CN115906980B - Pedestrian detection method - Google Patents

Pedestrian detection method Download PDF

Info

Publication number
CN115906980B
CN115906980B CN202211412927.7A CN202211412927A CN115906980B CN 115906980 B CN115906980 B CN 115906980B CN 202211412927 A CN202211412927 A CN 202211412927A CN 115906980 B CN115906980 B CN 115906980B
Authority
CN
China
Prior art keywords
neural network
gat
graph neural
node
nodes
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
Application number
CN202211412927.7A
Other languages
Chinese (zh)
Other versions
CN115906980A (en
Inventor
鲁鸣鸣
郭清明
常佳宇
欧阳凯
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Central South University
Original Assignee
Central South University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Central South University filed Critical Central South University
Priority to CN202211412927.7A priority Critical patent/CN115906980B/en
Publication of CN115906980A publication Critical patent/CN115906980A/en
Application granted granted Critical
Publication of CN115906980B publication Critical patent/CN115906980B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Traffic Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a pedestrian detection method, which comprises the steps of determining an initial GAT graph neural network model adopted by pedestrian detection in an automatic driving process; adopting a determined initial GAT graph neural network model, and adopting a construction method to construct and obtain a GAT graph neural network; and adopting the obtained GAT graph neural network to detect pedestrians in the automatic driving process. The invention not only realizes the defending of the GAT graph neural network through an innovative robust message transmission mechanism, but also increases the robustness of the model greatly, and has good defending effect, high reliability and good stability.

