CN112464994B - PointNet network-based boat tail wave recognition and removal method - Google Patents
PointNet network-based boat tail wave recognition and removal method Download PDFInfo
- Publication number
- CN112464994B CN112464994B CN202011225901.2A CN202011225901A CN112464994B CN 112464994 B CN112464994 B CN 112464994B CN 202011225901 A CN202011225901 A CN 202011225901A CN 112464994 B CN112464994 B CN 112464994B
- Authority
- CN
- China
- Prior art keywords
- point cloud
- target
- point
- tail
- pointnet
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 33
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims abstract description 46
- 238000013136 deep learning model Methods 0.000 claims abstract description 33
- 238000001514 detection method Methods 0.000 claims abstract description 31
- 238000001914 filtration Methods 0.000 claims description 9
- 238000004347 surface barrier Methods 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 description 3
- 230000011218 segmentation Effects 0.000 description 3
- 230000006978 adaptation Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002779 inactivation Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000007493 shaping process Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Optical Radar Systems And Details Thereof (AREA)
Abstract
The method for identifying and removing the stern waves of the ships and the boats based on the PointNet network comprises the following steps: s1: acquiring point cloud data of the unmanned ship in 360 degrees in real time by utilizing a laser radar; s2: denoising the point cloud data; s3: performing Euclidean distance clustering on the point cloud data subjected to denoising processing to obtain a suspicious target list; s4: constructing a PointNet deep learning model and loading trained weight parameters; s5: inputting the clustered point cloud subsets into a PointNet deep learning model for classification and identification; s6: removing the point cloud subset identified as the tail wave target from the suspicious target list; s7: judging whether each tail wave target is the tail wave generated by the unmanned ship or the tail wave generated by other ships, and positioning the ship target on the water surface according to the tail wave position under the condition of not self tail waves; s8: and sending the detection result of the water surface obstacle to the unmanned ship control platform. According to the invention, the PointNet deep learning model is utilized to realize rapid recognition and removal of the stern waves of the boat, and the target boat is positioned based on the stern wave recognition result, so that the adaptability of the laser radar to the application on the water surface is improved.
Description
Technical Field
The invention relates to a ship stern wave recognition and removal method based on a PointNet network, and belongs to the technical field of environmental awareness.
Background
Lidar is one of the most important and indispensable sensors in the implementation of automatic driving, and its importance is self-evident, such as obstacle detection, road edge detection, map construction, etc. is not separated. The stern wave of the boat can be detected by the laser radar by reflecting the detection laser, but the stern wave does not belong to a water surface obstacle and does not influence the navigation route of the boat, so that the stern wave needs to be filtered from the targets detected by the laser radar. The tail waves generated by the unmanned ship can be determined to be positioned behind the ship, but the tail waves also change in real time along with the change of the navigation speed of the ship. The tail waves generated by the unmanned ship can be removed by filtering the point cloud in a certain range behind the ship, but the filtering range needs to be set to be far larger than the fixed range of the actual tail waves to ensure the accurate filtering of the tail waves, so that the water surface target detection capability of the unmanned ship under the close range condition is greatly reduced. The tail waves generated by the water surface target boat can improve the detection probability and the detection distance of the target, but also can improve the positioning error of the water surface target boat, thereby affecting the accurate planning of the obstacle avoidance path of the unmanned boat. Unmanned ship is urgent to obtain the detection and recognition capability of the navigation tail wave of the ship based on the laser radar, and the tail wave is accurately filtered, so that the close-range detection capability of the laser radar equipment in the water surface environment and the positioning capability of a long-range target are improved.
Disclosure of Invention
The technical solution of the invention is as follows: overcomes the defects of the prior art and provides a method for identifying and removing the stern waves of a boat based on a PointNet network.
