinates of the original center of mass trajectory, respectively, to obtain the horizontal and vertical fluctuations. The trajectory fluctuations are then spectrally analyzed using a fast Fourier transform, taking the maximum fluctuation amplitude within three frequency ranges: 0-10 Hz, 10-20 Hz, and 20-30 Hz.

[0053] [S05] Using machine learning, a velocity classifier, an acceleration classifier, and a trajectory fluctuation classifier are respectively established based on the speed, acceleration, and trajectory fluctuation characteristics of the bird and the drone. The present invention preferably uses a naive Bayesian algorithm in machine learning to establish a velocity classifier and an acceleration classifier based on all the speeds and accelerations and their corresponding categories obtained in [S04]. A support vector machine algorithm in machine learning is used to establish a trajectory fluctuation classifier based on the trajectory fluctuation amplitude extracted in [S04].

[0054] [S06] Based on the three classifiers obtained in step [S05], for any 200ms trajectory segment, the velocity classifier calculates the class probability of each velocity value in the trajectory segment. The mean of these class probabilities is the velocity class probability of the trajectory segment. The acceleration classifier calculates the class probability of each acceleration value in the trajectory segment. The mean of these class probabilities is the acceleration class probability. The trajectory fluctuation classifier calculates the fluctuation class probability based on the amplitude characteristics. The prediction results of the three classifiers are weighted and integrated using a weighted integration method. The optimal weight is obtained through cross-validation. The motion trajectory classification model is established, that is, the final target classifier is established, and recognition is performed on the bird and / or drone video dataset in the test set.

[0055] The present invention preferably uses a five-fold cross-validation method to calculate the optimal weight. Specifically, the training set is divided into five equal subsets, four of which are used for training and one for validation. This process is repeated five times, and an accuracy rate is calculated each time. The average of the five accuracy rates is used as an estimate of the accuracy rate corresponding to the weight. The optimal weight is the weight that gives the highest accuracy rate.

[0056] In order to apply the established motion trajectory classification model to real-time classification tasks, such as Figure 2 As shown, the present invention also simulates the working mechanism of humans in step [S07] and constructs a machine observer to classify the trajectories in the test set. Figure 5 The machine observer in the invention includes two modules. The first module is the constructed motion trajectory classification model (i.e., bird-drone classifier), which simulates the feature detection network in the brain, continuously extracts the motion features of the trajectory segments, and assigns different weights to them. The motion trajectory classification model in the present invention calculates the category probability of a trajectory segment of fixed length (200ms, 24 frames in total). When applied in the test set, it is tested by inputting trajectory data in real time. When the input trajectory length is less than 200 milliseconds, it is not classified; when the input trajectory length reaches 200 milliseconds, the 200-millisecond trajectory segment is judged and the category probability is output; for each new frame of data input, the latest 200-millisecond trajectory segment is obtained and the new category probability is output. In actual classification tasks, it is not enough to just calculate the category probability, but also requires the help of working memory and decision-making mechanism. Therefore, the present invention simulates the working principle of the human brain and also constructs a second module, namely the stability assessment module. The motion trajectory classification model continuously inputs updated probability information into the stability assessment module. If at least seven of the ten consecutive probabilities of the same category are greater than 0.8, and the three classifiers agree on the category, the model is considered stable and reliable, and the final classification result is output. If the three classifiers disagree, no classification result is output. If the motion trajectory classification model does not make a judgment at the end of the trajectory, indicating that it has not reached stability, the accuracy is recorded as 0.5, and the reaction time is recorded as the average reaction time of the other trajectories for which the judgment was made.

[0057] The trajectory data in the test set is input into the motion trajectory classification model established in [S06] to obtain the predicted category of each trajectory; the consistency between the predicted category and the true category is compared to obtain the classification accuracy of the motion trajectory classification model.

