CN113570005A - Long-distance ship type identification method based on airborne photon radar - Google Patents
Long-distance ship type identification method based on airborne photon radar Download PDFInfo
- Publication number
- CN113570005A CN113570005A CN202111125216.7A CN202111125216A CN113570005A CN 113570005 A CN113570005 A CN 113570005A CN 202111125216 A CN202111125216 A CN 202111125216A CN 113570005 A CN113570005 A CN 113570005A
- Authority
- CN
- China
- Prior art keywords
- ship
- point cloud
- point
- steps
- algorithm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
- G01S17/88—Lidar systems specially adapted for specific applications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/48—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S17/00
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
Abstract
The invention provides a long-distance ship type identification method based on an airborne photon counting radar, which comprises the following steps: receiving target original information of an airborne photon counting radar for detecting a marine target, and performing preprocessing on the target original information, wherein the preprocessing comprises the following steps: point cloud denoising, sea surface point cloud fitting and point cloud clustering; for the preprocessed target signal, extracting ship features, wherein the ship features comprise: vessel dimension characteristics, vessel geometric characteristics, deck shape characteristics and normal vector histograms; the dimension characteristics of the ship comprise the length, the width and the height of the ship; the geometric characteristics of the ship refer to the spatial distribution characteristics of buildings on the deck, and the shape characteristics of the deck refer to the included angle of the ship head; inputting the extracted ship features into a classifier to perform ship identification; and outputting a ship identification result.
Description
Technical Field
The invention belongs to the technical field of optical radars, and relates to a long-distance ship type identification method based on an airborne photon radar.
Background
In recent years, with the development of society, activities on the sea of various countries are increasing, the safety and stability of the countries are related to the ocean safety, and the strengthening of the real-time identification of marine ships has important significance in the fields of marine rescue, ship management and the like. At present, almost all of the coastal countries put more energy into the field of ship identification, and seek to develop farther and more real-time identification technology.
Photon counting laser radar is an active three-dimensional detection technology developed in recent years, and is one of the research hotspots in the detection field at present. Compared with the conventional laser radar, the photon counting radar has the advantages that the detection sensitivity is improved by 2-3 orders of magnitude, and photon level weak signals returned by a long-distance target can be detected, so that a target three-dimensional point cloud model is generated. The point cloud model contains three-dimensional characteristics of the target, characteristics of the target can be well described, the photon counting radar is slightly affected by weather, the detection distance is long, and the point cloud model has great advantages in the field of recognition.
The ship identification firstly needs to obtain target related information by using a certain detection means, then extracts characteristics from the target related information, and realizes the ship type identification through comparison with a characteristic library. The detection means commonly used at present include optical shooting, radar detection, laser radar and the like, but the technologies have the problems of large weather influence, poor real-time performance, short working distance, complex identification process, poor characteristic description and the like.
In conclusion, compared with the conventional ship identification method, the ship identification method by the airborne photon counting radar has the advantages of all-weather work, strong real-time performance and long working distance, and the identification precision is greatly improved.
As can be seen from research on documents in the field of ship identification, the current ship identification technologies mainly include: 1. the ship identification is realized by deep learning aiming at the two-dimensional image extraction characteristics of the ship; 2. the characteristics of the echo signals of the ships are extracted through a radar, and the ship identification is realized through a traditional classifier; 3. a three-dimensional point cloud model is generated for a target through a laser radar, and ship identification is achieved through a traditional classifier after three-dimensional features are extracted.
The algorithm steps common to ship identification techniques in the prior art are shown in fig. 1.
Current identification techniques follow the above-described procedure. Firstly, detecting a target by the existing detection means to obtain target original information, and then preprocessing the target original information to remove part of useless information or adjust the format of the target information. And extracting features from the preprocessed information, inputting the extracted features into a classifier, comparing the features with a feature library, and outputting a result with the maximum possibility.
The defects of the prior art are as follows:
1. ship identification by means of images: in practical application, the influence of adverse weather is large, and the neural network needs large-scale data to train, so that the process is complex and long in time consumption.
2. Rely on radar echo signal discernment naval vessel: most of the features extracted from the radar echo signals are one-dimensional features and two-dimensional features, the target description is not strong, or the target dynamic needs to be monitored for a long time, and the real-time performance is poor.
3. Rely on laser radar echo signal discernment naval vessel: the method is limited by the working distance of the radar, and can not identify ships with long distance beyond 2 km.
Disclosure of Invention
In order to solve the technical problem, the invention provides a long-distance ship type identification method based on an airborne photon counting radar, which comprises the following steps
Step 1, receiving target original information of a marine target detected by an airborne photon counting radar,
step 2, preprocessing the target original information, wherein the preprocessing comprises the following steps: point cloud denoising, sea surface point cloud fitting and point cloud clustering;
step 3, aiming at the preprocessed target signal, extracting ship features, wherein the ship features comprise: vessel dimension characteristics, vessel geometric characteristics, deck shape characteristics and normal vector histograms; the dimension characteristics of the ship comprise the length, the width and the height of the ship; the geometric characteristics of the ship refer to the spatial distribution characteristics of buildings on the deck, and the shape characteristics of the deck refer to the included angle of the ship head;
step 4, inputting the extracted ship features into a classifier to execute ship identification;
and 5, outputting a ship identification result.