The technical scheme of the invention is as follows:
the method for identifying and removing the stern waves of the ships and the boats based on the PointNet network comprises the following steps:
s1: the laser radar equipment is arranged at the non-shielding position at the top of the unmanned ship, and the point cloud data of the unmanned ship in the 360-degree range is acquired in real time;
s2: denoising processing based on point cloud density statistics is carried out on the current point cloud data, and discrete noise points in the point cloud data are removed;
s3: performing Euclidean distance clustering on point cloud data with noise points removed according to a pre-designed tolerance distance, dividing the point cloud data into a plurality of independent point cloud subsets, wherein each point cloud subset is a suspicious target, and counting and determining target point clouds, boundary ranges, target center points and target radiuses of each suspicious target to obtain a suspicious target list;
s4: constructing a PointNet deep learning model and loading trained weight parameters;
s5: in the S3, searching point cloud subsets with point cloud points and target radius larger than the preset data, and inputting the found point cloud subsets into a PointNet deep learning model for classification and identification;
s6: if a certain point cloud subset is identified as a tail wave target through the PointNet deep learning model, removing the point cloud subset from a suspicious target list;
s7: judging whether each tail wave target is the tail wave generated by the unmanned ship or the tail wave generated by other ships, if the tail wave generated by other ships, determining the water surface navigation ship target corresponding to the tail wave target, and forming a water surface barrier together with other suspicious targets;
s8: according to the position change of the water surface obstacle in the continuous frames, the current position, the movement speed and the movement direction of the water surface obstacle are calculated, the Kalman filtering is utilized to optimize detection information and predict the movement trend of the obstacle, and finally the information is used as a detection result to be sent to the unmanned ship control platform to assist in realizing the automatic driving and the autonomous obstacle avoidance of the unmanned ship.
In the step S2, the step of denoising the current point cloud data is as follows:
(2.1) setting the number k of neighborhood points and a standard deviation multiple threshold std_mul;
(2.2) calculating the distance average value d of each point in the current point cloud data and k field points to form a distance vector;
(2.3) estimating a distance mean mu and a standard deviation sigma in the whole point cloud data according to the distance vector, wherein the distance vector accords with Gaussian distribution;
(2.4) calculating a distance threshold value with a calculation formula of t=μ+σx std_mul; the method comprises the steps of carrying out a first treatment on the surface of the
(2.5) if the distance average d is greater than the distance threshold t, the laser detection point is an outlier and is filtered from the current point cloud data.
In the step S3, the tolerance distance is a distance threshold of the point cloud distance cluster, and if the distance between two points is smaller than the tolerance distance, the two points are considered to belong to one target, otherwise, the two points are considered to not belong to one target.
The implementation manner of the step S7 is as follows:
for each point cloud subset identified as a tail wave target, the following operations are performed:
according to the point Yun Ziji, performing shape fitting on a tail wave fan of a tail wave target, wherein the top of the fan after shape fitting is the position of the target boat which causes the tail wave, and if the position is coincident with the position of the unmanned boat, the tail wave is the tail wave generated by the unmanned boat; otherwise, the tail waves generated by other boats are used as the tail waves, the fan-shaped top is a water surface navigation boat target, and the tail waves and other suspicious targets form a water surface barrier together.
The method for performing shape fitting on the tail wave sector of the tail wave target according to the point Yun Ziji is as follows:
(5.1) randomly selecting three points A, B, C from the tail wave point cloud subset;
(5.2) taking B, C as a fixed point, searching a point farthest from B, C points from the tail wave point cloud subset, and updating the position of the point A by using the point;
(5.3) taking C, A as a fixed point, searching a point farthest from C, A points from the tail wave point cloud subset, and updating the position of the point B by using the point;
(5.4) taking A, B as a fixed point, searching a point farthest from A, B points from the tail wave point cloud subset, and updating the position of the point C by using the point;
(5.5) repeating the steps (5.2) - (5.4) until the positions of the three points A, B, C are not changed any more, and entering the step (5.6);
(5.6) A, B, C, calculating angles of three vertex angles A, B and C of the triangle, selecting a vertex corresponding to the minimum angle as a circle center, setting 1.2 times of the minimum angle as a fan-shaped angle, setting 1.2 times of the maximum distance from the vertex corresponding to the minimum angle to other two points as a radius, and constructing a fan-shaped shape, wherein the fan-shaped shape covers all tail wave point clouds.
