A high-altitude projectile detection method, device, equipment and storage medium

Abstract

The application discloses a high-altitude throwing detection method and device, equipment and storage medium, and relates to the technical field of video monitoring security and protection, and comprises the following steps: determining whether a detection target is a high-altitude throwing target based on a current track of the detection target; if yes, determining again whether the detection target is the high-altitude throwing target based on an EDR distance between a standard track and the current track based on an angle sequence; if still yes, obtaining a slice image by transversely slicing a relevant area of the detection target in a collected high-altitude throwing monitoring image, splicing and cutting all slice images related to the high-altitude throwing target in continuous frames to obtain a detection target area image, and inputting the detection target area image into a pre-trained deep convolutional neural network for further detection to obtain a judgment result of whether the high-altitude throwing target is a real high-altitude throwing target. Through the strategy of hierarchical detection and identification, the false detection rate of the high-altitude throwing can be reduced.

Description

Technical Field

[0001] This invention relates to the field of video surveillance and security technology, and in particular to a method, device, equipment, and storage medium for detecting objects thrown from heights. Background Technology

[0002] Currently, with urbanization, objects thrown from high-rise buildings have increasingly become a safety hazard for community residents. The crime of throwing objects from high-rise buildings was officially implemented in March 2021 to ensure residents' safety. However, the concealed, rapid, and small nature of these objects leads to excessively high costs for manual enforcement and difficulties in supervision. Currently, intelligent image analysis-based algorithms for detecting objects thrown from high-rise buildings have become a viable method. However, these algorithms need to maximize sensitivity when facing small, fast-moving targets, which places high demands on equipment and deployment environments. To capture high-speed objects, current intelligent algorithms often sacrifice some accuracy. In other words, due to the extremely small size of the objects, their high-speed movement, and the complex background, there is a trade-off between detection accuracy and false positives; achieving high accuracy can result in a large number of false positives. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a method, apparatus, device, and storage medium for detecting objects thrown from heights, which can reduce the false detection rate of objects thrown from heights. The specific solution is as follows:

[0004] Firstly, this application discloses a method for detecting objects thrown from heights, including:

[0005] The motion trajectory corresponding to the detected target is determined to obtain the current trajectory, and based on the current trajectory, it is determined whether the detected target is a high-altitude projectile target;

[0006] If the current trajectory indicates that the detected target is the high-altitude projectile target, then the corresponding standard trajectory is determined, and the EDR distance based on the angle sequence between the standard trajectory and the current trajectory is determined;

[0007] Based on the EDR distance, it is determined again whether the detected target is the high-altitude projectile target;

[0008] If the detected target is the high-altitude object thrown from above, then the relevant area of ​​the high-altitude object throwing monitoring image showing the detected target is horizontally sliced ​​to obtain the corresponding slice image. Then, all the slice images related to the high-altitude object thrown from above in consecutive frames are stitched and cropped to obtain the corresponding detection target area image. The detection target area image is input into a pre-trained deep convolutional neural network for further detection to obtain the judgment result of whether the high-altitude object thrown from above is a real high-altitude object thrown from above, as output by the deep convolutional neural network.

[0009] Optionally, determining whether the detected target is a high-altitude projectile based on the current trajectory includes:

[0010] The category of the current trajectory is determined, and a preset feature value corresponding to the category is obtained; the category is any one of the following: high parabola point high standard deviation, high parabola point low standard deviation, low parabola point high standard deviation, or low parabola point low standard deviation.

[0011] Determine the 4-dimensional feature values ​​of the current trajectory; the 4-dimensional features include the first-order difference mean of the x-axis coordinate sequence, the first-order difference mean of the y-axis coordinate sequence, the second-order difference mean of the x-axis coordinate sequence, and the second-order difference mean of the y-axis coordinate sequence;

[0012] Calculate the percentage difference between the 4-dimensional feature values ​​of the current trajectory and the preset feature values;

[0013] Based on the comparison between the percentage difference and the first preset threshold, it is determined whether the detected target corresponding to the current trajectory is a high-altitude projectile target.

[0014] Optionally, determining whether the detected target corresponding to the current trajectory is a high-altitude projectile target based on the comparison result between the difference percentage and the first preset threshold includes:

[0015] If the comparison result indicates that the percentage difference is greater than the first preset threshold, then the detected target corresponding to the current trajectory is determined to be a non-high-altitude projectile target, and the current trajectory is filtered out.

[0016] If the comparison result indicates that the percentage difference is not greater than the first preset threshold, then the detected target corresponding to the current trajectory is determined to be a high-altitude projectile target.

[0017] Optionally, determining the corresponding standard trajectory includes:

[0018] The corresponding standard trajectory is determined based on the 4-dimensional feature values ​​and the preset feature values.

[0019] Optionally, determining the angle sequence-based EDR distance between the standard trajectory and the current trajectory includes:

[0020] Obtain the first angle sequence corresponding to the current trajectory and the second angle sequence corresponding to the standard trajectory; the first angle sequence is a sequence formed by arranging the inferior angles formed by three adjacent points in the current trajectory in order, and the second angle sequence is a sequence formed by arranging the inferior angles formed by three adjacent points in the standard trajectory in order;

[0021] The angle sequence-based EDR distance between the standard trajectory and the current trajectory is determined based on the first angle sequence and the second angle sequence.

[0022] Optionally, the step of determining again whether the detected target is the high-altitude projectile target based on the EDR distance includes:

[0023] The EDR distance is compared with a comparison threshold, which is a threshold determined based on the number of elements in the longer sequence of the current trajectory and the standard trajectory.

[0024] If the EDR distance is greater than the comparison threshold, then the detected target is determined to be the non-high-altitude projectile target;

[0025] If the EDR distance is not greater than the comparison threshold, then the detected target is determined to be the high-altitude projectile target.

[0026] Optionally, the step of horizontally slicing the relevant area of ​​the detected target in the acquired high-altitude object throwing monitoring image to obtain a corresponding slice image, and stitching and cropping all the slice images related to the high-altitude object throwing target in consecutive frames to obtain a corresponding detection target area image, includes:

[0027] Based on the coordinates of the trajectory points of the detected target in the high-altitude object throwing monitoring image, the high-altitude object throwing monitoring image is horizontally sliced ​​to obtain a slice image containing the detected target, and the slice image is height-scaled according to the time sequence to obtain a scaled slice image.

[0028] All the scaled slice images are stitched together to obtain the stitched slice image;

[0029] The spliced ​​sliced ​​image is cropped to obtain the image of the target detection region.