Further, step 2 comprises the sub-steps of:
step 2.1, point cloud denoising, wherein the point cloud denoising comprises the following steps: removing noise of target original information by setting a density threshold and using a discrete point cloud extraction algorithm;
step 2.2, fitting the sea surface point cloud, setting a height threshold value aiming at the sea surface wave fluctuation, and removing the sea surface point cloud by using a plane fitting algorithm;
and 2.3, clustering the point clouds of the targets, and classifying the point clouds of the same ship into one class by using a clustering algorithm so as to extract ship features in the subsequent steps.
Further, the plane fitting algorithm comprises a RANSAC plane fitting algorithm; the clustering algorithm comprises a DBSCAN algorithm.
Preferably, the DBSCAN algorithm comprises the sub-steps of:
step 2.31, establishing a spatial topological relation for the point cloud data,
step 2.32, taking any point as the center of circle, calculating the number minpts of the adjacent points in the radius espi，minptsiIs a point piThe density of (a) of (b),
wherein p isiAs a coordinate of the center point, pjIs a pointSetting the coordinates of other points of the cloud set as fixed values, calculating the mean value minpts of the point cloud density of the data set in a self-adaptive manner,
and 2.33, substituting minpts into a DBSCAN algorithm for calculation to obtain a classification result.
Further, step 3 comprises the sub-steps of:
step 3.1, extracting the length, the width and the height of the point cloud of the same ship as the dimension characteristics of the three-dimensional model;
3.2, extracting the geometric characteristics of the upper-layer building by using a concentric circle segmentation method, wherein the optimal number of rings in the concentric circle segmentation method is selected according to the size of the ship body;
and 3.3, extracting the ship bow angle.
Preferably, step 3.2 comprises the sub-steps of:
3.21, carrying out coordinate deflection on the detected scene to enable the scene plane to coincide with the earth horizontal plane;
step 3.22, dividing the z interval into a plurality of sub-blocks, and counting the point cloud number of each sub-block;
step 3.23, counting the height of the sub-block in the interval with the highest proportion to determine the height of the deck, and extracting the point cloud with the z height larger than the height of the deck to obtain a data set F;
step 3.24, respectively calculating the mean values of all points in the data set F in the x axis, the y axis and the z axis to obtain a central point;
step 3.25, traversing each point, and calculating the distance from each point to the central point;
and 3.26, dividing each point into rings with different radiuses according to the distance from the central point, and calculating the ratio of the number of the central points in each ring to obtain the geometric characteristics.
Further, step 3.3 comprises the sub-steps of:
step 3.31, extracting a point cloud of the bow or the stern of the ship;
step 3.32, dividing the bow point cloud into a plurality of sections in parallel to the long axis direction of the ship, and taking the point of each section closest to the radar;
step 3.33, dividing the extracted points into 2 groups by taking the median as a boundary, projecting the points to an x-y plane, fitting a straight line by using a least square method, selecting vectors a and b on the straight line, and calculating an angle, wherein an angle calculation formula is as follows:
wherein | represents an absolute value.
Further, step 4 comprises the sub-steps of:
step 4.1, executing the steps 1 to 3 to obtain known ship feature data sets distinguished according to types, and storing the data sets into a ship feature database;
step 4.2, inputting the target original information of the marine target acquired in real time at sea into a classifier to perform classification;
and 4.3, comparing the classified information with corresponding information of a ship characteristic database, and determining the ship type of the detected marine target.
Preferably, the classifier is a classifier which adopts a random forest algorithm to carry out deep training.
The method of the invention provides a simple and effective method for accurately identifying ships in a long distance. The problems of complex training, short working distance, low recognition rate and poor real-time performance of the conventional ship recognition algorithm are solved, accurate ship type recognition under a long-distance condition is successfully realized, and the method has guiding significance on the practical application of the photon radar.
Drawings
FIG. 1 is a diagram of the algorithm steps common to prior art ship identification techniques;
FIG. 2 is an algorithmic flow chart of the present invention;
FIG. 3 is a denoised point cloud model of an offshore exploration scene;
FIG. 4 is a scene after the sea surface is removed;
FIG. 5 is a clustered scene;
FIG. 6 is a concentric circle segmentation method;
FIG. 7 is an angle between two ends of a ship;
FIG. 8 is a flow chart of an embodiment of the present invention.
Detailed Description
The invention starts from the practical requirements and application angles, researches the type identification problem of the long-distance sea surface ship under the detection condition of the airborne photon counting radar, mainly depends on the weak signal receiving capacity of the photon counting radar, researches how to extract three-dimensional characteristics and optimize the characteristics under the condition that the point cloud density of a target model is sparse, improves the identification capacity of characteristic combination, and increases the algorithm identification precision.
The following detailed description of embodiments of the invention refers to the accompanying drawings.
The general technical scheme of the invention is shown in figure 2.
