CN-121977584-A - Depth-space information combined guidance-based real-time path planning method and system for robot

CN121977584ACN 121977584 ACN121977584 ACN 121977584ACN-121977584-A

Abstract

The invention discloses a robot real-time path planning method and a system based on depth-space information combined guidance, wherein the robot real-time path planning method constructs a path planning model, and the depth image, the gesture information and the expected speed are input into a path planning model, so that the mobile robot can efficiently avoid the obstacle. The path planning model generates an obstacle avoidance correlation weight matrix in a perception network through depth-space information coding, guides a vision transducer network to focus effective information and restrain redundant interference, guides a scene representation to be input into a decision network through the generated depth-space information, outputs a motion control instruction, improves the representation intensity of obstacle avoidance related features through targeted distribution of feature weights, reduces interference of irrelevant features, is suitable for various mobile robots needing autonomous obstacle avoidance, such as unmanned aerial vehicles, ground mobile robots and the like, and remarkably improves the obstacle avoidance success rate and real-time response speed of the robots in complex environments.

Inventors

QIAN MING
ZHOU JIE
LOU TIECHENG
YE XIA
CUI WENFENG

Assignees

浙江飞航智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. A robot real-time path planning method based on depth-space information combined guidance is characterized by comprising the following steps: constructing a data set containing depth images, attitude information and expected speed, and carrying out labeling treatment on the data set; The method comprises the steps of constructing a path planning model, wherein the path planning model comprises a perception network and a decision network, and the perception network is used for converting a depth image into a front scene representation, inputting the front scene representation, gesture information and expected speed into the decision network for processing to obtain a predicted action instruction; In the depth-space information guiding module, carrying out element-by-element weighted scaling on an input feature map through a depth-space guiding weight matrix, and carrying out residual connection on the weighted feature map and the input feature map to obtain an output feature map of the depth-space information guiding module; the method comprises the steps of training a path planning model by using a data set, predicting the actual movement speed of the robot by using the trained path planning model, and controlling the movement of the robot according to the actual movement speed, so that the real-time path planning of the robot is realized.
2. The method for planning the real-time path of the robot based on the combined guidance of the depth and the space information, which is disclosed in claim 1, is characterized in that the perception network comprises a plurality of coding modules and a feature fusion module which are connected in series, the plurality of coding modules process the depth image in sequence, and the feature fusion module fuses the output feature graphs of the coding modules to obtain the front scene representation output by the perception network.
3. The method for planning the real-time path of the robot based on the combined guidance of the depth and the space information, which is disclosed in claim 2, is characterized in that the coding module comprises a downsampling layer and a plurality of feature extraction units which are sequentially connected, and the feature extraction units comprise a depth-space information guiding module and a vision transducer multi-head attention module which are connected in series.
4. The method for planning the real-time path of the robot based on the depth-space information joint guidance of claim 2 is characterized in that in a feature fusion module, pixel recombination operation is adopted to up-sample an output feature map except for a first coding module, and after channel splicing is carried out on the output feature map of each coding module, dimension fusion is carried out through a convolution layer to obtain the output feature map of the feature fusion module.
5. The method for planning the real-time path of the robot based on the combined guidance of the depth and the space information, which is disclosed in claim 1, is characterized in that the depth guidance weight matrix is constructed based on depth image information, and the space guidance weight matrix is generated based on two-dimensional space position information of an obstacle relative to the robot.
6. The method for planning a real-time path of a robot based on joint guidance of depth-space information as set forth in claim 1, wherein the depth-space guidance weight matrix is The construction method of (2) is as follows: Wherein, the Is a weight coefficient; A depth value for each pixel; x and y are the abscissa and ordinate of each pixel, respectively; And Respectively the abscissa and the ordinate of the Gaussian distribution center point; And Standard deviations of Gaussian distribution in the width direction and the height direction are respectively shown; Is an error compensation coefficient.
7. The method for planning the real-time path of the robot based on the depth-space information joint guidance is characterized in that in a vision transducer multi-head attention module, an input feature map is split into a plurality of feature blocks, the feature blocks are mapped into queries, keys and values respectively through independent linear layers, after the keys and the values are downsampled, the multi-head attention is jointly calculated through the queries and fused to obtain attention output, and nonlinear transformation is carried out on the attention output through a mixed feedforward layer to obtain an output feature map of the vision transducer multi-head attention module.
8. The method for planning the real-time path of the robot based on the depth-space information joint guidance of claim 1, wherein the method is characterized in that in a decision network, front scene representation, gesture information and expected speed are flattened and vector spliced, and then input into a plurality of serial full-connection layers to obtain an output result of the decision network.
9. The real-time path planning method for the robot based on the depth-space information joint guidance is characterized by training a path planning model by using a preprocessed data set, wherein the preprocessing process comprises the steps of carrying out normalization processing on depth images in the data set, mapping depth values I d to a [0,1] interval, carrying out filtering denoising on gesture information by adopting a Kalman filtering algorithm, and carrying out time stamp alignment on all data.
10. The robot real-time path planning system based on the depth-space information joint guidance is characterized by being used for executing the robot real-time path planning method based on the depth-space information joint guidance, and comprises an image acquisition module, a map construction module and a motion planning module, wherein the image acquisition module is used for acquiring depth image data, the map construction module is used for constructing a three-dimensional point cloud map and acquiring gesture information of a robot, and the motion planning module is used for determining the actual motion speed of the robot in the three-dimensional point cloud map according to the depth image, the gesture information and the expected speed.