[0030] Optionally, the method for detecting objects thrown from heights further includes:

[0031] The image of the detected target area corresponding to the real high-altitude projectile target is used as a time-series stitched image of positive samples;

[0032] The high-altitude littering monitoring images are collected at preset time intervals, and the building outlines in the high-altitude littering monitoring images are determined as the selected area;

[0033] Within the selected area, a corresponding cropping region is determined to obtain multiple target cropping regions, and all the target cropping regions are stitched together to obtain a time-series stitched image as a negative sample;

[0034] A training dataset is constructed using temporal stitched images that serve as positive samples and temporal stitched images that serve as negative samples.

[0035] The trained deep convolutional neural network is obtained by training the initial deep convolutional neural network using the training dataset.

[0036] Optionally, determining the corresponding cropping region within the selected area to obtain multiple target cropping regions includes:

[0037] Within the selected area, the first point is determined as the x-axis coordinate by randomly selecting points on average in the x-axis direction, and a point is randomly selected as the y-axis coordinate from the normal distribution established based on the selected area in the y-axis direction. The coordinates of the upper left corner of the first target cropping area are determined based on the x-axis coordinate and the y-axis coordinate.

[0038] The first target cropping region is determined based on the coordinates of the top left corner of the first target cropping region, the height of the scaled slice image, and the image input width specified by the initial depth convolutional neural network.

[0039] The extreme value intervals corresponding to the first difference of the x-axis coordinates of the high-altitude parabolic targets in the positive samples are statistically analyzed, and a corresponding uniform distribution is established based on the extreme value intervals.

[0040] The result of random sampling within the uniform distribution is determined as the initial downward velocity of the pseudo-parabola along the x-axis.

[0041] The average initial y-axis velocity of the high-altitude projectile target in the positive sample is statistically analyzed, and the average initial y-axis velocity is determined as the downward y-axis initial velocity of the dummy projectile.

[0042] Based on the upper left corner coordinates corresponding to the first target cropping region, the lower left corner coordinates of multiple target cropping regions are determined by the lower x-axis initial velocity and the lower y-axis initial velocity.

[0043] The multiple target cropping regions are determined based on the top-left corner coordinates of the multiple target cropping regions, the height of the scaled slice image, and the image input width specified by the initial deep convolutional neural network.

[0044] Secondly, this application discloses a high-altitude object throwing detection device, comprising:

[0045] The trajectory determination module is used to determine the motion trajectory corresponding to the detected target to obtain the current trajectory;

[0046] The first detection module is used to determine whether the detected target is a high-altitude projectile target based on the current trajectory;

[0047] The EDR distance determination module is used to determine the corresponding standard trajectory and the angle sequence-based EDR distance between the standard trajectory and the current trajectory if the current trajectory indicates that the detected target is the high-altitude projectile target.

[0048] The second detection module is used to determine again whether the detected target is the high-altitude projectile target based on the EDR distance;

[0049] The third detection module is used to, if the detected target is the high-altitude object throwing target, horizontally slice the relevant area of ​​the detected target in the acquired high-altitude object throwing monitoring image to obtain the corresponding slice image, and stitch and crop all the slice images related to the high-altitude object throwing target in consecutive frames to obtain the corresponding detection target area image, and input the detection target area image into a pre-trained deep convolutional neural network for further detection to obtain the judgment result of whether the high-altitude object throwing target is a real high-altitude object thrown by the deep convolutional neural network.

[0050] Thirdly, this application discloses an electronic device, including:

[0051] Memory, used to store computer programs;

[0052] A processor is used to execute the computer program to implement the steps of the aforementioned disclosed method for detecting objects thrown from heights.

[0053] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed method for detecting objects thrown from heights.

[0054] As can be seen, this application provides a method for detecting objects thrown from high altitudes, comprising: determining the motion trajectory corresponding to the target to obtain the current trajectory, and judging whether the target is a target thrown from high altitudes based on the current trajectory; if the current trajectory indicates that the target is a target thrown from high altitudes, determining a corresponding standard trajectory, and determining the EDR distance between the standard trajectory and the current trajectory based on the angle sequence; judging whether the target is a target thrown from high altitudes again based on the EDR distance; if the target is a target thrown from high altitudes, horizontally slicing the relevant area of ​​the target in the acquired high-altitude object throwing monitoring image to obtain a corresponding slice image, and stitching and cropping all the slice images related to the target in consecutive frames to obtain a corresponding target area image, and inputting the target area image into a pre-trained deep convolutional neural network for further detection to obtain the judgment result of whether the target thrown from high altitudes is a real target thrown from high altitudes output by the deep convolutional neural network. Therefore, this application adopts a hierarchical detection and identification strategy, prioritizing the detection of objects thrown from high altitudes based on trajectory information and EDR distance, which have low computational complexity. Then, it inputs strong interference targets that cannot be filtered out into a deep convolutional neural network for further detection. This can solve the problem of high detection rate and high false detection rate of current high-altitude object throwing algorithms, and reduce the false detection rate of high-altitude object throwing. Attached Figure Description

[0055] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1 This is a flowchart of a method for detecting objects thrown from heights disclosed in this application;

[0057] Figure 2 This is a flowchart of a specific trajectory-based high-altitude object detection method disclosed in this application;

[0058] Figure 3 This is a schematic diagram of an angle sequence composition disclosed in this application;

[0059] Figure 4 This is a schematic diagram of a time-series clipping and splicing of a parabolic target disclosed in this application;

[0060] Figure 5 This is a schematic diagram of the timing splicing of a parabolic target disclosed in this application;

[0061] Figure 6This is a schematic diagram of a negative sample clipping and splicing method disclosed in this application;

[0062] Figure 7 This is a schematic diagram of the structure of a high-altitude object throwing detection device disclosed in this application;

[0063] Figure 8 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0064] 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 embodiments of the present invention, and not all embodiments. 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.

[0065] Currently, high-altitude object throwing detection algorithms based on intelligent image analysis have become a favorable method for detecting objects thrown from heights. However, when facing small, fast-moving targets, these algorithms need to maximize sensitivity, which places high demands on equipment and deployment environments. To capture high-speed objects, current intelligent algorithms often sacrifice some accuracy. In other words, due to the extremely small size of high-altitude objects, their high-speed movement, and the complexity of the detection background, there is a trade-off between detection accuracy and false positives; ensuring high accuracy can lead to a large number of false positives. Therefore, this application provides a new high-altitude object throwing detection scheme that can reduce the false positive rate.

[0066] This invention discloses a method for detecting objects thrown from heights, see [link to relevant documentation]. Figure 1 As shown, the method includes:

[0067] Step S11: Determine the motion trajectory corresponding to the detected target to obtain the current trajectory, and determine whether the detected target is a high-altitude projectile target based on the current trajectory.