Point cloud denoising:
in photon detection data, the signal point cloud and the noise point cloud have a difference of 2-3 orders of magnitude in density, most of noise can be removed by setting a density threshold and a discrete point cloud extraction algorithm, formula (1) is a defined radial basis function, and formula (2) is a density value calculation formula.
Operator in the formula (1)
Representing a point of proximity within a predetermined areaxTo the center point of the regioncThe distance of (c).
In the formula (2)
And D is a previously given area.
By the processing of the step, the background noise in the sea surface detection data of the photon counting radar is almost completely removed, and only a few residual noise point clouds can be ignored in the following ship extraction and model identification.
Fitting of sea surface point cloud
After the point cloud is denoised, the sea point cloud and the ship point cloud are still left in the detection scene point cloud model, and fig. 3 is the denoised sea detection scene point cloud model. 3 ships are distributed in the graph, the graph is mainly used for detecting the fitting effect of the point cloud of the sea surface, and the types of the ships in practical experiments are far larger than 3.
The sea surface point cloud is similar to a plane in spatial distribution, under the normal weather condition, the fluctuation of sea waves is between 5 meters, the sea surface point cloud can be obtained and distributed in a bounding box range with the height of 10 meters, therefore, the distance threshold value is set to be 10 meters, the sea surface is fitted by using the classical RANSAC plane fitting algorithm, the fitted plane point cloud is removed, the rest is the ship point cloud, and the figure 4 is the result after the sea surface point cloud is removed.
Part of the formula for the plane fit is as follows:
the formula (3) is a plane equation,a 0、a 1、a 2is the coefficient to be found. Equation (4) is the sum of the distances of each point from the plane, and when C is minimum, the plane coefficient is obtained.
Point cloud clustering
After the preprocessing, background noise point clouds and sea surface point clouds are basically and completely removed, only one or more ship point clouds are left in a scene, but for machine vision, the point clouds in the scene are an integral body, clustering is needed before extracting features, the same ship point clouds are classified into one type, and then the features are gradually extracted.
The DBSCAN algorithm is a density-based clustering algorithm, parameters of the traditional DBSCAN algorithm depend on manual selection, and over-segmentation or under-segmentation results can occur when the DBSCAN algorithm is applied to different data sets. In order to make up for the defects of the DBSCAN algorithm, the invention firstly establishes a spatial topological relation for point cloud data, and then takes any point as the pointThe number minpts of the adjacent points in the radius eps is calculated at the center of the circlei，minptsiIs a pointp i The density of (c) is calculated as in equation (5).
Whereinp i Is the coordinate of the central point, and the central point,p j and coordinates of other points of the point cloud set.
In order to reduce variable errors, eps is set as a fixed value 20, and the mean value minpts of the data set point cloud density is calculated in a self-adaptive calculation mode to obtain a classification result. The improved algorithm can automatically calculate parameters according to the density of the data set and perform segmentation, and the problem that the original DBSCAN algorithm is segmented by manually setting the parameters is solved.
The modified DBSCAN algorithm is applied to fig. 4, and the segmented result is shown in fig. 5. The classification result is represented by color in fig. 5, and it can be seen that the scene is divided into 3 clustering results, wherein the entire point clouds of the ships are almost classified into one category, which is consistent with the expected result, and few discrete point clouds are determined as noise, but considering that the amount of the ships and the characteristics contained in the discrete point clouds are very few, the main characteristic extraction of the ships is not affected, so the improved clustering algorithm has a good clustering effect on the ships, and achieves the expected target.
Feature extraction:
after clustering, different types of ship point clouds are already classified into a point cloud set, so that the clustered result can be directly used for feature extraction. Under the condition of remote detection, the obtained target point clouds are sparse, the number of ship model point clouds with the length less than 100 meters is very small, and the ship model point clouds are easily submerged in noise, so that the ship length distribution researched by the invention is between 100 meters and 300 meters.
The length, width and height characteristics of the ship target shot in the air hardly change along with the change of the shooting angle, and in the experimental data of the invention, the average change rate of the length, width and height of the ship under the detection of different azimuth angles is below 10 percent, so the length, width and height of the three-dimensional modelThe high feature has strong stability. In summary, the length of the present invention is selectedl) Width (a)w) High, high (h) Reflecting the volume of the target as a dimension characteristic, and calculating the formula in a photon counting radar coordinate system as follows:
the main characteristics of the ship target are concentrated on the upper half part, and under the detection condition of the airborne photon counting radar, no matter what posture the ship is in, the deck and the buildings above the deck can be almost completely detected, so that the space distribution characteristics of the buildings above the deck must be excavated to achieve the target of accurately identifying the ship type. Fig. 6 is a concentric circle segmentation method, which includes firstly calculating the distance between each point and the center point in the data set, then dividing the data set into rings according to the different distances, and finally calculating the ratio of the number of the points in each ring, wherein the statistical result is the geometric characteristic.