Description

Pedestrian detection method
Technical Field
The invention belongs to the technical field of artificial intelligence, and particularly relates to a pedestrian detection method.
Background
Along with the development of economic technology and the improvement of living standard of people, the artificial intelligence technology is widely applied to the production and living of people, and brings endless convenience to the production and living of people.
The graphic neural network (Graph Neural Networks, GNN) is an important component of artificial intelligence technology. The graph neural network is widely applied to the fields of chemistry, physics, traffic, knowledge graph, recommendation systems and the like, and shows a strong application prospect. Therefore, for the study of the graph neural network, it has been an important work of researchers.
The GAT network (Graph attention networks, graph annotation meaning network) is one of the graph neural networks, and the GAT network processes graph structure data, and aggregates neighbor nodes through a attention mechanism (Attention Mechanism), so that the self-adaptive distribution of different neighbor weights is realized, and the expression capability of the graph neural network is greatly improved. However, recent studies have found that, similar to the deep neural networks such as the conventional convolutional neural networks (Convolutional Neural Networks, CNN), GAT networks also suffer from a lack of robustness. This problem makes GAT networks very vulnerable to attack, and an attacker can quickly degrade the performance of the model by adding only a small disturbance to the data. For fields with high security requirements, such as banks, finance and the like, the robustness of the GAT network model is particularly important.
Currently, studies indicate that attack methods of GAT networks tend to employ schemes that add edges and connect nodes with relatively large feature differences; for such attack schemes, researchers have proposed a corresponding defense scheme against GAT networks, i.e., preprocessing data using Jaccard similarity. However, these solutions can only process discrete data, artificially delete edges smaller than a threshold, but cannot learn in an end-to-end manner, and also do not establish a corresponding defense mechanism according to the characteristics of the attack method. This makes the current defense approach against GAT networks less reliable.
As such, when the existing GAT network is applied to the pedestrian detection field, the existing GAT network may be attacked due to poor defensive performance, so that a detection failure or a detection error occurs, and the reliability of the pedestrian detection process is seriously affected.
Disclosure of Invention
The invention aims to provide a GAT graph neural network defense method with high reliability and good stability.
The second purpose of the invention is to provide a construction method comprising the GAT graph neural network defense method.
It is a further object of the present invention to provide a pedestrian detection method including the construction method.
The GAT graph neural network defense method provided by the invention comprises the following steps:
s1, acquiring a network structure and network parameters of a GAT graph neural network;
s2, calculating cosine similarity of features among nodes in the GAT graph neural network according to the data information obtained in the step S1, converting, and modifying a function g in the GAT graph neural network by adopting the converted similarity;
s3, adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network based on the modification of the step S2;
s4, after abnormal node information is filtered, K neighbor operation is designed based on a K neighbor algorithm;
s5, in the training stage of the GAT graph neural network, randomly adding the K neighbor operation designed in the step S4, so as to realize characteristic enhancement of the nodes;
s6, finishing the defense of the GAT graph neural network.
And step S2, calculating cosine similarity of features among nodes in the GAT graph neural network, converting, and modifying a function g in the GAT graph neural network by adopting the converted similarity, wherein the method specifically comprises the following steps of:
the following equation is used as a modified function
Figure GDA0004242204400000031
Figure GDA0004242204400000032
In the middle of
Figure GDA0004242204400000033
Is node v in the graph neural network i Features at the first layer and
Figure GDA0004242204400000034
sigma () is a ReLU activation function, W 1 l For node v 1 Weight matrix at layer I, +.>
Figure GDA0004242204400000035
Is a weight coefficient and->
Figure GDA0004242204400000036
Figure GDA0004242204400000037
For node v 2 Weight matrix at layer I, +.>
Figure GDA0004242204400000038
Is the node v i A set of connected nodes; s () is a sigmoid function for adjusting the scaling, and +.>
Figure GDA0004242204400000039
Beta is a learnable hyper-parameter; sim () is a similarity calculation function, and +.>
Figure GDA00042422044000000310
Figure GDA00042422044000000311
For node v i Transpose of feature vector, ||h i The I is node v i The modulo length of the feature vector.
The step S3 of adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network specifically includes the following steps:
the weight coefficients are modified by the following formula, so that the attention of the previous layer of the GAT graph neural network is added to the weight of the current layer of the network:
Figure GDA00042422044000000312
in the middle of
Figure GDA00042422044000000313
The weight coefficient of the current layer; l is the first layer in the graph neural network; />
Figure GDA00042422044000000314
Modification obtained for step S2A subsequent function g; />
Figure GDA0004242204400000041
Is the weight coefficient of the previous layer.