The training process of the PointNet deep learning model is as follows:
(6.1) acquiring point cloud data in a 360-degree range of the unmanned ship by using laser radar equipment;
(6.2) manually marking the unmanned ship point cloud data with tail wave point cloud, wherein the tail wave target is marked as 1, and the non-tail wave target is marked as 0;
(6.3) combining unmanned ship point cloud data and a marking result into PointNet deep learning model training data;
(6.4) denoising processing based on point cloud density statistics is carried out on the training data;
(6.5) carrying out Euclidean distance clustering on the training data processed in the step (6.4) according to the pre-designed tolerance distance, and dividing the training data into a plurality of independent point cloud subsets;
(6.6) using the training data to accelerate training of the PointNet deep learning model on the multiple GPU servers.
Compared with the prior art, the invention has the following beneficial effects:
(1) According to the invention, the tail wave point cloud distribution in the point cloud data is learned and identified through the PointNet deep learning model, so that the tail wave is accurately removed, the tail wave presents a fan-shaped plane distribution and is easy to identify, and the tail wave shape analysis result can also assist in positioning a water surface target boat, so that the detection capability of the laser radar is improved. The PointNet deep learning model directly utilizes point cloud data to carry out learning training, can greatly improve the recognition speed of tail waves, and is matched with the rapid scanning capability of a laser radar. The tail waves of the unmanned ship are accurately identified and removed, interference of the tail waves to the close-range water surface obstacle can be eliminated, and the close-range water surface target detection discovery capability and the adaptation capability of the unmanned ship are effectively improved.
(2) According to the method, the original laser radar data is optimized by adopting a denoising algorithm based on the point cloud density statistics, and discrete noise points of the laser radar point cloud data are filtered, so that the accuracy of a point cloud segmentation method based on distance clustering is effectively ensured.
(3) According to the method, the laser radar point cloud is rapidly segmented into mutually independent targets by utilizing a distance clustering method, and the segmentation result is transmitted to a PointNet deep learning model for classification and identification. Because of the mirror reflection effect, the calm water surface is not imaged in the laser radar data, the laser radar water surface point cloud data only comprises targets such as water surface obstacles, the laser point number is usually kept at a lower level, the distance clustering segmentation method of the point cloud can achieve extremely high speed, and the real-time performance of the method for processing the laser radar data is ensured by combining with the PointNet deep learning model.
Drawings
FIG. 1 is a flow chart of the PointNet model training data creation provided by the present invention;
FIG. 2 is a flowchart of the PointNet model training provided by the present invention;
FIG. 3 is a flow chart of tail wave detection and identification based on PointNet provided by the invention.
FIG. 4 is a diagram showing the network configuration parameters of the deep learning PointNet model according to the present invention.
Detailed Description
The technical scheme of the invention is further described in detail through the attached drawings.
The invention aims to automatically detect and identify a water surface area and an unknown water surface obstacle based on a visible light camera, and is applicable to pixel-level detection distributed in the water surface area and automatic detection of the unknown water surface obstacle. Therefore, the invention designs a deep learning PointNet model, utilizes a laser radar to detect point cloud data around the unmanned ship, utilizes statistical noise filtering to remove discrete noise in the point cloud data, utilizes a distance clustering-based method to divide the point cloud into mutually independent targets, trains the PointNet deep learning model to learn and identify tail wave point clouds generated by navigation of the ship in the point cloud data, enables the tail wave to be in fan-shaped plane distribution and easy to identify, and can filter the tail wave point clouds in the point cloud data according to the identification result. And fitting the shape of the tail wave fan according to the identification result, wherein the top of the fan is a target boat position for causing the tail wave, the tail wave generated by the unmanned boat is obtained by overlapping the position of the tail wave fan and the unmanned boat, otherwise, the tail wave generated by other boats is obtained by combining the position of the tail wave fan with other water surface targets to form a water surface barrier. The aggregation result of the water surface obstacles detected by the laser radar can guide the unmanned ship to navigate and avoid the obstacle autonomously.
As shown in FIG. 3, the method for identifying and removing the stern waves of the ships based on the PointNet network comprises the following steps:
s1: and arranging the laser radar equipment at a non-shielding position at the top of the unmanned ship, and acquiring the point cloud data of the unmanned ship within 360 degrees in real time.
S2: and denoising the current point cloud data based on the point cloud density statistics to remove discrete noise points in the point cloud data.
