CN117253138A - Deep learning position recognition algorithm based on laser radar - Google Patents

Abstract

The point cloud position recognition technology in the dynamic environment has the advantages of high detection precision, high detection speed and small model calculation amount. The method adopts a characteristic point extraction module to greatly reduce the scale of the point cloud and retain the data characteristics so as to lighten the influence of environmental change on data acquisition, and designs a Point Transformer module which introduces Multi Head Attention capable of promoting local information exchange to control the calculated amount and extract local descriptors; feature Attention is designed to aggregate recognizable global descriptors; thereby improving the accuracy and speed of the point cloud position identification and reducing the network calculation amount.

Description

Deep learning position recognition algorithm based on laser radar

Technical Field

The deep learning position recognition algorithm based on the laser radar is used for improving the accuracy and speed of position recognition.

Technical Field

In recent years, some vision-based position recognition methods are proposed, but the vision sensor has high sensitivity to light, and is difficult to apply to dynamic environments; as a reliable alternative, 3D lidar can acquire accurate and detailed three-dimensional information that is inherently immune to illumination variations; however, the position recognition method based on the 3D lidar cannot obtain an effective and robust detection effect due to environmental factors such as sensor shielding and viewpoint change.

From the recent academic research, the accuracy and the robustness of the position recognition technology based on the point cloud in the dynamic environment are difficult to be ensured due to environmental factors such as sensor shielding, viewpoint change and the like; the existing method with high detection precision has the defects that a network model is too large, the detection speed of the method with small network model is not fast enough, and the method is difficult to be applied to a real scene; in order to solve the above problems, we propose an efficient and lightweight point cloud location recognition learning network combining feature point extraction and transformation.

Disclosure of Invention

The point cloud position recognition technology in the dynamic environment has the advantages of high detection precision, high detection speed and small model calculation amount.

The characteristic point extraction module is adopted to greatly reduce the scale of the point cloud and preserve the data characteristics so as to reduce the influence of environmental change on data acquisition, the Point Transformer module is designed to introduce a multi-head attention capable of promoting local information exchange to control the calculated amount, the expression capacity of the model is enhanced, and the accuracy and the speed of point cloud position identification are improved and the network calculated amount is reduced.

A deep learning position recognition algorithm based on a laser radar comprises the following steps:

step one: firstly, calculating a horizontal index and a scan index of each point in point cloud data acquired by a laser radar; the expression is as follows:

wherein x, y and z are three-dimensional coordinate values of each point; delta is the absolute value of the angle which can be scanned under the horizontal axis of the laser radar;

step two: each point is cycled, and the point r is used _i Calculating the curvature value of each point by using a sub-point cloud S consisting of five points which are positioned in front of and behind the same scan index and closest to the horizontal index; discarding if the curvature value of the point is smaller than a certain threshold value; finally obtaining characteristic points; the formula is expressed as:

C _i <σ (4)

step three: the feature points are copied into 3 parts of data K, Q and V with equal size, and the data K, Q and V are processed by Multi Head Attention to obtain local descriptors; the formula is expressed as follows:

MultiHead(Q,K,V)＝Concat(head ₁ ,...,head _h )W ^O (4)

wherein W is _i ^Q ，W _i ^K And W is _i ^V ∈R ^D×(D/h) ,W ^O ∈R ^D×D Are all a matrix of learnable parameters;

step four: aggregating global descriptors by assigning local descriptor weights to Feature Attention; the formula is expressed as follows:

wherein N is the number of local descriptors, u is the number of local descriptors, and J is a learnable matrix parameter;

step five: obtaining corresponding scene similarity by comparing the global descriptor similarity; the formula is described as follows:

wherein omega ^[1:s] ∈R ^D×D×S ，α∈R ^2D×S Is a learnable matrix parameter b epsilon R ^S Is a learnable vector parameter, S is the dimension size of the global descriptor.

Step six: in the training stage, the model has the function of learning by obtaining scene similarity output, returning to the input position to compare the true value, comparing the difference value between the scene similarity output and the true value, and modifying the weight parameters in the next training again.

The method is mainly characterized in that redundant original point clouds firstly extract characteristic points, meaningless points are removed, the scale of the point clouds is greatly reduced, and the calculated amount is further reduced; the Point Transformer module learns the local features of the feature points well and aggregates the discernable global descriptors.

Drawings

Description of the drawings figure 1 is a general flow chart of the invention

FIG. 2 is a flow chart of the Muti-Head addition

Description of the drawings figure 3 is a flow chart of Feature Attention

Description of the drawings figure 4 is a flow chart of Feature Similarity Network

FIG. 5 is a partial experimental result diagram

Detailed description of the preferred embodiments

As shown in the overall flow of FIG. 1, the invention herein provides a deep learning position based on laser radar

An identification algorithm.

Firstly, extracting point cloud data from two paired scenes through a Feature Point Extrarction module to obtain sub point clouds with rich characteristics. Then extracting the local descriptor F through the Muti-Head attribute layer in the Point Transformer module _L Feature Attention layer aggregates distinct global descriptors. Finally, two global descriptors e to be paired ₁ And e ₂ And the incoming Feature similarity network module obtains the scene similarity. The method specifically comprises the following steps:

an origin point cloud is acquired in a scene using a lidar sensor.

The original point cloud is first projected as a Range image and then the scan index and the horizontal index for each point are calculated.

Each point in the original point cloud is traversed circularly, the curvature value of the point is calculated through a sub point cloud S formed by five points which are positioned in the same scan index and closest to the point, and if the curvature value of the point is smaller than a certain threshold value, the curvature value of the point is omitted; the final rest is the feature points;

as shown in fig. 2, the feature points are mapped into three equal-sized parts, denoted by K, Q and V, using a linear projection layer; and then dividing K, Q and V into h parts respectively and transmitting the h parts into Multi Head Attention for extracting local descriptors.

As shown in fig. 3, the global context is constructed by multiplying the learning matrix J with the average value of all the local descriptors to provide global structure and feature information; and multiplying the local descriptor by using the global context, ensuring the obtained weight value to be 0,1 by using a sigmoid function, and finally giving the weight to the local descriptor to obtain the final global descriptor through summation.

As shown in fig. 4, feature Similarity Network is used to obtain the similarity of two global descriptors, so as to obtain the similarity of the scene.

As shown in fig. 5, our algorithm can well distinguish subtle differences in scenes, giving a suitable score for the similarity of two scenes; scene 1 and the compared scene are very similar in terms of both whole and detail, so that a higher similarity score is obtained, and scene 2 is completely opposite; scene 3, while similar to the compared scene as a whole, is not identical in detail, and therefore results in a lower similarity score.

Claims (1)

1. A deep learning position recognition algorithm based on laser radar is characterized by comprising the following steps,

step one: firstly, calculating a horizontal index and a scan index of each point in point cloud data acquired by a laser radar; the expression is as follows:

wherein x, y and z are three-dimensional coordinate values of each point; delta is the absolute value of the angle which can be scanned under the horizontal axis of the laser radar;

step two: each point is cycled, and the point r is used _i Calculating the curvature value of each point by using a sub-point cloud S consisting of five points which are positioned in front of and behind the same scan index and closest to the horizontal index; discarding if the curvature value of the point is smaller than a certain threshold value; finally obtaining characteristic points; the formula is expressed as:

C _i <σ (4)

step three: the feature points are copied into 3 parts of data K, Q and V with equal size, and the data K, Q and V are processed by Multi Head Attention to obtain local descriptors; the formula is expressed as follows:

MultiHead(Q,K,V)＝Concat(head ₁ ,...,head _h )W ^O (4)

wherein W is _i ^Q ，W _i ^K And W is _i ^V ∈R ^D×(D/h) ,W ^O ∈R ^D×D Are all a matrix of learnable parameters;

step four: aggregating global descriptors by assigning local descriptor weights to Feature Attention; the formula is expressed as follows:

wherein N is the number of local descriptors, u is the number of local descriptors, and J is a learnable matrix parameter;

step five: obtaining corresponding scene similarity by comparing the global descriptor similarity; the formula is described as follows:

wherein omega ^[1:s] ∈R ^D×D×S ，α∈R ^2D×S Is a learnable matrix parameter b epsilon R ^S Is a learnable vector parameter, S is the dimension size of the global descriptor.

Step six: in the training stage, the model has the function of learning by obtaining scene similarity output, returning to the input position to compare the true value, comparing the difference value between the scene similarity output and the true value, and modifying the weight parameters in the next training again.