[0058] On the other hand, Figure 6As shown, the present invention also provides a bird and drone classification model system based on centroid motion characteristics, including a camera and a preprocessing module built into a computer, a visual algorithm annotation module, a wavelet denoising module, a data processing module, a classifier construction module and a classification model construction module. The camera is used to shoot video data of birds and drones flying; the preprocessing module is used to preprocess the video data and cut it into several fixed-length trajectories to form a video data set containing several continuous trajectory segments; the visual algorithm annotation module is used to annotate the centroid position of each frame of the preprocessed bird and drone trajectory segments to form the centroid motion trajectory of the flying target; the wavelet denoising module is used to perform wavelet denoising preprocessing on the centroid motion trajectory to remove high-frequency noise and annotation errors in the centroid motion trajectory; the data processing module divides each trajectory segment into several fixed-length trajectory segments without overlapping. The velocity, acceleration and trajectory fluctuation motion characteristics of each trajectory segment are calculated to obtain all velocity and acceleration values, and the maximum fluctuation amplitude of the trajectory segments in different frequency ranges is extracted; the classifier construction module extracts the velocity values, acceleration values ​​and trajectory fluctuation amplitude from the data processing module, and uses machine learning to construct velocity classifiers, acceleration classifiers and trajectory fluctuation classifiers respectively; the classification model construction module uses the weighted integration method to weightedly integrate the prediction results of the velocity classifier, acceleration classifier and trajectory fluctuation classifier, and uses the cross-validation method to obtain the optimal weight to obtain the motion trajectory classification model.

[0059] In order to improve the classification accuracy, a stability evaluation module is also set up in the system. The motion trajectory classification model calculates the category probability of the trajectory segments in each trajectory, and outputs the classification result after the probability evaluation by the stability evaluation module.

[0060] like Figure 7 and Figure 8 Figure 2 shows a schematic diagram of the classification model constructed by the present invention. Birds and drones are distinguished based on differences in their speed and acceleration probability distributions, as well as their trajectory fluctuations. The figure shows that birds have a wider range of speed and acceleration distributions and greater trajectory fluctuations than drones.

[0061] Figure 7 and Figure 8The left side of the first row is the original center of mass motion trajectory, and the right side of the first row is the real-time output result of the classifier probability judgment. The white vertical line indicates the result output time (the average reaction time is 380.79ms), and the classification result is output in the reaction box. The second row is the extracted trajectory features, which are speed, acceleration, and trajectory fluctuations in the horizontal and vertical directions. Below the speed and acceleration are the probability distribution templates of the bird and drone speed and acceleration obtained based on the trajectory data in the training set. The speed and acceleration of the bird are distributed within the probability template of the bird, and the trajectory of the bird fluctuates greatly, such as Figure 7 As shown in Figure 2, the speed and acceleration of the drone are distributed within the probability template of the drone, and the trajectory of the drone has little fluctuation, as shown in Figure 2. Figure 8 shown.

[0062] After testing 30 trajectories in the test set, the accuracy rate reached 100%.

[0063] The present invention requires a smaller number of samples for training. It only needs to collect part of the flight trajectory, and there is no need to calculate all the flight trajectories. The present invention can obtain stable output results by using small sample training, which can be applied to the accurate identification of different types of birds and drones.

[0064] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for constructing a classification model for birds and drones based on centroid motion characteristics, characterized in that: The construction method includes the following steps: Step 1: film the bird flying and control the drone to simulate the bird's trajectory to obtain video data of the bird and the drone respectively; Step 2: Preprocess the video data and cut it into several fixed-length trajectories to form a video dataset; Step 3: Use computer vision algorithms to mark the center of mass position of the flying target in each frame of the video dataset, form the center of mass motion trajectory, and perform wavelet denoising preprocessing; Step 4: Divide each trajectory into several local trajectory segments of fixed duration without overlapping. Calculate the velocity, acceleration, and trajectory fluctuation characteristics of the trajectory segments, obtain all velocity and acceleration values, and extract the maximum fluctuation amplitude of the trajectory segments in different frequency intervals. Step 5: Use machine learning to establish a speed classifier, an acceleration classifier, and a trajectory fluctuation classifier based on the speed, acceleration, and trajectory fluctuation characteristics of birds and drones, respectively. Step 6: Use the weighted integration method to weight the prediction results of the three classifiers obtained in step 5, use the cross-validation method to obtain the optimal weight, establish a motion trajectory classification model, and identify the flying bird and / or drone video dataset in the test set; In step 7, the trajectory data is input frame by frame into the motion trajectory classification model. When the input trajectory length reaches 200 milliseconds, the motion trajectory classification model calculates the category probability of the trajectory segment. For each new frame of data input, the latest 200-millisecond trajectory segment is obtained and the new category probability is output. As the trajectory data is continuously input, the motion trajectory classification model continuously inputs updated probability information into the stability assessment module. The stability assessment module, based on the working memory and decision-making mechanism of the human brain, performs a stability assessment on the category probability of the trajectory segment and outputs the classification result. The classification result is output when at least 7 out of 10 consecutive input category probabilities have a category probability greater than 0.8, and the category judgments of the velocity classifier, acceleration classifier, and trajectory fluctuation classifier are consistent.