The denoising process comprises the following steps:
(2.1) setting the number k of neighborhood points and a standard deviation multiple threshold std_mul;
(2.2) calculating the distance average value d of each point in the current point cloud data and k field points to form a distance vector;
(2.3) estimating a distance mean mu and a standard deviation sigma in the whole point cloud data according to the distance vector, wherein the distance vector accords with Gaussian distribution;
(2.4) calculating a distance threshold value with a calculation formula of t=μ+σx std_mul;
(2.5) if the distance average d is greater than the distance threshold t, the laser detection point is an outlier and is filtered from the current point cloud data.
S3: and (3) carrying out Euclidean distance clustering on the point cloud data with noise points removed according to a pre-designed tolerance distance, dividing the point cloud data into a plurality of independent point cloud subsets, wherein each point cloud subset is a suspicious target, and counting and determining the target point cloud, the boundary range, the target center point and the target radius of each suspicious target to obtain a suspicious target list.
The tolerance distance is a distance threshold value of the point cloud distance cluster, if the distance between two points is smaller than the tolerance distance, the two points are considered to belong to one target, otherwise, the two points are considered to not belong to one target.
S4: and constructing a PointNet deep learning model and loading trained weight parameters.
S5: and in S3, searching for a point cloud subset with the point cloud point number and the target radius larger than the pre-specified data in each point cloud subset, and inputting the found point cloud subset into the PointNet deep learning model for classification and identification.
S6: if a certain point cloud subset is identified as a tail wave target through the PointNet deep learning model, the point cloud subset is removed from the suspicious target list.
S7: for each point cloud subset identified as a tail wave target, the following operations are performed:
according to the point Yun Ziji, performing shape fitting on a tail wave fan of a tail wave target, wherein the top of the fan after shape fitting is the position of the target boat which causes the tail wave, and if the position is coincident with the position of the unmanned boat, the tail wave is the tail wave generated by the unmanned boat; otherwise, the tail waves generated by other boats are used as the tail waves, the fan-shaped top is a water surface navigation boat target, and the tail waves and other suspicious targets form a water surface barrier together.
The method for performing shape fitting on the tail wave sector of the tail wave target according to the point Yun Ziji is as follows:
(5.1) randomly selecting three points A, B, C from the tail wave point cloud subset;
(5.2) taking B, C as a fixed point, searching a point farthest from B, C points from the tail wave point cloud subset, and updating the position of the point A by using the point;
(5.3) taking C, A as a fixed point, searching a point farthest from C, A points from the tail wave point cloud subset, and updating the position of the point B by using the point;
(5.4) taking A, B as a fixed point, searching a point farthest from A, B points from the tail wave point cloud subset, and updating the position of the point C by using the point;
(5.5) repeating the steps (5.2) - (5.4) until the positions of the three points A, B, C are not changed any more, and entering the step (5.6);
and (5.6) calculating angles of three vertex angles A, B and C of the triangle ABC, selecting a vertex corresponding to the minimum angle as a circle center, wherein 1.2 times of the minimum angle is a fan-shaped angle, 1.2 times of the maximum distance from the vertex corresponding to the minimum angle to other two points is a radius, and constructing a fan-shaped shape, wherein the fan-shaped shape covers all tail wave point clouds.
S8: according to the position change of the water surface obstacle in the continuous frames, the current position, the movement speed and the movement direction of the water surface obstacle are calculated, the Kalman filtering is utilized to optimize detection information and predict the movement trend of the obstacle, and finally the information is used as a detection result to be sent to the unmanned ship control platform to assist in realizing the automatic driving and the autonomous obstacle avoidance of the unmanned ship.
A PointNet deep learning model is built on a deep learning platform (TensorFlow or Pytorch), and the PointNet model can realize classification domain identification of three-dimensional point clouds, and detailed design parameters of the model are shown in fig. 4. In fig. 4, inputLayer represents an input layer, conv-bn-relu represents a convolution module, maxPooling represents a max pooling layer, dense-bn-relu represents a fully connected layer, reshape represents a data shaping layer, dropout represents a random inactivation layer, and softmax represents a classification layer.
As shown in fig. 1, training data of the PointNet deep learning model is obtained as follows:
(1) The laser radar is arranged at the high position of the mast of the unmanned ship, the situation that the equipment work is influenced by shielding of 360 degrees around the laser radar is guaranteed, and the point cloud data of the unmanned ship in the 360-degree range are acquired in real time according to a laser radar equipment protocol.