[0068] In this embodiment, the device determines the detection target in each video frame (i.e., each high-altitude object throwing monitoring image) of the acquired high-altitude object throwing monitoring video, and then determines the motion trajectory corresponding to the detection target to obtain the current trajectory of the detection target. Specifically, a preset moving target detection algorithm is used to extract the foreground image corresponding to the acquired high-altitude object throwing monitoring image to determine the corresponding detection target, and the detection box of the detection target is determined. A preset tracking algorithm is used to predict and match the detection box to obtain continuous detection boxes corresponding to the same detection target, and the center point of the detection box is determined as the center point of the detection target. The motion trajectory corresponding to the detection target is determined based on all the center points of the detection target to obtain the current trajectory of the detection target. It can be understood that by taking the center point corresponding to the detection box as the center point of the detection target, and then arranging all the center points in sequence, the trajectory sequence corresponding to the detection target can be formed to obtain the current trajectory. For example, background modeling is used to extract the foreground image of each video frame in a high-altitude object throwing surveillance video to determine the corresponding detection target. Then, traditional filtering, dilation, erosion, and region matching methods are used to determine the detection box of the target. Subsequently, the SORT (Simple Online and Realtime Tracking) algorithm is used to predict and match the detection box to obtain continuous detection boxes corresponding to the target under the same ID (Identity Document). The center point of these continuous detection boxes is used as the center point of the target. The coordinates of all the center points are arranged sequentially to form the trajectory sequence of the target and obtain the current trajectory.

[0069] It should be noted that the preset moving target detection algorithm can be not only a background modeling algorithm, but also other moving target detection algorithms, and the preset tracking algorithm can not only be the SORT tracking algorithm, but also other tracking algorithms with better timeliness and effect.

[0070] In this embodiment, after determining the current trajectory of the target, it is determined whether the target is a high-altitude object thrown from a height based on the current trajectory. That is, based on the current trajectory of the target, trajectories that are clearly not high-altitude objects thrown from a height can be filtered out, i.e., current trajectories that do not meet the conditions are filtered out. For the trajectories that are not filtered out, which may be high-altitude objects thrown from a height, they are detected again. Figure 2 As shown, the specific steps for determining whether the detected target is a high-altitude projectile based on the current trajectory are as follows:

[0071] Step S111: Determine the category of the current trajectory and obtain the preset feature value corresponding to the category; the category is any one of the following: high parabolic point high standard deviation, high parabolic point low standard deviation, low parabolic point high standard deviation, or low parabolic point low standard deviation.

[0072] Understandably, since the initial parabolic height has a significant impact on the number of trajectory points, velocity, and acceleration, different threshold data are needed for parabolas of different heights. This application sets two types based on the dataset: high parabolic points and low parabolic points. To increase accuracy and reduce the difficulty of subsequent threshold adjustments, the number of types can be appropriately increased, but each type must ensure sufficient data differentiation. Similarly, the standard deviation of the first difference of the parabolic y-axis coordinate varies significantly between different types of parabolas, such as a lightweight plastic bag versus heavy wet waste. Therefore, the current trajectory is classified into two dimensions: the height of the first parabolic point and the standard deviation of the first difference of the y-axis coordinates of all center points in the trajectory. After determining the category of the current trajectory, a preset feature value corresponding to that category is obtained.

[0073] Step S112: Determine the 4-dimensional feature values ​​of the current trajectory; the 4-dimensional features include the first-order difference mean of the x-axis coordinate sequence, the first-order difference mean of the y-axis coordinate sequence, the second-order difference mean of the x-axis coordinate sequence, and the second-order difference mean of the y-axis coordinate sequence.

[0074] In this embodiment, the 4-dimensional feature values ​​of the current trajectory are determined. That is, the first-order difference mean of the x-axis coordinate sequence, the first-order difference mean of the y-axis coordinate sequence, the second-order difference mean of the x-axis coordinate sequence, and the second-order difference mean of the y-axis coordinate sequence are calculated based on the coordinate sequence of the corresponding trajectory points. For example, a trajectory S of length n, with time slices of equal length, i.e., a current trajectory S containing n trajectory points, is represented as:

[0075] S={(x1, y1), (x2, y2), (x3, y3),...(x n y n )};

[0076] Calculate the 4-dimensional features corresponding to the current trajectory S under this category, i.e.

[0077] in, Let represent the first-order difference mean of the x-axis coordinate sequence, and the formula for calculating the first-order difference mean of the x-axis coordinate sequence is as follows:

[0078]

[0079] X′={x′1,x′2,...,x′ n-1}={x2-x1,x3-x2,...,x n -x n-1};

[0080] Let represent the first-order difference mean of the y-axis coordinate sequence, and the formula for calculating the first-order difference mean of the y-axis coordinate sequence is as follows:

[0081]

[0082] Y′={y′1, y′2,...,y′ n-1}={y2-y1,y3-y2,...,y n -y n-1};

[0083] Let represent the second difference mean of the x-axis coordinate sequence, and the formula for calculating the second difference mean of the x-axis coordinate sequence is as follows:

[0084]

[0085] X″={x″1,x″2,…,x″ n-1}={x′2-x′1, x′3-x′2,...,x′ n-1 -x′ n-2};

[0086] Let represent the second difference mean of the y-axis coordinate sequence, and the formula for calculating the second difference mean of the y-axis coordinate sequence is as follows:

[0087]

[0088] Y″={y″1,y″2,…,y″ n-1}={y′2-y′1, y′3-y′2,...,y′ n-1 -y′ n-2};

[0089] Step S113: Calculate the percentage difference between the 4-dimensional feature value of the current trajectory and the preset feature value.

[0090] In this embodiment, after determining the 4-dimensional feature value of the current trajectory and the preset feature value of the corresponding category of the current trajectory, the percentage difference between the 4-dimensional feature value of the current trajectory and the preset feature value is calculated. For example, the percentage difference p between the feature value of the current trajectory and the feature value of a preset trajectory under the corresponding category is calculated, where the feature value of the preset trajectory is the preset feature value. The formula for calculating the percentage difference p is as follows:

[0091]

[0092] Where, t∈{x′ c y′ c ,x″ c ,y″ c} represents the 4-dimensional feature value corresponding to the current trajectory. This represents the 4-dimensional feature value calculated for the preset trajectory, i.e., the preset feature value.