2. The method for constructing a classification model of birds and drones based on centroid motion characteristics according to claim 1, characterized in that: The constructed motion trajectory classification model calculates the category probability of the trajectory segment containing 24 frames and a duration of 200ms.

3. The method for constructing a classification model of flying birds and drones based on centroid motion characteristics according to claim 1, characterized in that: In step 2, when pre-processing the original video data, the motion features of the birds and drones during takeoff, landing, and hovering in the original video data are removed to form trajectories of at least 2 seconds each.

4. The method for constructing a classification model of birds and drones based on centroid motion characteristics according to claim 1, characterized in that: In step 3, the ECO target tracking algorithm is used to mark the center of mass position of the flying target in each frame of the collected pigeon video dataset, and the DIMP target tracking algorithm is used to mark the center of mass position of the flying target in each frame of the collected topaz, pearl bird and drone video dataset.

5. The method for constructing a classification model of flying birds and drones based on centroid motion characteristics according to claim 1, characterized in that: In step 4, each trajectory is divided into several trajectory segments with a duration of 200ms without overlap. The method for calculating the trajectory fluctuation motion characteristics in the trajectory segments is: performing moving mean smoothing processing on the center of mass motion trajectory in step 3 to obtain a smooth trajectory; subtracting the horizontal coordinates and vertical coordinates of the smooth trajectory from the horizontal coordinates and vertical coordinates of the center of mass motion trajectory to obtain the trajectory fluctuations in the horizontal and vertical directions; using fast Fourier transform to perform spectral analysis on the trajectory fluctuations to obtain the maximum fluctuation amplitudes in the frequency ranges of 0-10Hz, 10-20Hz and 20-30Hz respectively.

6. The method for constructing a classification model of birds and drones based on centroid motion characteristics according to claim 1, characterized in that: In step 5, the naive Bayes algorithm is used to establish a velocity classifier and an acceleration classifier respectively, and the support vector machine algorithm is used to establish a trajectory fluctuation classifier.

7. A bird and drone classification model system based on centroid motion characteristics, characterized by: The system comprises: Cameras for capturing video footage of birds and drones flying; The preprocessing module is used to preprocess the video data and cut it into several fixed-length trajectories to form a video dataset; The visual algorithm annotation module is used to annotate the center of mass position of the flying target in each frame of the pre-processed bird and drone trajectory to form the center of mass motion trajectory of the flying target; The wavelet denoising module is used to perform wavelet denoising preprocessing on the centroid motion trajectory to remove high-frequency noise and labeling errors in the centroid motion trajectory; The data processing module divides each trajectory into several fixed-length segments without overlap, calculates the velocity, acceleration, and trajectory fluctuation characteristics of each segment, obtains all velocity and acceleration values, and extracts the maximum fluctuation amplitude of the segment in different frequency ranges; A classifier construction module extracts the velocity value, acceleration value and trajectory fluctuation amplitude from the data processing module and uses machine learning to construct a velocity classifier, an acceleration classifier and a trajectory fluctuation classifier respectively; The classification model construction module uses the weighted integration method to weight the prediction results of the velocity classifier, acceleration classifier, and trajectory fluctuation classifier, and uses the cross-validation method to obtain the optimal weight to obtain the motion trajectory classification model; Machine observer, including a motion trajectory classification model and a stability assessment module; The motion trajectory classification model simulates the feature detection network in the brain, continuously extracting motion features of trajectory segments, assigning them different weights, and making judgments on the trajectory segments based on the trajectory classification model. Trajectory segment information is input frame by frame. When the segment accumulates to 200 milliseconds, the classifier will make its first judgment. Subsequently, as new trajectory data continues to be input, the classifier will continuously generate new classification results based on the updated trajectory segments. The stability assessment module simulates the working memory and decision-making mechanism of the brain. It continuously receives updated probability information from the motion trajectory classification model. When at least 7 out of 10 consecutive input probabilities of the same category are greater than 0.8, and the category judgments of the three classifiers are consistent at this time, it is considered a stable and reliable judgment, and the final classification result is output.

Images