Before extracting geometric features, the upper-level building point cloud needs to be extracted from the whole ship model, and the steps are as follows: 1. according to the previously fitted plane, carrying out coordinate deflection on the detected scene to enable the scene plane to coincide with the horizontal plane of the earth; 2. dividing the z interval into a plurality of sub-blocks, and counting the point cloud number of each sub-block; 3. and (4) counting the height of the subblock in the interval with the highest ratio, namely the height of the deck, and extracting the point cloud with the z height larger than the height of the deck to obtain a data set F. The following is the calculation of the geometric features:
1. and respectively calculating the mean values of all points in the data set F in the x axis, the y axis and the z axis to obtain a central point.
2. And traversing each point, and calculating the distance from each point to the central point.
3. Dividing each point into different rings according to the distance, and then calculating the ratio of the number of the middle points in each ring to obtain the geometric characteristics.
In order to obtain the optimal number of ring divisions, 6 targets of a cruiser, an aircraft carrier, a guard ship, a cargo ship, a container ship and a mail ship are selected as experimental objects, and a random forest classifier is used for researching the optimal number of rings of a circular space division method under the condition of sparse point cloud.
The number of the divided rings is respectively set to be 3 to 10, the correct recognition rate of each division method is tested, and the test results are shown in table 1.
TABLE 1 round space division method experimental results
It can be seen from table 1 that, as the number of the dividing rings increases, the correct identification rates of 6 ships also increase and tend to be stable, wherein the 3-ring division is unstable in the experiment, the experiment result is basically stable and at a higher level from the 4-ring division, and the 4-ring division is adopted to represent the geometric characteristics of the target under the consideration of the data dimension and the identification correct rate.
In the case of the on-board photon counting radar illumination mentioned above, the deck is fully detectable in addition to the upper buildings. The size characteristics of the deck are repeated with the above characteristics, so the invention extracts the included angle characteristic of the shape characteristics of the deck, namely the ship head included angle.
The bow angles of different types of ships are often different, and the bow angle of a warship is smaller than that of a civil ship. In radar detection, under the condition that no manual work participates, the radar cannot distinguish the bow and the stern, so that the invention extracts point clouds at the foremost end and the rearmost end of a target along the length of a ship (firstly, the end close to the radar is defaulted to be the bow), calculates corresponding angles, and uses a smaller angle as the bow angle. The method comprises the following steps:
1. extracting a bow (stern) point cloud;
2. the point cloud at the bow (stern) is divided into a plurality of sections in the direction parallel to the long axis of the ship, and the point with the minimum y value (closer to the radar) in each section is taken.
3. Dividing the extracted points into 2 groups by taking the median as a boundary, projecting the points to an x-y surface, fitting a straight line by using a least square method, and selecting vectors on the straight linea、bAnd calculating the angle according to the following calculation formula.
Fig. 7 is a fitting straight line of the ship bow and the ship stern, the left image is a model obtained by detecting the ship with the airborne photon counting radar from the right rear direction in the sky, the right image is a model obtained by detecting the cargo ship with the airborne photon counting radar from the right front direction in the sky, and a red line segment is a fitting straight line. It can be seen from the figure that the fitted straight line better reflects the curves of the bow and the stern, and the included angle between the bow and the stern is successfully obtained.
The normal vector histogram can display the distribution of normal vectors of the target surface, and can better display the three-dimensional characteristics of the target surface in the short-distance radar recognition. In the long-distance photon counting radar detection, the normal vector histogram is still a good classification basis through analysis and optimization, and the calculation formula of the normal vector histogram is as follows.
nIs the total number of the point clouds, m i is the frequency count of the different frequency bands,H i is in proportion to satisfy
，iIs the frequency band number.
The normal vector histogram is obtained by counting the whole detected target, but under the condition of sparse point cloud, many detailed features of the ship are fuzzified, and the target is difficult to distinguish through the conventional normal vector histogram. Similar to the research process, the method selects 6 targets of a cruiser, an aircraft carrier, a protective carrier, a cargo ship, a container ship and a mail ship as experimental objects, and uses a random forest classifier to research the classification capability of a conventional normal vector histogram under the condition of sparse point cloud.
As most ship targets are symmetrical about the long axis of the ship targets, the ship target selecting method selects the vector in the long axis direction as an initial line, calculates the included angle between the normal vector of each point and the initial line, and the included angle is distributed between 0 and 180 degrees. In order to select the histogram feature of the optimal dimension, the method takes 20 degrees, 30 degrees, 45 degrees and 60 degrees as intervals respectively, and calculates the correct recognition rate under different intervals. The following table shows the correct recognition rate for 5 experiments at different intervals.
TABLE 2 Classification of conventional normal vector histograms
As can be seen from table 2, the classification capability of the conventional normal vector histogram under the sparse point cloud condition is very poor, and even if the histogram is divided into 9 dimensions at intervals of 20 degrees, the highest correct recognition rate only reaches 33%, so that it can be seen that the detailed features of the ship target can hardly be displayed under the remote detection condition, and the conventional normal vector histogram can not meet the requirements of the present invention.
In the detection models with different azimuth angles, compared with a side deck, the shielding change of an upper-layer building of a target is the largest, in order to enhance the recognition capability of a normal vector histogram, histogram statistics is carried out on the upper half part of a ship, the normal vector histograms under different intervals are adopted for classification training, and the training results of 5 times are shown in table 3.