[0093] It should be noted that the calculation process for the preset feature values ​​corresponding to different trajectory categories is as follows: Trajectory sequences are collected to obtain corresponding trajectory sequence data, i.e., multiple trajectories are collected. This trajectory sequence data can cover all trajectory categories in this application. Then, the category of each trajectory is determined, and the 4-dimensional feature values ​​of each trajectory in each trajectory category are calculated. This yields the 4-dimensional feature values ​​of all trajectories in each trajectory category. Finally, DBSCAN (Density-Based Spatial Clustering of Applications with...) is used... The density-based noisy spatial clustering (DBSCAN) algorithm removes outlier features from each trajectory category. Specifically, for each of the four trajectory categories, the feature values ​​corresponding to each category are used as input, and the DBSCAN algorithm is employed to eliminate outlier features. For example, since the feature value data has already been divided into single classes based on the trajectory category, the hyperparameter K of the minimum number of neighborhood points in the DBSCAN algorithm is less meaningful; therefore, K is set to 5. The algorithm then calculates the maximum L2 distance d among the feature values ​​of all trajectories in that trajectory category. The ratio of the L2 distance d to the number of trajectories L in that category is used as the interval, i.e., d / L. Starting from d / L, the interval is increased by one each time, forming the search space for the neighborhood radius.

[0094] L2 distance is the Euclidean distance.

[0095] Within the search space, neighborhood radii are selected sequentially from smallest to largest to search for outliers. If the number of outliers is less than [a certain value], [the search will proceed]. Then stop the search, remove the abnormal feature values ​​under that trajectory category, calculate the centroid of that trajectory category, and use the centroid as the preset feature value corresponding to that trajectory category, thereby obtaining the preset feature value corresponding to each trajectory category, i.e.

[0096] in, And m is the number of trajectory sequences obtained after outlier exclusion in this trajectory category.

[0097] Step S114: Determine whether the detected target corresponding to the current trajectory is a high-altitude projectile target based on the comparison result between the difference percentage and the first preset threshold.

[0098] In this embodiment, after calculating the percentage difference between the 4-dimensional feature value of the current trajectory and the feature value of a preset trajectory under the corresponding category, it is determined whether the detected target corresponding to the current trajectory is a high-altitude projectile target based on the comparison result between the percentage difference and a first preset threshold. It can be understood that if the comparison result shows that the percentage difference is greater than the first preset threshold, the detected target corresponding to the current trajectory is determined to be a non-high-altitude projectile target, and the current trajectory is filtered; if the comparison result shows that the percentage difference is not greater than the first preset threshold, the detected target corresponding to the current trajectory is determined to be a high-altitude projectile target.

[0099] The specific process for determining the category of the current trajectory can be as follows: determining the standard deviation between the height of the first occurrence of the parabola and the first difference of the y-axis coordinates of all the center points in the current trajectory; determining whether the standard deviation is less than a second preset threshold, and determining whether the trajectory of the first occurrence of the parabola is located in the upper half of a regular region; the regular region is the rectangular area occupied by buildings in the target frame image; if the standard deviation is not less than the second preset threshold and the trajectory of the first occurrence of the parabola is located in the upper half of the regular region, then the category of the current trajectory is determined to be the high parabola high standard deviation category; if If the standard deviation is less than the second preset threshold and the trajectory where the parabolic point first appears is located in the upper half of the regular region, then the current trajectory is classified as a high parabolic point, low standard deviation category. If the standard deviation is not less than the second preset threshold and the trajectory where the parabolic point first appears is not located in the upper half of the regular region, then the current trajectory is classified as a low parabolic point, high standard deviation category. If the standard deviation is less than the second preset threshold and the trajectory where the parabolic point first appears is not located in the upper half of the regular region, then the current trajectory is classified as a low parabolic point, low standard deviation category. In other words, if the trajectory where the parabolic point first appears is located in the upper half of the regular region, it is a high parabolic point; otherwise, it is a low parabolic point. If the standard deviation between the height of the first parabolic point and the first difference of the y-axis coordinates of all the center points in the current trajectory is less than the second preset threshold, it is a low standard deviation; otherwise, it is a high standard deviation. The regular region refers to the area occupied by buildings in the image, generally selected using a rectangular frame, and this regular region can be appropriately expanded or contracted.

[0100] Step S12: If the current trajectory indicates that the detected target is the high-altitude projectile target, then determine the corresponding standard trajectory and determine the EDR distance based on the angle sequence between the standard trajectory and the current trajectory.

[0101] In this embodiment, if the detected target is determined to be a high-altitude projectile based on the trajectory information of the current trajectory, that is, the current trajectory has not been filtered out based on its trajectory information, then a corresponding standard trajectory is determined, and the EDR (Edit Distance on Real Sequence) based on the angle sequence between the standard trajectory and the current trajectory is determined. It can be understood that a standard trajectory is obtained based on preset feature values, and the EDR distance based on the angle sequence between the standard trajectory and the current trajectory is calculated for subsequent further filtering of the current trajectory. The corresponding standard trajectory can be determined based on the 4-dimensional feature values ​​and the preset feature values; that is, the trajectory closest to the 4-dimensional feature values ​​and the preset feature values ​​under the corresponding category of the current trajectory is calculated, and this trajectory is used as the standard trajectory under the corresponding category of the current trajectory. The method for selecting the closest trajectory is as follows:

[0102]

[0103]

[0104] In this embodiment, determining the EDR distance based on the angle sequence between the standard trajectory and the current trajectory may include: obtaining a first angle sequence corresponding to the current trajectory and a second angle sequence corresponding to the standard trajectory; the first angle sequence is a sequence formed by arranging the inferior angles formed by three adjacent points in the current trajectory in order, and the second angle sequence is a sequence formed by arranging the inferior angles formed by three adjacent points in the standard trajectory in order; determining the EDR distance based on the angle sequence between the standard trajectory and the current trajectory based on the first angle sequence and the second angle sequence. It can be understood that obtaining the angle sequences of the current trajectory and the standard trajectory, such as... Figure 3 As shown, the angle sequence is formed by arranging the inferior angles formed by three adjacent points in the trajectory sequence in order. For example, the angle sequences of the current trajectory and the standard trajectory are represented as A. cur and A std 'a' represents the angle value. Then, the EDR distance based on the angle sequence between the standard trajectory and the current trajectory is calculated, where D is the EDR distance based on the angle sequence. edr (A cur A std The formula is as follows:

[0105]

[0106]

[0107] The formula for the EDR distance based on the angle sequence is a recursive formula, where len(·) represents the number of elements in the angle sequence, and Rest(·) represents the new angle sequence formed after removing the first angle from the original angle sequence. For example, if the original angle sequence is A={a1,a2,...,a...} n Given a sequence of angles} and a length of n, the new angle sequence is Rest(A) = {a2, a3, ..., a...}. n}, and its length is n-1; a1 and b1 are A in this recursive formula. cur A std The first angle value; ε is the preset threshold.

[0108] Step S13: Determine again whether the detected target is the high-altitude projectile target based on the EDR distance.