The K nearest neighbor operation is designed based on the K nearest neighbor algorithm in the step S4, and specifically comprises the following steps:
the K nearest neighbor operation specifically comprises the following steps: based on a K neighbor algorithm, K nodes with the maximum similarity with a center node are selected through calculation of K neighbor nodes, and the center node is connected with the K nodes;
when in connection, directional edges are adopted for connection, and the direction of the directional edges points to a central node.
In the training stage of the GAT graph neural network, the step S5 adds the K nearest neighbor operation designed in the step S4 randomly, and specifically includes the following steps:
the following formula is adopted as a characteristic updating formula after K-nearest neighbor operation designed in the step S4:
Figure GDA0004242204400000042
in the middle of
Figure GDA0004242204400000043
Is the updated feature; KNN (x) i ) An operating function of the K-nearest neighbor operation designed for step S4, and KNN (x i )＝RANK(sim(x i (·), K), RANK () is the computation and node v i Function of the top K node index sets, sim (x i And,) is the compute node v i Feature x of (2) i K represents the number of nodes with the maximum similarity with the central node; p is the probability that the model randomly adds KNN to carry out node enhancement in the training stage; e-shaped article n Is a threshold of probability.
The invention also discloses a construction method of the GAT graph neural network defense method, which comprises the following steps:
A. determining a target GAT graph neural network;
B. c, defending the target GAT graph neural network determined in the step A by adopting the GAT graph neural network defending method;
C. and after the defense, obtaining the final defended GAT graph neural network, and completing the construction of the GAT graph neural network.
The invention also discloses a pedestrian detection method comprising the construction method, which comprises the following steps:
a. determining an initial GAT graph neural network model adopted by pedestrian detection in an automatic driving process;
b. adopting the initial GAT graph neural network model determined in the step a, and adopting the construction method to construct and obtain a GAT graph neural network;
c. and c, adopting the GAT graph neural network obtained in the step b to detect pedestrians in the automatic driving process.
According to the pedestrian detection method provided by the invention, through an innovative robust message transmission mechanism, the defending of the GAT graph neural network is realized, the robustness of the model is increased, the defending effect is good, and the reliability and the stability are high.
Drawings
Fig. 1 is a schematic flow chart of the defending method of the invention.
FIG. 2 is a schematic diagram showing the comparison of the effectiveness test of the defense method of the present invention with the prior art.
Figure 3 is a data schematic diagram comparing the effectiveness of the defense method of the present invention with the prior art.
FIG. 4 is a schematic flow chart of the construction method of the present invention.
Fig. 5 is a flow chart of the pedestrian detection method of the present invention.
Detailed Description
Fig. 1 is a schematic flow chart of the defending method of the present invention: the GAT graph neural network defense method provided by the invention comprises the following steps:
s1, acquiring a network structure and network parameters of a GAT graph neural network;
s2, calculating cosine similarity of features among nodes in the GAT graph neural network according to the data information obtained in the step S1, converting, and modifying a function g in the GAT graph neural network by adopting the converted similarity; the method specifically comprises the following steps:
the core of the step is to calculate cosine similarity of features among nodes in the graph, convert an output result of the similarity into a value between 0 and 1 through an s function of automatic learning and scaling adjustment, and design a function g in GAT into attention based on the feature similarity by utilizing the similarity so as to realize robust message transmission;
the following equation is used as a modified function
Figure GDA0004242204400000061
Figure GDA0004242204400000062
In the middle of
Figure GDA0004242204400000063
Is node v in the graph neural network i Features at the first layer and
Figure GDA0004242204400000064
sigma () is a ReLU activation function, W 1 l For node v 1 Weight matrix at layer I, +.>
Figure GDA0004242204400000065
Is a weight coefficient and->
Figure GDA0004242204400000066
Figure GDA0004242204400000067
For node v 2 Weight matrix at layer I, +.>
Figure GDA0004242204400000068
Is the node v i A set of connected nodes; s () is a sigmoid function for adjusting the scaleCount and->
Figure GDA0004242204400000069
Beta is a learnable hyper-parameter; sim () is a similarity calculation function, and +.>
Figure GDA00042422044000000610
Figure GDA00042422044000000611
For node v i Transpose of feature vector, ||h i The I is node v i The modular length of the feature vector; giving corresponding weight according to the feature similarity among the nodes;
s3, adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network based on the modification of the step S2; the method specifically comprises the following steps:
it is assumed that if two nodes are dissimilar, after feature transformation and propagation, the node features of the next hidden layer should also be dissimilar, i.e. the weight between the two nodes should still be small. However, weights are given according to node similarity, at each layer of the network
Figure GDA00042422044000000612
Based on the input features of the current layer only, and does not take into account information that has been obtained before;
thus, the weight coefficients are modified using the following formula to add the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network:
Figure GDA0004242204400000071
in the middle of
Figure GDA0004242204400000072
The weight coefficient of the current layer; l is the first layer in the graph neural network; />
Figure GDA0004242204400000073
The modified function g obtained in the step S2; />
Figure GDA0004242204400000074
The weight coefficient of the previous layer;
s4, after abnormal node information is filtered, K neighbor operation is designed based on a K neighbor algorithm; the method specifically comprises the following steps:
the K nearest neighbor operation specifically comprises the following steps: based on a K neighbor algorithm, K nodes with the maximum similarity with a center node are selected through calculation of K neighbor nodes, and the center node is connected with the K nodes;
when in connection, directional edges are adopted for connection, and the direction of the directional edges points to a central node;
s5, in the training stage of the GAT graph neural network, K neighbor operation designed in the step S4 is randomly added, feature enhancement of the nodes is achieved, and the model focuses on information of similar nodes; the method specifically comprises the following steps:
the following formula is adopted as a characteristic updating formula after K-nearest neighbor operation designed in the step S4:
Figure GDA0004242204400000075
in the middle of
Figure GDA0004242204400000076
Is the updated feature; KNN (x) i ) An operating function of the K-nearest neighbor operation designed for step S4, and KNN (x i )＝RANK(sim(x i (·), K), RANK () is the computation and node v i Function of the top K node index sets, sim (x i And,) is the compute node v i Feature x of (2) i K represents the number of nodes with the maximum similarity with the central node; p is the probability that the model randomly adds KNN to carry out node enhancement in the training stage; e-shaped article n Is a threshold of probability;
firstly, when the model starts training, KNN is added, and then model characteristics are updated; however, as models are trained, KNN increasingly does not help with performance improvement, possibly because newly added edges are not truly present in the data, and the models may overfit these artificially added data; therefore, after training for a certain period, the influence of KNN can be gradually reduced, so that the model is more fit with the original data;
s6, finishing the defense of the GAT graph neural network.
The defensive performance of the method of the invention is illustrated below by means of an example.
In order to evaluate the defending performance of the method, the embodiment performs the experiment of node classification under popular target attack and non-target attack. The target attack is an attack against a particular target node, and an attacker attacks nodes on a portion of the test set, causing the accuracy of the model on these nodes to drop. Under this setting, the present embodiment employs a popular Nettack attack approach. The purpose of the non-target attack is to reduce the classification accuracy of the model on all test sets through the attack, and the model is global attack, and under the setting, the embodiment adopts a popular Mettack attack method.
Against target attacks:
the target attack is classified into a direct attack and an indirect attack, wherein the direct attack changes the connection between the target node and the neighboring node, and the indirect attack changes the connection between the neighboring node of the target node and the neighboring node thereof. Because the direct attack has the strongest aggressiveness, in order to verify the robustness of the method of the present invention, the present embodiment adopts the direct attack in the experiment. Nettack is a popular target attack method, and this embodiment uses default parameter settings in its original paper. Specifically, after GCN (surrogate model) training is completed, in this embodiment, 10 nodes with highest scores, 10 nodes with lowest scores and 20 nodes with random nodes with highest scores in the correct classification node set in the test set are selected to respectively attack, and finally, the accuracy of 40 nodes after attack is counted. This embodiment attacks both edges and features. The model accuracy after Nettack direct attack is shown in figure 2. It can be seen that the accuracy of the GCN-based model on the various data sets is very low, largely because the alternative model for Nettack is GCN. Nonetheless, the baseline model SAGE-MP of this example performed poorly on 3 data sets, but the performance of the method of the present invention far exceeded SAGE-MP.
Against non-target attacks:
in the embodiment, mettack attack is adopted to test the performance of different models against non-target attack under the node classification task, the parameters follow the setting in the original paper, and the most aggressive version Meta-Self is adopted. The disturbance rate (Perturbation Rate, PR) is the change rate of the back edge of the structural data of the attack graph of an attacker, and the change range of the value is 0 to 25 percent, and the step length is 5 percent. The performance of the different models under Mettack attack on Cora, citeseer and Cora_ML data sets is shown in FIG. 3, where the best performance is shown in bold.
From the data in fig. 3, it can be known that:
1) Mettack attack causes the performance of all models to be reduced to different degrees;
2) Compared with GAT based on feature attention and RGCN based on variance attention, the method based on similarity attention has better robustness;
3) The SAGE-MP also has greatly reduced performance under Mettack attack, and the SAGE-MP-based method can greatly improve the robustness.
Experiments show that compared with other methods, the method has the advantages of higher defense performance improvement, more robust defense capability and balance between robustness and accuracy.
FIG. 4 is a schematic flow chart of the construction method of the present invention: the construction method of the neural network defense method comprising the GAT graph disclosed by the invention comprises the following steps of:
A. determining a target GAT graph neural network;
B. c, defending the target GAT graph neural network determined in the step A by adopting the defending method of the GAT graph neural network;
C. and after the defense, obtaining the final defended GAT graph neural network, and completing the construction of the GAT graph neural network.
The GAT graph neural network defense method described in the step B comprises the following steps:
B1. acquiring a network structure and network parameters of a GAT graph neural network;
B2. according to the data information obtained in the step B1, calculating cosine similarity of features among nodes in the GAT graph neural network, converting, and modifying a function g in the GAT graph neural network by adopting the converted similarity;
B3. based on the modification of the step B2, adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network;
B4. after filtering abnormal node information, designing K neighbor operation based on a K neighbor algorithm;
B5. in the training stage of the GAT graph neural network, randomly adding the K neighbor operation designed in the step B4 to realize the characteristic enhancement of the node;
B6. and (5) finishing the defense of the GAT graph neural network.
And B2, calculating cosine similarity of features among nodes in the GAT graph neural network, converting, and modifying a function g in the GAT graph neural network by adopting the converted similarity, wherein the method specifically comprises the following steps of:
the following equation is used as a modified function
Figure GDA0004242204400000101
Figure GDA0004242204400000102
In the middle of
Figure GDA0004242204400000103
Is node v in the graph neural network i Features at the first layer and
Figure GDA0004242204400000104
sigma () is a ReLU activation function, W 1 l For node v 1 Weight matrix at layer I, +.>
Figure GDA0004242204400000105
Is a weight coefficient and->
Figure GDA0004242204400000106
Figure GDA0004242204400000107
For node v 2 Weight matrix at layer I, +.>
Figure GDA0004242204400000108
Is the node v i A set of connected nodes; s () is a sigmoid function for adjusting the scaling, and +.>
Figure GDA0004242204400000109
Beta is a learnable hyper-parameter; sim () is a similarity calculation function, and +.>
Figure GDA0004242204400000111
Figure GDA0004242204400000112
For node v i Transpose of feature vector, ||h i The I is node v i The modulo length of the feature vector.
The step B3 of adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network specifically includes the following steps:
the weight coefficients are modified by the following formula, so that the attention of the previous layer of the GAT graph neural network is added to the weight of the current layer of the network:
Figure GDA0004242204400000113
in the middle of
Figure GDA0004242204400000114
The weight coefficient of the current layer; l is the first layer in the graph neural network; />
Figure GDA0004242204400000115
The modified function g obtained in the step S2; />
Figure GDA0004242204400000116
Is the weight coefficient of the previous layer.
The K neighbor operation is designed based on the K neighbor algorithm in the step B4, and specifically comprises the following steps:
the K nearest neighbor operation specifically comprises the following steps: based on a K neighbor algorithm, K nodes with the maximum similarity with a center node are selected through calculation of K neighbor nodes, and the center node is connected with the K nodes;
when in connection, directional edges are adopted for connection, and the direction of the directional edges points to a central node.
In the training stage of the GAT graph neural network, the step B5 is implemented by randomly adding the K neighbor operation designed in the step B4, and specifically includes the following steps:
the following formula is adopted as a characteristic updating formula after K-nearest neighbor operation designed in the step S4:
Figure GDA0004242204400000117
in the middle of
Figure GDA0004242204400000118
Is the updated feature; KNN (x) i ) An operating function of the K-nearest neighbor operation designed for step S4, and KNN (x i )＝RANK(sim(x i (·), K), RANK () is the computation and node v i Function of the top K node index sets, sim (x i And,) is the compute node v i Feature x of (2) i K represents the number of nodes with the maximum similarity with the central node; p is the probability that the model randomly adds KNN to carry out node enhancement in the training stage; e-shaped article n Is a threshold of probability.
Fig. 5 is a flow chart of a pedestrian detection method according to the present invention: the pedestrian detection method comprising the construction method disclosed by the invention comprises the following steps of:
a. determining an initial GAT graph neural network model adopted by pedestrian detection in an automatic driving process;
b. adopting the initial GAT graph neural network model determined in the step a, and adopting the construction method to construct and obtain a GAT graph neural network;
c. and c, adopting the GAT graph neural network obtained in the step b to detect pedestrians in the automatic driving process.
The building method in the step b specifically comprises the following steps:
b1. determining a target GAT graph neural network;
b2. c, defending the target GAT graph neural network determined in the step b1 by adopting the GAT graph neural network defending method;
b3. and after the defense, obtaining the final defended GAT graph neural network, and completing the construction of the GAT graph neural network.
The GAT graph neural network defense method described in the step b2 comprises the following steps:
b21. acquiring a network structure and network parameters of a GAT graph neural network;
b22. c, calculating cosine similarity of features among nodes in the GAT graph neural network and converting the cosine similarity according to the data information obtained in the step b21, and modifying a function g in the GAT graph neural network by adopting the converted similarity;
b23. based on the modification of step b22, adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network;
b24. after filtering abnormal node information, designing K neighbor operation based on a K neighbor algorithm;
b25. in the training stage of the GAT graph neural network, randomly adding the K neighbor operation designed in the step b24 to realize the characteristic enhancement of the node;
b26. and (5) finishing the defense of the GAT graph neural network.
The step b22 of calculating and converting cosine similarity of features among nodes in the GAT graph neural network, and modifying a function g in the GAT graph neural network by adopting the converted similarity, specifically comprising the following steps:
the following equation is used as a modified function
Figure GDA0004242204400000131
Figure GDA0004242204400000132
In the middle of
Figure GDA0004242204400000133
Is node v in the graph neural network i Features in the first layer and->
Figure GDA0004242204400000134
Sigma () is a ReLU activation function, W 1 l For node v 1 Weight matrix at layer I, +.>
Figure GDA0004242204400000135
Is a weight coefficient and->
Figure GDA0004242204400000136
Figure GDA0004242204400000137
For node v 2 Weight matrix at layer I, +.>
Figure GDA0004242204400000138
Is the node v i A set of connected nodes; s () is a sigmoid function for adjusting the scaling, and +.>
Figure GDA0004242204400000139
Beta is a learnable hyper-parameter; sim () is a similarity calculation function, and
Figure GDA00042422044000001310
Figure GDA00042422044000001311
for node v i Transpose of feature vector, ||h i The I is node v i The modulo length of the feature vector.
The weight of adding the attention of the previous layer of the GAT graph neural network to the current layer of the network in the step b23 specifically comprises the following steps:
the weight coefficients are modified by the following formula, so that the attention of the previous layer of the GAT graph neural network is added to the weight of the current layer of the network:
Figure GDA0004242204400000141
in the middle of
Figure GDA0004242204400000142
The weight coefficient of the current layer; l is the first layer in the graph neural network; />
Figure GDA0004242204400000143
The modified function g obtained in the step S2; />
Figure GDA0004242204400000144
Is the weight coefficient of the previous layer.
The K-nearest neighbor operation based on the K-nearest neighbor algorithm described in the step b24 specifically includes the following steps:
the K nearest neighbor operation specifically comprises the following steps: based on a K neighbor algorithm, K nodes with the maximum similarity with a center node are selected through calculation of K neighbor nodes, and the center node is connected with the K nodes;
when in connection, directional edges are adopted for connection, and the direction of the directional edges points to a central node.
In the training stage of the GAT graph neural network, the step b25 is implemented by randomly adding the K-nearest neighbor operation designed in the step b24, and specifically includes the following steps:
the following formula is adopted as a characteristic updating formula after K-nearest neighbor operation designed in the step S4:
Figure GDA0004242204400000145
in the middle of
Figure GDA0004242204400000146
Is the updated feature; KNN (x) i ) An operating function of the K-nearest neighbor operation designed for step S4, and KNN (x i )＝RANK(sim(x i (·), K), RANK () is the computation and node v i Function of the top K node index sets, sim (x i And,) is the compute node v i Feature x of (2) i K represents the number of nodes with the maximum similarity with the central node; p is the probability that the model randomly adds KNN to carry out node enhancement in the training stage; e-shaped article n Is a threshold of probability.
The human detection method provided by the invention is particularly suitable for pedestrian detection of automatic driving of vehicles in the prior industry; the targets such as people, vehicles, objects and the like on the road need to be closely focused in the automatic driving process, and a target identification technology based on vision or vehicle-mounted laser radar is needed. In order to improve accuracy and interpretability of recognition, many current target recognition technologies adopt a graph neural network technology (i.e., a GAT graph neural network) fused with an attention mechanism, because GAT can model relationships among various parts of targets such as people, vehicles, objects and the like, for example, aiming at target recognition of pedestrians, a vehicle-mounted vision/laser radar sensor can acquire pictures of the targets of the pedestrians, and model key points of a skeleton of the pedestrians in the pictures and relationships among the key points in the skeleton of the pedestrians in the pictures in the form of nodes and edges in the GAT graph neural network. Because the requirements of automatic driving on safety are high, if an attacker tries to attack by adding disturbance to the picture, misjudgment occurs in the target identification of people, vehicles and objects facing the automatic driving, and serious traffic accidents are caused. Therefore, the pedestrian detection method provided by the invention is used for detecting pedestrians in real time, not only can ensure the detection precision and reliability, but also the detection model has better defending performance, can realize the graph neural network defending of robust message transmission, and improves the capability of defending attacks for the target identification of the automatic driving person, vehicle and object.

Claims (3)

1. A pedestrian detection method specifically comprises the following steps:
a. determining an initial GAT graph neural network model adopted by pedestrian detection in an automatic driving process;
b. adopting the initial GAT graph neural network model determined in the step a, and adopting a construction method to construct and obtain a GAT graph neural network;
c. b, acquiring a picture of a pedestrian target acquired by a vehicle-mounted vision/laser radar sensor by adopting the GAT graph neural network obtained in the step b, modeling key points of a pedestrian body skeleton in the picture and the relation between the key points by using nodes and edges in the GAT graph neural network obtained in the step b, and detecting pedestrians in an automatic driving process;
the construction method specifically comprises the following steps:
A. taking the initial GAT graph neural network model determined in the step a as a target GAT graph neural network;
B. c, defending the target GAT graph neural network determined in the step A by adopting a GAT graph neural network defending method;
C. after the defense, obtaining a final defended GAT graph neural network, and completing the construction of the GAT graph neural network;
the GAT graph neural network defense method comprises the following steps:
s1, acquiring a network structure and network parameters of a target GAT graph neural network;
s2, calculating cosine similarity of features among nodes in the GAT graph neural network according to the data information obtained in the step S1, converting, and modifying a function g in the GAT graph neural network by adopting the converted similarity; the method specifically comprises the following steps:
the following equation is used as a modified function
Figure FDA0004242204340000011
Figure FDA0004242204340000012
In the middle of
Figure FDA0004242204340000013
Is node v in the graph neural network i Features at the first layer and
Figure FDA0004242204340000021
sigma () is a ReLU activation function, +.>
Figure FDA0004242204340000022
For node v 1 Weight matrix at layer I, +.>
Figure FDA0004242204340000023
Is a weight coefficient and->
Figure FDA0004242204340000024
Figure FDA0004242204340000025
For node v 2 Weight matrix at layer I, +.>
Figure FDA0004242204340000026
Is the node v i A set of connected nodes; s () is a sigmoid function for adjusting the scaling, and +.>
Figure FDA0004242204340000027
Beta is a learnable hyper-parameter; sim () is a similarity calculation function, and +.>
Figure FDA0004242204340000028
Figure FDA0004242204340000029
For node v i Transpose of feature vector, ||h i The I is node v i The modular length of the feature vector;
s3, adding the attention of the previous layer of the GAT graph neural network to the weight of the current layer of the network based on the modification of the step S2; the method specifically comprises the following steps:
the weight coefficients are modified by the following formula, so that the attention of the previous layer of the GAT graph neural network is added to the weight of the current layer of the network:
Figure FDA00042422043400000210
in the middle of
Figure FDA00042422043400000211
The weight coefficient of the current layer; l is the first layer in the graph neural network; />
Figure FDA00042422043400000212
The modified function g obtained in the step S2; />
Figure FDA00042422043400000213
The weight coefficient of the previous layer;
s4, after abnormal node information is filtered, K neighbor operation is designed based on a K neighbor algorithm;
s5, in the training stage of the GAT graph neural network, randomly adding the K neighbor operation designed in the step S4, so as to realize characteristic enhancement of the nodes;
s6, finishing the defense of the GAT graph neural network.
2. The pedestrian detection method according to claim 1, wherein the K-nearest neighbor operation is designed based on the K-nearest neighbor algorithm in step S4, and specifically comprises the following steps:
the K nearest neighbor operation specifically comprises the following steps: based on a K neighbor algorithm, K nodes with the maximum similarity with a center node are selected through calculation of K neighbor nodes, and the center node is connected with the K nodes;
when in connection, directional edges are adopted for connection, and the direction of the directional edges points to a central node.
3. The pedestrian detection method according to claim 2, wherein in the training stage of the GAT graph neural network in step S5, the K-nearest neighbor operation designed in step S4 is randomly added, and the method specifically includes the following steps:
the following formula is adopted as a characteristic updating formula after K-nearest neighbor operation designed in the step S4:
Figure FDA0004242204340000031
in the middle of
Figure FDA0004242204340000032
Is the updated feature; KNN (x) i ) An operating function of the K-nearest neighbor operation designed for step S4, and KNN (x i )＝RANK(sim(x i (·), K), RANK () is the computation and node v i Function of the top K node index sets, sim (x i And,) is the compute node v i Feature x of (2) i K represents the number of nodes with the maximum similarity with the central node; p is the probability that the model randomly adds KNN to carry out node enhancement in the training stage; e-shaped article n Is a threshold of probability.
CN202211412927.7A 2022-11-11 2022-11-11 Pedestrian detection method Active CN115906980B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211412927.7A CN115906980B (en) 2022-11-11 2022-11-11 Pedestrian detection method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211412927.7A CN115906980B (en) 2022-11-11 2022-11-11 Pedestrian detection method

Publications (2)

Publication Number Publication Date
CN115906980A CN115906980A (en) 2023-04-04
CN115906980B true CN115906980B (en) 2023-06-30

Family

ID=86472155

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211412927.7A Active CN115906980B (en) 2022-11-11 2022-11-11 Pedestrian detection method

Country Status (1)

Country Link
CN (1) CN115906980B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113806546A (en) * 2021-09-30 2021-12-17 中国人民解放军国防科技大学 Cooperative training-based method and system for defending confrontation of graph neural network
CN114708479A (en) * 2022-03-31 2022-07-05 杭州电子科技大学 Self-adaptive defense method based on graph structure and characteristics

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113806546A (en) * 2021-09-30 2021-12-17 中国人民解放军国防科技大学 Cooperative training-based method and system for defending confrontation of graph neural network
CN114708479A (en) * 2022-03-31 2022-07-05 杭州电子科技大学 Self-adaptive defense method based on graph structure and characteristics

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Urban Traffic Flow Forecast Based on FastGCRNN;Ya Zhang等;Journal of Advanced Transportation;第2020卷;1-9 *

Also Published As

Publication number Publication date
CN115906980A (en) 2023-04-04

Similar Documents

Publication Publication Date Title
CN110781776B (en) Road extraction method based on prediction and residual refinement network
CN110941794A (en) Anti-attack defense method based on universal inverse disturbance defense matrix
CN103226820B (en) The two-dimensional maximum entropy division night vision image fusion target detection algorithm improved
CN111967343A (en) Detection method based on simple neural network and extreme gradient lifting model fusion
CN104820997B (en) A kind of method for tracking target based on piecemeal sparse expression Yu HSV Feature Fusion
CN109635763B (en) Crowd density estimation method
CN112489054A (en) Remote sensing image semantic segmentation method based on deep learning
CN112241724A (en) Automatic identification method and system based on double-path convolution long-term and short-term neural network
CN113657491A (en) Neural network design method for signal modulation type recognition
CN106411572A (en) Community discovery method combining node information and network structure
CN107944354A (en) A kind of vehicle checking method based on deep learning
CN103955950A (en) Image tracking method utilizing key point feature matching
CN112084895B (en) Pedestrian re-identification method based on deep learning
CN111178504B (en) Information processing method and system of robust compression model based on deep neural network
CN108093213B (en) Target track fuzzy data fusion method based on video monitoring
CN114708479B (en) Self-adaptive defense method based on graph structure and characteristics
CN113962281A (en) Unmanned aerial vehicle target tracking method based on Siamese-RFB
CN113807214B (en) Small target face recognition method based on deit affiliated network knowledge distillation
CN105894012B (en) Based on the object identification method for cascading micro- neural network
CN115906980B (en) Pedestrian detection method
CN109492816B (en) Coal and gas outburst dynamic prediction method based on hybrid intelligence
CN111027542A (en) Target detection method improved based on fast RCNN algorithm
CN109800517A (en) Improved reverse modeling method for magnetorheological damper
CN112396137B (en) Point cloud semantic segmentation method integrating context semantics
CN111563513B (en) Defocus blur detection method based on attention mechanism

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
GR01 Patent grant
GR01 Patent grant