TABLE 3 Classification results of Normal vector histograms after optimization
Compared with the table 2, the correct recognition rate of the table 3 is obviously greatly improved, and the optimized normal vector feature histogram is more suitable for sparse point cloud classification under a remote condition. In addition, the average correct recognition rates of different interval angles are respectively 68.4%, 59.8%, 76.8% and 60%, wherein the correct recognition rate of the 4-dimensional normal vector histogram with the interval of 45 degrees is higher, so that the 4-dimensional normal vector histogram is selected as the classification basis.
Random forest classifier
Because no point cloud data set of the airborne photon counting radar to the ship exists at present, the ship model simulated by the method can not meet the requirement of neural network algorithm training amount, so that the method selects a random forest algorithm to train and classify the extracted features, and verifies the classification result.
The random forest algorithm can be regarded as a simplified neural network algorithm and consists of a plurality of decision trees, each decision tree is a weak classifier and outputs a result to input data, the random forest analyzes all output results, and the result with the highest probability is selected as output, and the algorithm comprises the following steps:
1. randomly sampling n samples from the sample set;
2. randomly selecting k features from all the features, and establishing a decision tree by using the features;
3. repeating the steps for m times to form a random forest consisting of m decision trees;
4. for new data, through each tree decision, the decision is finally voted to confirm which category is assigned.
Example 1
The invention relates to a ship type, which adopts 13 ships including a cruiser, an aircraft carrier, a guard ship, a destroyer, a medium-sized aircraft carrier, a medical ship, a landing ship, a crane ship, a cargo ship, a container ship, a scientific investigation ship, a mail ship and a fishing ship to establish a model, wherein the ships comprise the types of ships common to people, and samples are representative.
Point cloud density: the simulated airborne photon counting radar scans the ship model beyond 5km, the horizontal resolution and the vertical resolution of the photon counting radar are both 0.5mrad, and therefore the mean value M of the distances between adjacent point clouds of the ship model is more than or equal to 2.5 meters. Table 4 shows the length, width, and height and M for different types of ships.
Table 4 partial data of different types of ship models
It can be seen from table 4 that the average value of the distance between adjacent point clouds of the ship model is greater than 2.5 meters, which meets the sparse point cloud condition, the last column in table 4 is the height of the ship model, part of the models have 2 heights, the data outside the bracket is the real height of the ship model, but the height of part of the ship is determined by the height of the extremely thin and long mast, which is difficult to detect in the actual detection, and even if the height is detected, the height can be removed as noise in the denoising process, and the height of the data ship main body in the bracket is determined.
It should be noted that, when navigating on the sea, the ship body has draft and occlusion of buildings above the deck, so the dimensional characteristics in the actual sample are smaller than those in table 4, and the dimensional data in table 4 is used for reference only and represents the size relationship between the models.
3) Ship sample: the invention establishes 13 models of ships in different types, and only part of the ship model can be obtained in different directions in actual detection, so that the model needs to be divided along different angles, and the divided ship model is a real sample.
Selecting a point in the air beyond 5km of the ship as the position of the radar, drawing a line segment with corresponding density according to the horizontal resolution and the vertical resolution of the photon counting radar, calculating the intersection point of the line and the model as a detected ship scene, if a plurality of intersection points exist, selecting the closest intersection point of the clustering radar, and obtaining ship samples in different directions by changing the radar position.
Aiming at each ship model, the photon counting radar irradiation simulation is respectively carried out in 7 directions of the ship, namely front right, right back, right left and left back to obtain 7 samples, so that the total number of the ship model is 13, and the number of the samples is 91.
The number of decision trees in the random forest algorithm is set to be 40, wherein samples of 5 directions of front, front right, rear right and rear right are used as training sets, and samples of front left and rear left are used as testing sets. The features extracted from each sample form a group of 12-dimensional arrays, wherein the dimensional features are 3-dimensional data, the normal vector histogram is 4-dimensional data, the geometric features are 4-dimensional data, and the bow included angle is 1-dimensional data. And after a one-dimensional label is added to each sample, establishing a one-to-one relationship between the sample and the 13-dimensional feature array, and inputting the feature array into a random forest for training or testing.
The invention has 13 output results in total, and the table 5 shows the accuracy of 7 tests by using the random forest algorithm.
TABLE 5 random forest Algorithm 7 Classification results
As can be seen from Table 5, the correct recognition rates of 7 operations of the classification algorithm are all higher than 95%, the correct rates of most classification results reach 100%, and the average correct recognition rate is 98.90%, so that the extracted features of the invention can meet the requirement of remote ship recognition.
The 4 steps of point cloud denoising, sea surface fitting, point cloud clustering and classifier category are point cloud preprocessing for the invention, and the denoising algorithm, the fitting algorithm and the clustering algorithm are not only density denoising, RANSAC algorithm, DBSCAN algorithm, random forest classifier and the like mentioned in the invention, but also the four steps are replaceable, and detailed steps are not required to be described in detail in the invention.