(2) Manually marking the unmanned ship point cloud data with tail wave point cloud, wherein the tail wave target is marked as 1, and the non-tail wave target is marked as 0;
(3) And forming the unmanned ship point cloud data and the marking result into point cloud data trained by a PointNet deep learning model, and storing the point cloud data to a specified path.
As shown in fig. 2, the training method of the PointNet deep learning model is as follows:
and automatically scanning all training point cloud data from the point cloud folder, randomly selecting the training point cloud data to generate model training data, wherein the model training data comprises label data corresponding to the model training data.
(1) Denoising the training data based on the point cloud density statistics;
(2) Carrying out Euclidean distance clustering on the training data after denoising treatment according to a pre-designed tolerance distance, dividing the training data into a plurality of independent subsets, wherein each subset is a suspicious target, and statistically determining a target point cloud, a boundary range, a target center point and a target radius of each suspicious target to obtain a suspicious target list;
(3) Utilizing a multi-GPU server to accelerate training of the PointNet deep learning model;
(4) And storing the trained PointNet deep learning model network structure and weight parameters, and copying the network structure and the weight parameters to an unmanned ship-borne computer for identifying tail waves and surface ships in the unmanned ship navigation process.
The invention can be executed by a shipborne environment sensing system based on laser radar equipment, and can realize rapid detection and identification of self tail waves of unmanned ships and tail waves of other ships.
Examples:
s310, installing the laser radar at the high position of the mast of the unmanned ship, and enabling the laser radar to be free of shielding within the range of 360 degrees.
And S320, the shipborne computer runs a laser radar driving program to communicate with laser radar equipment, and three-dimensional point cloud data of the unmanned ship in a 360-degree range are collected in real time.
S330, denoising processing based on point cloud density statistics is carried out on the current point cloud data, and discrete noise points in the laser radar point cloud data are removed.
S340, performing Euclidean distance clustering on the laser radar point cloud with noise removed, dividing the point cloud into a plurality of independent point cloud handwriting as suspicious targets, and further performing statistical analysis on information such as target point cloud, boundary range, target center point, target radius and the like of the suspicious targets.
S350, the shipborne computer automatically loads the trained PointNet deep learning model and weight parameters, sets the training model as an inference mode, searches for a point cloud subset with the point cloud number and the target radius larger than the pre-specified data in each point cloud subset, and inputs the found point cloud subset into the PointNet deep learning model for classification and identification; and inputting the found point cloud subset, and outputting class labels of suspicious targets on the water surface.
S360, if a certain point cloud subset is identified as a tail wave target through the PointNet deep learning model, the point cloud subset is removed from the suspicious target list.
S370, according to the point Yun Ziji, fitting the shape of the tail wave fan of the tail wave target, wherein the top of the fan is the target boat position for causing the tail wave, the position is overlapped with the position of the unmanned boat, namely the tail wave generated by the unmanned boat, otherwise, the position is the tail wave generated by other boats, and the position is a water surface navigation boat target and forms a water surface barrier together with other water surface targets.
And S380, calculating the current position, the movement speed and the movement direction of the water surface target according to the position change of the water surface target in the continuous frames, optimizing detection information and predicting the movement trend of the target by utilizing Kalman filtering, and finally counting the detection result and sending the detection result to the unmanned ship for autonomous navigation and autonomous obstacle avoidance.
According to the method, the PointNet deep learning model is utilized to realize rapid recognition and removal of the stern waves of the ship, and the target ship is positioned based on the stern wave recognition result, so that the adaptability of the laser radar equipment to the water surface application is improved.
The method can also realize the detection and identification of the non-tail wave targets on the water surface, but is limited by the lower detection resolution of the existing laser radar equipment, the laser radar can only identify the targets with larger short distance and larger size, and the targets need to acquire enough points to reflect the shape of the targets. Therefore, the method only identifies and filters tail wave targets at present, and the detection identification range of the method can be expanded to various water surface targets along with the improvement of the detection resolution of the laser radar in the future.
What is not described in detail in the present specification belongs to the known technology of those skilled in the art.