[0109] In this embodiment, after determining the EDR distance based on the angle sequence between the standard trajectory and the current trajectory, the detection target is again determined to be the high-altitude projectile target based on the EDR distance. That is, the current trajectory is re-detected and identified based on the EDR distance. Specifically, the EDR distance is compared with a comparison threshold; the comparison threshold is a threshold determined based on the number of elements in the longer sequence of the current trajectory and the standard trajectory. If the EDR distance is greater than the comparison threshold, the detection target is determined to be the non-high-altitude projectile target; if the EDR distance is not greater than the comparison threshold, the detection target is determined to be the high-altitude projectile target.

[0110] In this embodiment, determining the comparison threshold based on the number of elements in the longer sequence of the current trajectory and the standard trajectory may include: determining the first number of elements in the first angle sequence and the second number of elements in the second angle sequence; comparing the first number of elements and the second number of elements; if the first number of elements is greater than the second number of elements, then determining a corresponding comparison threshold based on the first number of elements; if the first number of elements is not greater than the second number of elements, then determining a corresponding comparison threshold based on the second number of elements. It can be understood that the longer sequence in the angle sequences corresponding to the current trajectory and the standard trajectory is determined, and then the comparison threshold for comparing with the EDR distance is determined based on the number of elements in the longer sequence. For example, 50% of the number of elements in the longer sequence can be determined as the comparison threshold.

[0111] Step S14: If the detected target is the high-altitude object throwing target, then the relevant area of ​​the detected target appearing in the acquired high-altitude object throwing monitoring image is horizontally sliced ​​to obtain the corresponding slice image, and all the slice images related to the high-altitude object throwing target in consecutive frames are stitched and cropped to obtain the corresponding detection target area image. The detection target area image is input into a pre-trained deep convolutional neural network for further detection to obtain the judgment result of whether the high-altitude object throwing target is a real high-altitude object throwing target output by the deep convolutional neural network.

[0112] Understandably, high-altitude object detection based on trajectory information and EDR distance, which has relatively low computational complexity, can eliminate most false detections of targets with dissimilar trajectories, such as fluttering clothing, falling leaves, birds and insects, and some rainy or snowy weather. However, there are still some complex targets that cannot be identified using trajectory information and EDR distance, mainly in rainy or snowy weather. Therefore, it is necessary to eliminate them by connecting and matching inter-frame features. That is, the relevant regions of the detected target appearing in the acquired high-altitude object monitoring images are cropped and stitched together to obtain the corresponding target region image. The high-altitude object and the target region image in consecutive frames are then input into a pre-trained deep convolutional neural network for further detection to obtain the judgment result of whether the high-altitude object is a real high-altitude object output by the deep convolutional neural network. In other words, the target in consecutive frames and the relevant regions of each frame in the consecutive frames are cropped and stitched together and then input into the trained deep convolutional neural network to determine whether the detected target is a real high-altitude object.

[0113] In this embodiment, the step of cropping and stitching together the relevant area of ​​the detected target appearing in the acquired high-altitude object throwing monitoring image to obtain the corresponding detection target area image may specifically include: horizontally slicing the high-altitude object throwing monitoring image according to the coordinates of the trajectory points of the detected target appearing in the acquired high-altitude object throwing monitoring image to obtain slice images containing the detected target, and scaling the slice images in chronological order to obtain scaled slice images; stitching together all the scaled slice images to obtain a stitched slice image; and cropping the stitched slice image based on the detection box to obtain the detection target area image. For example, as... Figure 4As shown, taking n=4 as an example, horizontal slices are made according to the coordinates of the trajectory points in the corresponding frames where the detected target appears. Then, all slices are scaled according to the time sequence, that is, the width of the slices remains unchanged while the height of the slices is compressed to h. Then, n consecutive scaled slices are stitched together to obtain a stitched slice with a stitched height of n*h. Finally, the horizontal excess pixels are cropped according to the detection box to obtain the image of the detected target region. Here, the value of h is determined by the average pixel height of the parabolic in the dataset. Since a value of n that is too small may cause the network to fail to learn the continuous parabolic features, and a value of n that is too large may cause the tangent of the image fed into the network to be too obvious, and at the same time, it may cause the number of pixels occupied by the parabolic to be too small, the setting of the value of n should be tried multiple times to select a suitable value. This process is applicable to the acquisition of positive samples during network training.

[0114] It should be noted that various practical application scenarios were considered when constructing the training sample set. Therefore, it is necessary to have as many and rich building backgrounds as possible, including parabolic scenarios during the day or night, images of parabolic objects at different heights, and parabolic objects of different sizes and types. Furthermore, considering computational performance and model transfer issues during deployment, the network cannot be too large. Therefore, the acquisition of training materials and the training process of the deep convolutional neural network are as follows: The image of the detected target area corresponding to the real high-altitude parabolic target is used as a temporal stitched image of positive samples; high-altitude parabolic monitoring images are collected at preset time intervals, and the building outlines in the high-altitude parabolic monitoring images are determined as the bounding box range; corresponding cropping regions are determined within the bounding box range to obtain multiple target cropping regions, and all the target cropping regions are stitched together to obtain a temporal stitched image of negative samples; a training dataset is constructed using the temporal stitched images as positive samples and the temporal stitched images as negative samples; the initial deep convolutional neural network is trained using the training dataset to obtain the trained deep convolutional neural network. Typically, because parabolic targets are too small and feature points are not obvious, the existing methods of directly scaling the detection target for training and judgment are very inaccurate. Therefore, this application constructs and trains the network by combining the image stitching method with inter-frame target information and the negative sample extraction method, as described above, which increases the pixel ratio of falling targets in a single frame, thereby improving the accuracy of classification.

[0115] For positive samples, i.e., real objects thrown from high altitudes, the processing involves the aforementioned slicing, stitching, and cropping process. Specifically, video frames of real objects thrown from high altitudes are acquired, and the real objects are identified within consecutive video frames. Then, based on the coordinates of the trajectory points, horizontal slices are created in the corresponding video frames where the real objects appear. All slices are then scaled according to time sequence, and the scaled slices from consecutive frames are stitched together to obtain a stitched image. Finally, based on the detection bounding boxes, any horizontally redundant pixels in the stitched image are cropped to obtain a cropped stitched image. Figure 5 As shown, the cropped and stitched sliced ​​image is finally scaled to the model input size to obtain a time-series stitched image of the parabolic target for positive samples used in network training. Since obtaining high-altitude parabolic material is difficult, this application also employs data augmentation to reasonably expand the training dataset. In addition to conventional data augmentation methods, for positive samples, parabolic segmentation is used to extract the parabolic objects from all frames during the parabolic process. Based on the absolute position of the parabolic object in the image input to the network, the parabolic object is stitched onto multiple selected building background images, thus ensuring the richness of the background.