The detailed steps of feature extraction:
three-dimensional characteristics:
in summary, the length of the present invention is selectedl) Width (a)w) High, high (h) Reflecting the volume of the target as a dimension characteristic, and calculating the formula in a photon counting radar coordinate system as follows:
geometric characteristics:
before extracting geometric features, the upper-level building point cloud needs to be extracted from the whole ship model, and the steps are as follows: 1. according to the previously fitted plane, carrying out coordinate deflection on the detected scene to enable the scene plane to coincide with the horizontal plane of the earth; 2. dividing the z interval into a plurality of sub-blocks, and counting the point cloud number of each sub-block; 3. and (4) counting the height of the subblock in the interval with the highest ratio, namely the height of the deck, and extracting the point cloud with the z height larger than the height of the deck to obtain a data set F. The following is the calculation of the geometric features:
1. and respectively calculating the mean values of all points in the data set F in the x axis, the y axis and the z axis to obtain a central point.
2. And traversing each point, and calculating the distance from each point to the central point.
3. The farthest distance from each point to the central point is averagely divided into 4 sections, each point is divided into 4 rings according to the distance, and then the ratio of the number of the central points in each ring is calculated to obtain the geometric characteristics.
The included angle of the bow:
1. extracting a bow (stern) point cloud;
2. the point cloud at the bow (stern) is divided into a plurality of sections in the direction parallel to the long axis of the ship, and the point with the minimum y value (closer to the radar) in each section is taken.
3. Dividing the extracted points into 2 groups by taking the median as a boundary, projecting the points to an x-y surface, fitting a straight line by using a least square method, and selecting vectors on the straight linea、bAnd calculating the angle according to the following calculation formula.
Normal vector histogram:
for each point of the data set. The 3 closest points to this point are selected. These 3 points fit a plane whose normal vector is the normal vector of the point.
As most ship targets are symmetrical about the long axis of the ship targets, the vector in the long axis direction is selected as an initial line, the included angle between the normal vector of each point in a ship storey building and the initial line is calculated, and the included angle is distributed between 0 and 180 degrees. Dividing 0-180 into 4 sections at intervals of 45 degrees, calculating the number of points in each section, calculating the ratio of the number of points in each section to the total number of the upper-layer buildings of the ship, and finally calculating the 4 ratios to obtain the 4-dimensional characteristics.
Random forest classifier
The random forest algorithm can be regarded as a simplified neural network algorithm and consists of a plurality of decision trees, each decision tree is a weak classifier and outputs a result to input data, the random forest analyzes all output results, and the result with the highest probability is selected as output, and the algorithm comprises the following steps:
1. randomly sampling n samples from the sample set;
2. randomly selecting k features from all the features, and establishing a decision tree by using the features;
3. repeating the steps for m times to form a random forest consisting of m decision trees;
4. for new data, through each tree decision, the decision is finally voted to confirm which category is assigned.
To obtain the result
The number of decision trees in the random forest algorithm is set to be 40, wherein samples of 5 directions of front, front right, rear right and rear right are used as training sets, and samples of front left and rear left are used as testing sets. The features extracted from each sample form a group of 12-dimensional arrays, wherein the dimensional features are 3-dimensional data, the normal vector histogram is 4-dimensional data, the geometric features are 4-dimensional data, and the bow included angle is 1-dimensional data. And after a one-dimensional label is added to each sample, establishing a one-to-one relationship between the sample and the 13-dimensional feature array, and inputting the feature array into a random forest for training or testing. The results obtained are shown in Table 5.
The ship identification technology by the airborne photon counting radar can break through the limitation of the traditional laser radar ship identification working distance and realize the identification of a long-distance ship target; the three-dimensional features extracted by the technology are stronger in description power of the target features compared with the two-dimensional features extracted from the images, so that high identification accuracy can be achieved without extremely large-scale training; the photon counting radar has high imaging speed and does not need to detect a target for a long time, so that the efficiency of extracting and identifying the features is high.
The 4 steps of point cloud denoising, sea surface fitting, point cloud clustering and classifier category are point cloud preprocessing for the invention, and the denoising algorithm, the fitting algorithm and the clustering algorithm are not only density denoising, RANSAC algorithm, DBSCAN algorithm, random forest classifier and the like mentioned in the invention, but also can be replaced.
Finally, it should be noted that the above embodiments are only used for illustrating the technical solutions of the embodiments of the present invention and not for limiting, and although the embodiments of the present invention are described in detail with reference to the above preferred embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions can be made on the technical solutions of the embodiments of the present invention without departing from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (9)
1. A long-distance ship type identification method based on an airborne photon counting radar is characterized by comprising the following steps
Step 1, receiving target original information of a marine target detected by an airborne photon counting radar,
step 2, preprocessing the target original information, wherein the preprocessing comprises the following steps: point cloud denoising, sea surface point cloud fitting and point cloud clustering;
step 3, aiming at the preprocessed target signal, extracting ship features, wherein the ship features comprise: vessel dimension characteristics, vessel geometric characteristics, deck shape characteristics and normal vector histograms; the dimension characteristics of the ship comprise the length, the width and the height of the ship; the geometric characteristics of the ship refer to the spatial distribution characteristics of buildings on the deck, and the shape characteristics of the deck refer to the included angle of the ship head;
step 4, inputting the extracted ship features into a classifier to execute ship identification;
and 5, outputting a ship identification result.