Description

Depth-space information combined guidance-based real-time path planning method and system for robot Technical Field The invention belongs to the technical field of artificial intelligence and robot control intersection, and particularly relates to a depth-space information combined guidance-based real-time path planning method and system for a robot. Background Along with the popularization of the robot technology, the application scene of the robot is expanded to the complex fields of electric power inspection, post-disaster search and rescue, urban air traffic, warehouse logistics, ground inspection and the like, and higher requirements are put forward on the real-time obstacle avoidance capability of the mobile robot in a dense obstacle environment. The traditional obstacle avoidance algorithm depends on a high-precision priori map and a layered architecture of 'perception-positioning-planning-control', is high in system complexity and weak in anti-interference capability, is difficult to meet real-time response requirements during high-speed movement, is poor in universality, and is difficult to adapt to different types of mobile robots. In recent years, an end-to-end learning-based method has become a main technical route for autonomous navigation of a robot by virtue of direct mapping from sensor data to control instructions. The lightweight depth camera, acting as a core sensor, can provide dense spatial depth information, providing data support for end-to-end control. However, the conventional end-to-end method still has obvious limitations that a Convolutional Neural Network (CNN) is limited by a local receptive field, global spatial relation of a remote obstacle is difficult to model, traditional vision Transformer (ViT) can capture long-range dependence, but does not distinguish obstacle avoidance related and irrelevant features through an effective mechanism, namely depth images contain a large amount of threat-free background areas, remote ineffective spaces and other irrelevant information, the information is mixed with obstacle avoidance related features such as obstacle outlines, distances, spatial distribution and the like, so that model feature extraction efficiency is low, effective information weight is insufficient, accuracy and instantaneity of obstacle avoidance decision are directly influenced, and in addition, the conventional method is mostly designed aiming at a specific type of robot (such as a single unmanned aerial vehicle), and has insufficient universality and is difficult to quickly migrate to other mobile robot platforms. Therefore, how to integrate depth-space information guidance into a network architecture, strengthen obstacle avoidance related features and inhibit obstacle avoidance unrelated features through differential distribution of feature weights, simultaneously maintain global perceptibility and improve method generality, adapt to obstacle avoidance requirements of various mobile robots, and become a technical problem to be solved urgently in the field of high-speed obstacle avoidance of robots. Disclosure of Invention The invention provides a depth-space information combined guidance-based real-time path planning method and system for robots, which aim to solve the core problems of confusion of relevant obstacle avoidance features and irrelevant features and insufficient effective information weight in the existing end-to-end obstacle avoidance method, and simultaneously improve the universality of the method, adapt to various mobile robots such as unmanned aerial vehicles, ground mobile robots and the like, ensure the global perception capability and improve the obstacle avoidance robustness and instantaneity of the robots in the unknown complex environment during high-speed movement. In a first aspect, the present invention provides a method for planning a real-time path of a robot based on depth-space information joint guidance, the method comprising: constructing a data set containing depth images, attitude information and expected speed, and carrying out labeling treatment on the data set; The method comprises the steps of constructing a path planning model, wherein the path planning model comprises a perception network and a decision network, and the perception network is used for converting a depth image into a front scene representation, inputting the front scene representation, gesture information and expected speed into the decision network for processing to obtain a predicted action instruction; In the depth-space information guiding module, carrying out element-by-element weighted scaling on an input feature map through a depth-space guiding weight matrix, and carrying out residual connection on the weighted feature map and the input feature map to obtain an output feature map of the depth-space information guiding module; the method comprises the steps of training a path planning model by using a data set, predicting the actual movement speed of th