[0116] Furthermore, the step of determining the corresponding cropping region within the selected area to obtain multiple target cropping regions specifically includes: within the selected area, determining a first point as the x-axis coordinate by averaging random point selection in the x-axis direction, and randomly selecting a point as the y-axis coordinate in the y-axis direction from a normal distribution established based on the selected area; determining the upper left corner coordinate of the first target cropping region based on the x-axis coordinate and the y-axis coordinate; determining the first target cropping region based on the upper left corner coordinate of the first target cropping region, the height of the scaled slice image, and the image input width specified by the initial deep convolutional neural network; and statistically analyzing the first-order difference of the x-axis coordinates of the high-altitude projectile target in the positive samples. The corresponding extreme value interval is determined, and a corresponding uniform distribution is established based on the extreme value interval; the result of random sampling within the uniform distribution is determined as the initial downward velocity of the x-axis of the pseudo parabola; the average initial velocity of the y-axis of the high-altitude parabola target in the positive sample is statistically analyzed, and the average initial velocity of the y-axis is determined as the initial downward velocity of the y-axis of the pseudo parabola; based on the upper left corner coordinates corresponding to the first target cropping region, the initial downward velocity of the x-axis and the initial downward velocity of the y-axis, the upper left corner coordinates of multiple target cropping regions are determined; multiple target cropping regions are determined based on the upper left corner coordinates of multiple target cropping regions, the height of the scaled slice image and the image input width specified by the initial deep convolutional neural network.

[0117] In other words, for negative samples, which consist of a parabolic background and a rainy or snowy background, the above-mentioned slicing, cropping, and splicing process for positive samples cannot be applied because there is no moving target to detect. Therefore, a "pseudo-parabolic" negative sample simulation method is adopted.

[0118] For example, the background of high-altitude object throwing surveillance images is unlikely to occur within a short period. Therefore, in the same surveillance video segment, images are taken from the background at relatively long time intervals. The extracted images are then framed around the building outline. Within this frame, the top-left corner coordinates of the first target cropping area are determined first. Specifically, on the x-axis, points are randomly sampled using an average method to determine the x-coordinate of the top-left corner of the first target cropping area. On the y-axis, 3 / 4 of the frame height is used as the mean, and 1 / 6 of the frame height is used as the standard deviation. A normal distribution is established based on this mean and standard deviation. Random points are then sampled within this normal distribution to determine the y-coordinate of the top-left corner of the first target cropping area. Thus, the top-left corner coordinates of the first target cropping area are obtained from these x-coordinates and y-coordinates. The upper left corner coordinates are used to determine the first target clipping region. The height h of this first target clipping region is consistent with that of the positive sample, and the width is consistent with the network input. Then, the average initial y-axis velocity of the real high-altitude projectile targets in the positive sample is statistically analyzed. This average initial y-axis velocity is used as the downward y-axis initial velocity of the "fake projectile". The extreme value intervals corresponding to the first difference of the x-coordinate in the positive sample are statistically analyzed. A uniform distribution is established based on this extreme value interval. The results of random sampling in this uniform distribution are used as the downward y-axis initial velocity of the "fake projectile". The acceleration in the y-axis direction is determined by the average second difference of the y-coordinate sequence in the positive sample. Then, starting from the upper left corner coordinates of the first target clipping region, n points are established sequentially, and the same method is used for clipping. Figure 6 As shown, the final image is a time-series stitched image used as a negative sample.

[0119] As can be seen, in this embodiment of the application, the hierarchical detection and identification strategy prioritizes the detection of objects thrown from high altitudes based on trajectory information and EDR distance, which have low computational complexity. Then, the strong interference targets that cannot be filtered out are input into a deep convolutional neural network for further detection. This can solve the problem of high detection rate and high false detection rate of current high-altitude object throwing algorithms and reduce the false detection rate of high-altitude object throwing.

[0120] Accordingly, this application also discloses a high-altitude object throwing detection device, see [link to relevant documentation]. Figure 7 As shown, the device includes:

[0121] The trajectory determination module 11 is used to determine the motion trajectory corresponding to the detected target to obtain the current trajectory;

[0122] The first detection module 12 is used to determine whether the detected target is a high-altitude projectile target based on the current trajectory;

[0123] EDR distance determination module 13 is used to determine the corresponding standard trajectory and the angle sequence-based EDR distance between the standard trajectory and the current trajectory if the current trajectory indicates that the detected target is the high-altitude projectile target.

[0124] The second detection module 14 is used to determine again whether the detected target is the high-altitude projectile target based on the EDR distance;

[0125] The third detection module 15 is used to, if the detected target is the high-altitude object throwing target, horizontally slice the relevant area of ​​the detected target in the acquired high-altitude object throwing monitoring image to obtain the corresponding slice image, and stitch and crop all the slice images related to the high-altitude object throwing target in consecutive frames to obtain the corresponding detection target area image, and input the detection target area image into a pre-trained deep convolutional neural network for further detection to obtain the judgment result of whether the high-altitude object throwing target output by the deep convolutional neural network is a real high-altitude object throwing target.

[0126] As can be seen from the above, in this embodiment of the application, the hierarchical detection and identification strategy prioritizes the detection of objects thrown from high altitudes based on trajectory information and EDR distance, which have low computational complexity. Then, the strong interference targets that cannot be filtered out are input into a deep convolutional neural network for further detection. This can solve the problem of high detection rate and high false detection rate of current high-altitude object throwing algorithms and reduce the false detection rate of high-altitude object throwing.

[0127] In some specific embodiments, the first detection module 12 may specifically include:

[0128] The category determination submodule is used to determine the category of the current trajectory and obtain the preset feature value corresponding to the category; the category is any one of the following: high parabola point high standard deviation, high parabola point low standard deviation, low parabola point high standard deviation, or low parabola point low standard deviation.

[0129] The feature value determination submodule is used to determine the 4-dimensional feature values ​​of the current trajectory; the 4-dimensional features include the first-order difference mean of the x-axis coordinate sequence, the first-order difference mean of the y-axis coordinate sequence, the second-order difference mean of the x-axis coordinate sequence, and the second-order difference mean of the y-axis coordinate sequence.

[0130] The difference percentage determination module is used to calculate the difference percentage between the 4-dimensional feature value of the current trajectory and the preset feature value;

[0131] The first parabolic target detection submodule is used to determine whether the detected target corresponding to the current trajectory is a high-altitude parabolic target based on the comparison result between the difference percentage and the first preset threshold.

[0132] In some specific embodiments, the first parabolic target detection submodule may specifically include:

[0133] The fifth determination unit is used to determine that the detected target corresponding to the current trajectory is a non-high-altitude projectile target if the comparison result shows that the difference percentage is greater than the first preset threshold, and to filter the current trajectory.

[0134] The sixth determination unit is used to determine that the detected target corresponding to the current trajectory is a high-altitude projectile target if the comparison result shows that the percentage difference is not greater than the first preset threshold.

[0135] In some specific embodiments, the EDR distance determination module 13 may specifically include:

[0136] The standard trajectory determination submodule is used to determine the corresponding standard trajectory based on the 4-dimensional feature values ​​and the preset feature values.

[0137] In some specific embodiments, the EDR distance determination module 13 may specifically include:

[0138] An angle sequence acquisition submodule is used to acquire a first angle sequence corresponding to the current trajectory and a second angle sequence corresponding to the standard trajectory; the first angle sequence is a sequence formed by arranging the inferior angles formed by three adjacent points in the current trajectory in order, and the second angle sequence is a sequence formed by arranging the inferior angles formed by three adjacent points in the standard trajectory in order;

[0139] The EDR distance determination submodule is used to determine the angle sequence-based EDR distance between the standard trajectory and the current trajectory based on the first angle sequence and the second angle sequence.

[0140] In some specific embodiments, the second detection module 14 may specifically include:

[0141] The comparison submodule is used to compare the EDR distance with a comparison threshold; the comparison threshold is a threshold determined based on the number of elements in the longer sequence of the current trajectory and the standard trajectory.

[0142] The first determination submodule is used to determine that the detected target is the non-high-altitude projectile target if the EDR distance is greater than the comparison threshold.

[0143] The second determination submodule is used to determine the detected target as the high-altitude projectile target if the EDR distance is not greater than the comparison threshold.

[0144] In some specific embodiments, the third detection module 15 may specifically include:

[0145] The image slicing submodule is used to perform horizontal slicing of the high-altitude object throwing monitoring image based on the coordinates of the trajectory points of the detected target appearing in the acquired high-altitude object throwing monitoring image to obtain a sliced ​​image containing the detected target;

[0146] The image scaling submodule is used to scale the sliced ​​image in time sequence to obtain scaled sliced ​​images.

[0147] The image stitching submodule is used to stitch together all the scaled slice images to obtain a stitched slice image;

[0148] The image cropping submodule is used to crop the stitched sliced ​​image based on the detection box to obtain the detection target region image.

[0149] In some specific embodiments, the high-altitude object throwing detection device may further include:

[0150] The positive sample determination module is used to use the image of the detected target area corresponding to the real high-altitude projectile target as a time-series stitched image of the positive sample;

[0151] The bounding box range determination module is used to collect the high-altitude littering monitoring images at preset time intervals and determine the building outlines in the high-altitude littering monitoring images as the bounding box range;

[0152] The cropping region determination module is used to determine the corresponding cropping region within the selected area to obtain multiple target cropping regions.

[0153] The negative sample determination module is used to stitch together all the target cropped regions to obtain a temporal stitched image as a negative sample;

[0154] A training set construction module is used to construct a training dataset using temporally stitched images as positive samples and temporally stitched images as negative samples.

[0155] The network training module is used to train the initial deep convolutional neural network using the training dataset to obtain the trained deep convolutional neural network.

[0156] In some specific embodiments, the cropping area determination module may specifically include:

[0157] The first coordinate determination submodule is used to determine the first point as the x-axis coordinate in the x-axis direction by means of average random point selection within the selected area, and to randomly select a point as the y-axis coordinate in the y-axis direction from the normal distribution established based on the selected area, and to determine the upper left corner coordinate of the first target clipping area based on the x-axis coordinate and the y-axis coordinate.

[0158] The first cropping region determination submodule is used to determine the first target cropping region based on the upper left corner coordinates of the first target cropping region, the height of the scaled slice image, and the image input width specified by the initial deep convolutional neural network.

[0159] The extreme value interval statistics submodule is used to count the extreme value intervals corresponding to the first difference of the x-axis coordinate of the high-altitude parabolic target in the positive sample, and to establish a corresponding uniform distribution based on the extreme value intervals.

[0160] The x-axis downward initial velocity determination submodule is used to determine the x-axis downward initial velocity of the pseudo-parabola based on the results of random sampling within the uniform distribution.

[0161] The y-axis downward initial velocity determination submodule is used to calculate the average y-axis initial velocity of the high-altitude projectile target in the positive sample, and determine the average y-axis initial velocity as the y-axis downward initial velocity of the false projectile.

[0162] The second coordinate determination submodule is used to determine the upper left corner coordinates of multiple target clipping regions based on the upper left corner coordinates corresponding to the first target clipping region, the initial downward velocity of the x-axis, and the initial downward velocity of the y-axis.

[0163] The second cropping region determination submodule is used to determine multiple target cropping regions based on the upper left corner coordinates of multiple target cropping regions, the height of the scaled slice image, and the image input width specified by the initial deep convolutional neural network.

[0164] Furthermore, embodiments of this application also provide an electronic device. Figure 8 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0165] Figure 8This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the high-altitude object throwing detection method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0166] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0167] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0168] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the high-altitude object throwing detection method disclosed in any of the foregoing embodiments by the electronic device 20, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0169] Furthermore, embodiments of this application also disclose a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the high-altitude object throwing detection method disclosed in any of the foregoing embodiments.

[0170] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0171] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0172] The above provides a detailed description of the high-altitude object throwing detection method, apparatus, equipment, and storage medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A high-altitude projectile detection method, characterized by, The method comprises the following steps: determining a current trajectory corresponding to a detection target, and judging whether the detection target is a high-altitude projectile target based on the current trajectory; if the current trajectory indicates that the detection target is the high-altitude projectile target, determining a corresponding standard trajectory, and determining an EDR distance based on an angle sequence between the standard trajectory and the current trajectory; based on the EDR distance, determining again whether the detection target is the high-altitude projectile target; if the detection target is the high-altitude projectile target, performing horizontal slicing on a relevant region of the detection target in a collected high-altitude projectile monitoring image to obtain a corresponding slice image, and performing splicing and cropping on all slice images related to the high-altitude projectile target in continuous frames to obtain a corresponding detection target region image, and inputting the detection target region image into a pre-trained deep convolutional neural network for further detection to obtain a judgment result of the deep convolutional neural network on whether the high-altitude projectile target is a real high-altitude projectile target; wherein the step of performing horizontal slicing on the relevant region of the detection target in the collected high-altitude projectile monitoring image to obtain the corresponding slice image, and performing splicing and cropping on all slice images related to the high-altitude projectile target in continuous frames to obtain the corresponding detection target region image comprises the following steps: performing horizontal slicing on the high-altitude projectile monitoring image according to the trajectory point coordinates of the detection target in the collected high-altitude projectile monitoring image to obtain a slice image containing the detection target, and performing height scaling on the slice images in time sequence to obtain scaled slice images; splicing all the scaled slice images to obtain spliced slice images; cropping the spliced slice images to obtain a detection target region image.

2. The high-altitude projectile detection method of claim 1, wherein, The step of judging whether the detection target is a high-altitude projectile target based on the current trajectory comprises the following steps: determining a category of the current trajectory, and obtaining a preset characteristic value corresponding to the category; the category is any one of high projectile point high standard deviation, high projectile point low standard deviation, low projectile point high standard deviation, or low projectile point low standard deviation; determining a 4-dimensional characteristic value of the current trajectory; the 4-dimensional characteristic includes a first-order difference mean of an x-axis coordinate sequence, a first-order difference mean of a y-axis coordinate sequence, a second-order difference mean of the x-axis coordinate sequence, and a second-order difference mean of the y-axis coordinate sequence; calculating a difference percentage between the 4-dimensional characteristic value of the current trajectory and the preset characteristic value; determining whether the detection target corresponding to the current trajectory is a high-altitude projectile target based on a comparison result between the difference percentage and a first preset threshold.

3. The high-altitude projectile detection method of claim 2, wherein, The step of determining whether the detection target corresponding to the current trajectory is a high-altitude projectile target based on the comparison result between the difference percentage and the first preset threshold comprises the following steps: if the comparison result indicates that the difference percentage is greater than the first preset threshold, determining that the detection target corresponding to the current trajectory is a non-high-altitude projectile target, and filtering the current trajectory; If the comparison result indicates that the difference percentage is not greater than the first preset threshold, it is determined that the detection target corresponding to the current trajectory is a high-altitude projectile target.

4. The high-altitude projectile detection method of claim 2, wherein, The corresponding standard trajectory is determined, including: The corresponding standard trajectory is determined based on the 4-dimensional characteristic value and the preset characteristic value.

5. The high-altitude projectile detection method of claim 4, wherein, The angle sequence-based EDR distance between the standard trajectory and the current trajectory is determined, including: A first angle sequence corresponding to the current trajectory and a second angle sequence corresponding to the standard trajectory are obtained; the first angle sequence is a sequence formed by sequentially arranging inferior angles composed of adjacent three points in the current trajectory, and the second angle sequence is a sequence formed by sequentially arranging inferior angles composed of adjacent three points in the standard trajectory; The angle sequence-based EDR distance between the standard trajectory and the current trajectory is determined based on the first angle sequence and the second angle sequence.

6. The high-altitude projectile detection method of claim 1, wherein, The EDR distance is compared with a comparison threshold; the comparison threshold is a threshold value determined based on the number of elements in the longer sequence of the current trajectory and the standard trajectory; If the EDR distance is greater than the comparison threshold, it is determined that the detection target is a non-high-altitude projectile target; If the EDR distance is not greater than the comparison threshold, it is determined that the detection target is the high-altitude projectile target. Further comprising:

7. The overhead object detection method of claim 1, wherein, The detection target area image corresponding to the real high-altitude projectile target is taken as a time sequence splicing image of a positive sample; The high-altitude projectile monitoring image is collected at a preset time interval, and the building contour in the high-altitude projectile monitoring image is determined as a frame selection range; A plurality of target cropping regions are determined in the frame selection range to obtain a plurality of target cropping regions, and all the target cropping regions are spliced to obtain a time sequence splicing image of a negative sample; A training data set is constructed using the time sequence splicing image of the positive sample and the time sequence splicing image of the negative sample; An initial deep convolutional neural network is trained using the training data set to obtain the trained deep convolutional neural network. The plurality of target cropping regions are determined in the frame selection range, including:

8. The high-altitude projectile detection method of claim 7, wherein, In the frame selection range, a first point is determined as an x-axis coordinate by an average random point selection method in the x-axis direction, a point is randomly selected as a y-axis coordinate in a normal distribution established based on the frame selection range in the y-axis direction, and the left upper corner coordinate of a first target cropping region is determined according to the x-axis coordinate and the y-axis coordinate; The first target cropping region is determined according to the left upper corner coordinate of the first target cropping region, the height of the scaled slice image, and the image input width specified by the initial deep convolutional neural network; The extreme value interval corresponding to the first-order difference of the x-axis coordinate of the high-altitude projectile target in the positive sample is counted, and a corresponding uniform distribution is established according to the extreme value interval; The result randomly sampled in the uniform distribution is determined as the x-axis downward initial velocity of a false projectile. ​ statistically average the initial y-axis velocities of the high-altitude projectile targets in the positive samples, and determine the average initial y-axis velocities as the initial y-axis velocities of the false projectiles; determine the top-left coordinates of the target cropping regions based on the top-left coordinates corresponding to the first target cropping region, the x-axis initial downward velocities, and the y-axis initial downward velocities; determine the target cropping regions based on the top-left coordinates of the target cropping regions, the height of the scaled slice images, and the image input width specified by the initial deep convolutional neural network.

9. A high-altitude projectile detection apparatus, characterized by, comprise: a trajectory determination module configured to determine a motion trajectory corresponding to a detection target to obtain a current trajectory; a first detection module configured to determine whether the detection target is a high-altitude projectile target based on the current trajectory; an EDR distance determination module configured to determine a standard trajectory corresponding to the detection target and an EDR distance based on an angle sequence between the standard trajectory and the current trajectory if the current trajectory indicates that the detection target is the high-altitude projectile target; a second detection module configured to determine whether the detection target is the high-altitude projectile target again based on the EDR distance; a third detection module configured to, if the detection target is the high-altitude projectile target, perform horizontal slicing on a region related to the detection target in a high-altitude projectile monitoring image collected to obtain a corresponding slice image, splice and crop all slice images related to the high-altitude projectile target in consecutive frames to obtain a corresponding detection target region image, and input the detection target region image into a pre-trained deep convolutional neural network for further detection to obtain a judgment result of the deep convolutional neural network on whether the high-altitude projectile target is a real high-altitude projectile target; wherein the third detection module comprises: an image slicing submodule configured to perform horizontal slicing on a high-altitude projectile monitoring image collected based on trajectory point coordinates of the detection target in the high-altitude projectile monitoring image to obtain a slice image containing the detection target; an image scaling submodule configured to scale the slice image in height in time sequence to obtain a scaled slice image; an image splicing submodule configured to splice all the scaled slice images to obtain a spliced slice image; an image cropping submodule configured to crop the spliced slice image based on a detection frame to obtain a detection target region image.

10. An electronic device, comprising: comprise: a memory configured to save a computer program; a processor configured to execute the computer program to implement the steps of the high-altitude projectile detection method according to any one of claims 1 to 8.

11. A computer readable storage medium, characterized in that, for storing a computer program; wherein the computer program is executed by a processor to implement the steps of the high-altitude projectile detection method according to any one of claims 1 to 8.

Images