2. The method of claim 1, wherein step 2 comprises the sub-steps of:
step 2.1, point cloud denoising, wherein the point cloud denoising comprises the following steps: removing noise of target original information by setting a density threshold and using a discrete point cloud extraction algorithm;
step 2.2, fitting the sea surface point cloud, setting a height threshold value aiming at the sea surface wave fluctuation, and removing the sea surface point cloud by using a plane fitting algorithm;
and 2.3, clustering the point clouds of the targets, and classifying the point clouds of the same ship into one class by using a clustering algorithm so as to extract ship features in the subsequent steps.
3. The method of claim 2, wherein said plane fitting algorithm comprises a RANSAC plane fitting algorithm; the clustering algorithm comprises a DBSCAN algorithm.
4. A method according to claim 3, wherein said DBSCAN algorithm comprises the sub-steps of:
step 2.31, establishing a spatial topological relation for the point cloud data,
step 2.32, taking any point as the center of circle, calculating the number minpts of the adjacent points in the radius espi，minptsiIs a point piThe density of (a) of (b),
wherein p isiAs a coordinate of the center point, pjSetting esp as fixed value for the coordinates of other points in the point cloud set, calculating the mean value minpts of the point cloud density of the data set in a self-adaptive manner,
and 2.33, substituting minpts into a DBSCAN algorithm for calculation to obtain a classification result.
5. The method of claim 1, wherein step 3 comprises the sub-steps of:
step 3.1, extracting the length, the width and the height of the point cloud of the same ship as the dimension characteristics of the three-dimensional model;
3.2, extracting the geometric characteristics of the upper-layer building by using a concentric circle segmentation method, wherein the optimal number of rings in the concentric circle segmentation method is selected according to the size of the ship body;
and 3.3, extracting the ship bow angle.
6. A method as claimed in claim 5, characterised in that step 3.2 comprises the sub-steps of:
3.21, carrying out coordinate deflection on the detected scene to enable the scene plane to coincide with the earth horizontal plane;
step 3.22, dividing the z interval into a plurality of sub-blocks, and counting the point cloud number of each sub-block;
step 3.23, counting the height of the sub-block in the interval with the highest proportion to determine the height of the deck, and extracting the point cloud with the z height larger than the height of the deck to obtain a data set F;
step 3.24, respectively calculating the mean values of all points in the data set F in the x axis, the y axis and the z axis to obtain a central point;
step 3.25, traversing each point, and calculating the distance from each point to the central point;
and 3.26, dividing each point into rings with different radiuses according to the distance from the central point, and calculating the ratio of the number of the central points in each ring to obtain the geometric characteristics.
7. A method as claimed in claim 1, characterized in that step 3.3 comprises the sub-steps of:
step 3.31, extracting a point cloud of the bow or the stern of the ship;
step 3.32, dividing the bow point cloud into a plurality of sections in parallel to the long axis direction of the ship, and taking the point of each section closest to the radar;
step 3.33, dividing the extracted points into 2 groups by taking the median as a boundary, projecting the points to an x-y plane, fitting a straight line by using a least square method, selecting vectors a and b on the straight line, and calculating an angle, wherein an angle calculation formula is as follows:
wherein | represents an absolute value.
8. The method of claim 1, wherein step 4 comprises the sub-steps of:
step 4.1, executing the steps 1 to 3 to obtain known ship feature data sets distinguished according to types, and storing the data sets into a ship feature database;
step 4.2, inputting the target original information of the marine target acquired in real time at sea into a classifier to perform classification;
and 4.3, comparing the classified information with corresponding information of a ship characteristic database, and determining the ship type of the detected marine target.
9. The method of claim 8, wherein the classifier is a classifier that is deep trained using a random forest algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111125216.7A CN113570005A (en) | 2021-09-26 | 2021-09-26 | Long-distance ship type identification method based on airborne photon radar |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111125216.7A CN113570005A (en) | 2021-09-26 | 2021-09-26 | Long-distance ship type identification method based on airborne photon radar |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113570005A true CN113570005A (en) | 2021-10-29 |
Family
ID=78174513
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111125216.7A Pending CN113570005A (en) | 2021-09-26 | 2021-09-26 | Long-distance ship type identification method based on airborne photon radar |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113570005A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114782868A (en) * | 2022-04-19 | 2022-07-22 | 网新百橙科技(杭州)有限公司 | Video image ship identification system and method based on AI artificial intelligence |
| CN116148878A (en) * | 2023-04-18 | 2023-05-23 | 浙江华是科技股份有限公司 | Ship starboard height identification method and system |
| CN117272086A (en) * | 2023-11-22 | 2023-12-22 | 中国电子科技集团公司第二十九研究所 | Radar signal scanning envelope segmentation method based on DBSCAN |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355194A (en) * | 2016-08-22 | 2017-01-25 | 广东华中科技大学工业技术研究院 | Treatment method for surface target of unmanned ship based on laser imaging radar |
| CN110515091A (en) * | 2019-08-28 | 2019-11-29 | 珠海达伽马科技有限公司 | A kind of water-surface areas detection method and water-surface areas detection device, computer readable storage medium for unmanned boat automatic Pilot |
| KR20200104436A (en) * | 2019-02-26 | 2020-09-04 | 경기과학기술대학교 산학협력단 | System for colleting marine waste |
| CN113343738A (en) * | 2020-02-17 | 2021-09-03 | 华为技术有限公司 | Detection method, device and storage medium |
-
2021
- 2021-09-26 CN CN202111125216.7A patent/CN113570005A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355194A (en) * | 2016-08-22 | 2017-01-25 | 广东华中科技大学工业技术研究院 | Treatment method for surface target of unmanned ship based on laser imaging radar |
| KR20200104436A (en) * | 2019-02-26 | 2020-09-04 | 경기과학기술대학교 산학협력단 | System for colleting marine waste |
| CN110515091A (en) * | 2019-08-28 | 2019-11-29 | 珠海达伽马科技有限公司 | A kind of water-surface areas detection method and water-surface areas detection device, computer readable storage medium for unmanned boat automatic Pilot |
| CN113343738A (en) * | 2020-02-17 | 2021-09-03 | 华为技术有限公司 | Detection method, device and storage medium |
Non-Patent Citations (6)
| Title |
|---|
| 丁宇星等: "基于激光二极管的光子计数激光测距技术", 《科学技术与工程》 * |
| 张怡等: "基于三维激光扫描技术的无图纸渔船船体型线获取方法研究与应用", 《河北渔业》 * |
| 方剑等: "一种基于光子计数激光雷达的去噪方法", 《舰船电子对抗》 * |
| 李铭等: "一种推扫式光子计数激光雷达点云滤波算法及其验证", 《科学技术与工程》 * |
| 蔡玉良等: "面向船舶检验的无人机关键技术解决方案", 《船海工程》 * |
| 陈鹏等: "《合成孔径雷达海上舰船遥感探测技术与应用》", 31 January 2012 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114782868A (en) * | 2022-04-19 | 2022-07-22 | 网新百橙科技(杭州)有限公司 | Video image ship identification system and method based on AI artificial intelligence |
| CN116148878A (en) * | 2023-04-18 | 2023-05-23 | 浙江华是科技股份有限公司 | Ship starboard height identification method and system |
| CN117272086A (en) * | 2023-11-22 | 2023-12-22 | 中国电子科技集团公司第二十九研究所 | Radar signal scanning envelope segmentation method based on DBSCAN |
| CN117272086B (en) * | 2023-11-22 | 2024-02-13 | 中国电子科技集团公司第二十九研究所 | Radar signal scanning envelope segmentation method based on DBSCAN |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113570005A (en) | Long-distance ship type identification method based on airborne photon radar | |
| CN105488770B (en) | A kind of airborne laser radar point cloud filtering method of object-oriented | |
| CN106599808B (en) | Hidden target extraction method based on full-waveform laser radar data | |
| CN109871902B (en) | SAR small sample identification method based on super-resolution countermeasure generation cascade network | |
| CN110163108B (en) | Robust sonar target detection method based on dual-path feature fusion network | |
| CN111444767B (en) | Pedestrian detection and tracking method based on laser radar | |
| CN111524084B (en) | Photon counting laser radar point cloud denoising method based on multi-peak Gaussian fitting | |
| US10497128B2 (en) | Method and system for sea background modeling and suppression on high-resolution remote sensing sea images | |
| CN112859011B (en) | Method for extracting waveform signals of single-wavelength airborne sounding radar | |
| CN112464994A (en) | Boat stern wave identification and removal method based on PointNet network | |
| CN110443201B (en) | Target identification method based on multi-source image joint shape analysis and multi-attribute fusion | |
| CN108830224A (en) | A kind of high-resolution remote sensing image Ship Target Detection method based on deep learning | |
| CN111444769A (en) | Laser radar human leg detection method based on multi-scale self-adaptive random forest | |
| Zhang et al. | Nearshore vessel detection based on Scene-mask R-CNN in remote sensing image | |
| CN116027349A (en) | Coral reef substrate classification method based on laser radar and side scan sonar data fusion | |
| CN115390040A (en) | Tree point cloud branch and leaf separation method based on segmentation geometric features | |
| CN105701856B (en) | A kind of vegetation extracting method and system | |
| CN109597044B (en) | Broadband polarization radar seeker target identification method based on hierarchical decision tree | |
| CN111796250A (en) | False trace point multi-dimensional hierarchical suppression method based on risk assessment | |
| CN111709986A (en) | Power transmission line forest tree statistical method based on laser point cloud | |
| CN115810144A (en) | Underwater suspended sonar target identification method based on area pre-detection | |
| CN115267827A (en) | Laser radar harbor area obstacle sensing method based on height density screening | |
| CN111507423B (en) | Engineering quantity measuring method for cleaning transmission line channel | |
| Kong et al. | Automatic detection technology of sonar image target based on the three-dimensional imaging | |
| Liang et al. | MVCNN: A Deep Learning-Based Ocean–Land Waveform Classification Network for Single-Wavelength LiDAR Bathymetry |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20211029 |