Claims (6)
1. The method for identifying and removing the stern waves of the ships and the boats based on the PointNet network is characterized by comprising the following steps of:
s1: the laser radar equipment is arranged at the non-shielding position at the top of the unmanned ship, and the point cloud data of the unmanned ship in the 360-degree range is acquired in real time;
s2: denoising processing based on point cloud density statistics is carried out on the current point cloud data, and discrete noise points in the point cloud data are removed;
s3: performing Euclidean distance clustering on point cloud data with noise points removed according to a pre-designed tolerance distance, dividing the point cloud data into a plurality of independent point cloud subsets, wherein each point cloud subset is a suspicious target, and counting and determining target point clouds, boundary ranges, target center points and target radiuses of each suspicious target to obtain a suspicious target list;
s4: constructing a PointNet deep learning model and loading trained weight parameters;
s5: in the S3, searching point cloud subsets with point cloud points and target radius larger than the preset data, and inputting the found point cloud subsets into a PointNet deep learning model for classification and identification;
s6: if a certain point cloud subset is identified as a tail wave target through the PointNet deep learning model, removing the point cloud subset from a suspicious target list;
s7: judging whether each tail wave target is the tail wave generated by the unmanned ship or the tail wave generated by other ships, if the tail wave generated by other ships, determining the water surface navigation ship target corresponding to the tail wave target, and forming a water surface barrier together with other suspicious targets;
s8: and calculating the current position, the movement speed and the movement direction of the water surface obstacle according to the position change of the water surface obstacle in the continuous frames, optimizing detection information and predicting the movement trend of the obstacle by utilizing Kalman filtering, and finally sending the obstacle information and the detection information to an unmanned ship control platform as detection results to assist in realizing the automatic driving and the autonomous obstacle avoidance of the unmanned ship.
2. The method for identifying and removing the stern waves of the ship based on the PointNet network according to claim 1, wherein in the step S2, the step of denoising the current point cloud data is as follows:
(2.1) setting the number k of neighborhood points and a standard deviation multiple threshold std_mul;
(2.2) calculating the distance average value d of each point in the current point cloud data and k field points to form a distance vector;
(2.3) estimating a distance mean mu and a standard deviation sigma in the whole point cloud data according to the distance vector, wherein the distance vector accords with Gaussian distribution;
(2.4) calculating a distance threshold value with a calculation formula of t=μ+σx std_mul;
and (2.5) if the distance average value d is greater than the distance threshold t, the laser detection point in the point cloud data corresponding to the distance average value d is an outlier point, and filtering out from the current point cloud data.
3. The method for identifying and removing stern waves of a ship based on a PointNet network according to claim 1, wherein in the step S3, the tolerance distance is a distance threshold of a point cloud distance cluster, and two points are considered to belong to one target if the distance between the two points is smaller than the tolerance distance, and are not considered to belong to one target if the distance between the two points is smaller than the tolerance distance.
4. The method for identifying and removing the stern waves of the ship based on the PointNet network according to claim 1, wherein the implementation manner of the step S7 is as follows:
for each point cloud subset identified as a tail wave target, the following operations are performed:
according to the point Yun Ziji, performing shape fitting on a tail wave fan of a tail wave target, wherein the top of the fan after shape fitting is the position of the target boat which causes the tail wave, and if the position is coincident with the position of the unmanned boat, the tail wave is the tail wave generated by the unmanned boat; otherwise, the tail waves generated by other boats are used as the tail waves, the fan-shaped top is a water surface navigation boat target, and the tail waves and other suspicious targets form a water surface barrier together.
5. The method for identifying and removing stern waves of a ship based on a PointNet network as set forth in claim 4, wherein the method for fitting the shape of the stern wave sector of the stern wave object according to the point Yun Ziji is as follows:
(5.1) randomly selecting three points A, B, C from the tail wave point cloud subset;
(5.2) taking B, C as a fixed point, searching a point farthest from B, C points from the tail wave point cloud subset, and updating the position of the point A by using the point;
(5.3) taking C, A as a fixed point, searching a point farthest from C, A points from the tail wave point cloud subset, and updating the position of the point B by using the point;
(5.4) taking A, B as a fixed point, searching a point farthest from A, B points from the tail wave point cloud subset, and updating the position of the point C by using the point;
(5.5) repeating the steps (5.2) - (5.4) until the positions of the three points A, B, C are not changed any more, and entering the step (5.6);
(5.6) A, B, C, calculating angles of three vertex angles A, B and C of the triangle, selecting a vertex corresponding to the minimum angle as a circle center, setting 1.2 times of the minimum angle as a fan-shaped angle, setting 1.2 times of the maximum distance from the vertex corresponding to the minimum angle to other two points as a radius, and constructing a fan-shaped shape, wherein the fan-shaped shape covers all tail wave point clouds.
6. The method for identifying and removing the stern waves of the ship based on the PointNet network according to claim 1, wherein the training process of the PointNet deep learning model is as follows:
(6.1) acquiring point cloud data in a 360-degree range of the unmanned ship by using laser radar equipment;
(6.2) manually marking the unmanned ship point cloud data with tail wave point cloud, wherein the tail wave target is marked as 1, and the non-tail wave target is marked as 0;
(6.3) combining unmanned ship point cloud data and a marking result into PointNet deep learning model training data;
(6.4) denoising processing based on point cloud density statistics is carried out on the training data;
(6.5) carrying out Euclidean distance clustering on the training data processed in the step (6.4) according to the pre-designed tolerance distance, and dividing the training data into a plurality of independent point cloud subsets;
(6.6) using the training data to accelerate training of the PointNet deep learning model on the multiple GPU servers.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011225901.2A CN112464994B (en) | 2020-11-05 | 2020-11-05 | PointNet network-based boat tail wave recognition and removal method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011225901.2A CN112464994B (en) | 2020-11-05 | 2020-11-05 | PointNet network-based boat tail wave recognition and removal method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112464994A CN112464994A (en) | 2021-03-09 |
| CN112464994B true CN112464994B (en) | 2024-03-26 |
Family
ID=74825029
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011225901.2A Active CN112464994B (en) | 2020-11-05 | 2020-11-05 | PointNet network-based boat tail wave recognition and removal method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112464994B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113191973B (en) * | 2021-04-29 | 2023-09-01 | 西北大学 | Cultural relic point cloud data denoising method based on unsupervised network framework |
| CN113408401B (en) * | 2021-06-16 | 2022-02-22 | 中国科学院南海海洋研究所 | Method and device for quickly and automatically identifying ship traveling wave based on machine learning |
| CN113389186B (en) * | 2021-07-07 | 2022-05-17 | 安徽工程大学 | Multi-unmanned-boat-cooperated sea surface oil stain cleaning method and system |
| CN113763263A (en) * | 2021-07-27 | 2021-12-07 | 华能伊敏煤电有限责任公司 | Water mist tail gas noise treatment method based on point cloud tail gas filtering technology |
| CN114384930A (en) * | 2021-12-21 | 2022-04-22 | 中国电子科技集团公司第五十四研究所 | Group intelligent decision making system and method based on global information sharing |
| CN114332647B (en) * | 2021-12-31 | 2022-10-21 | 合肥工业大学 | River channel boundary detection and tracking method and system for unmanned ship |
| CN115327568B (en) * | 2022-07-19 | 2023-10-20 | 哈尔滨工程大学 | PointNet network-based unmanned aerial vehicle cluster real-time target recognition method, system and map construction method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108152831A (en) * | 2017-12-06 | 2018-06-12 | 中国农业大学 | A kind of laser radar obstacle recognition method and system |
| KR101888170B1 (en) * | 2017-11-16 | 2018-08-13 | 엘아이지넥스원 주식회사 | Method and device for deleting noise in detecting obstacle by unmanned surface vessel |
| CN109283538A (en) * | 2018-07-13 | 2019-01-29 | 上海大学 | A kind of naval target size detection method of view-based access control model and laser sensor data fusion |
| CN110275153A (en) * | 2019-07-05 | 2019-09-24 | 上海大学 | A kind of waterborne target detection and tracking based on laser radar |
| CN110501719A (en) * | 2019-08-27 | 2019-11-26 | 王玉娇 | A kind of train obstacle detection method based on laser radar |
| CN111239766A (en) * | 2019-12-27 | 2020-06-05 | 北京航天控制仪器研究所 | Water surface multi-target rapid identification and tracking method based on laser radar |
| CN111680542A (en) * | 2020-04-17 | 2020-09-18 | 东南大学 | Steel coil point cloud identification and classification method based on multi-scale feature extraction and Pointernet neural network |
-
2020
- 2020-11-05 CN CN202011225901.2A patent/CN112464994B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101888170B1 (en) * | 2017-11-16 | 2018-08-13 | 엘아이지넥스원 주식회사 | Method and device for deleting noise in detecting obstacle by unmanned surface vessel |
| CN108152831A (en) * | 2017-12-06 | 2018-06-12 | 中国农业大学 | A kind of laser radar obstacle recognition method and system |
| CN109283538A (en) * | 2018-07-13 | 2019-01-29 | 上海大学 | A kind of naval target size detection method of view-based access control model and laser sensor data fusion |
| CN110275153A (en) * | 2019-07-05 | 2019-09-24 | 上海大学 | A kind of waterborne target detection and tracking based on laser radar |
| CN110501719A (en) * | 2019-08-27 | 2019-11-26 | 王玉娇 | A kind of train obstacle detection method based on laser radar |
| CN111239766A (en) * | 2019-12-27 | 2020-06-05 | 北京航天控制仪器研究所 | Water surface multi-target rapid identification and tracking method based on laser radar |
| CN111680542A (en) * | 2020-04-17 | 2020-09-18 | 东南大学 | Steel coil point cloud identification and classification method based on multi-scale feature extraction and Pointernet neural network |
Non-Patent Citations (1)
| Title |
|---|
| 吕德华.《 基于光学测量技术的水下船只尾流气泡幕探测技术研究》.《 机械设计与制造工程》.2020,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112464994A (en) | 2021-03-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112464994B (en) | PointNet network-based boat tail wave recognition and removal method | |
| Danzer et al. | 2d car detection in radar data with pointnets | |
| CN111239766B (en) | Water surface multi-target rapid identification and tracking method based on laser radar | |
| WO2020192431A1 (en) | System and method for ordered representation and feature extraction for point clouds obtained by detection and ranging sensor | |
| CN110275153B (en) | Water surface target detection and tracking method based on laser radar | |
| Haghbayan et al. | An efficient multi-sensor fusion approach for object detection in maritime environments | |
| CN109283538A (en) | A kind of naval target size detection method of view-based access control model and laser sensor data fusion | |
| CN112347999A (en) | Obstacle recognition model training method, obstacle recognition method, device and system | |
| CN112613504A (en) | Sonar underwater target detection method | |
| CN115861968A (en) | Dynamic obstacle removing method based on real-time point cloud data | |
| Wang et al. | An autonomous cooperative system of multi-AUV for underwater targets detection and localization | |
| CN111736167B (en) | Method and device for obtaining laser point cloud density | |
| US20220309794A1 (en) | Methods and electronic devices for detecting objects in surroundings of a self-driving car | |
| CN116503803A (en) | Obstacle detection method, obstacle detection device, electronic device and storage medium | |
| CN115100741A (en) | Point cloud pedestrian distance risk detection method, system, equipment and medium | |
| Thompson | Maritime object detection, tracking, and classification using lidar and vision-based sensor fusion | |
| CN114241448A (en) | Method and device for obtaining heading angle of obstacle, electronic equipment and vehicle | |
| CN113611008B (en) | Vehicle driving scene acquisition method, device, equipment and medium | |
| Li et al. | Underwater robot visual place recognition in the presence of dramatic appearance change | |
| CN115421160A (en) | Road edge detection method, device, equipment, vehicle and storage medium | |
| She et al. | Multi-obstacle detection based on monocular vision for UAV | |
| CN113077520A (en) | Collision prediction method and device and edge calculation server | |
| CN110895680A (en) | Unmanned ship water surface target detection method based on regional suggestion network | |
| CN112882058A (en) | Shipborne laser radar obstacle detection method based on variable-size grid map | |
| Wan et al. | Online Obstacle Detection for USV based on Improved RANSAC Algorithm |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20220408 Address after: 266200 aoshanwei sub district office, Jimo District, Qingdao City, Shandong Province Applicant after: Aerospace Times (Qingdao) marine equipment technology development Co.,Ltd. Address before: 142 box 403, box 100854, Beijing, Beijing, Haidian District Applicant before: BEIJIGN INSTITUTE OF AEROSPACE CONTROL DEVICES |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |