Visual target detection method for detection robot

Abstract

The invention discloses a visual target detection method for a detection robot. The method comprises the following steps of: constructing a hybrid attention adaptive sensing network; training the hybrid attention adaptive sensing network; and detecting the visual target of the detection robot by using the trained hybrid attention adaptive sensing network. The hybrid attention adaptive perception network comprises a global attention hierarchical network, a target space positioning network, a self-adaptive local attention sensing network and a target classification and identification network. According to the method of the invention, target positioning and identification tasks can be completed in a unified framework; the backbone networks are all based on weight sharing under a global attention mechanism, the advantages of a two-stage method and a one-stage method can be integrated; and high accuracy and low omission ratio are achieved while high real-time performance is achieved. The method and device are suitable for high-requirement working conditions of the detection robot.

Description

technical field [0001] The invention belongs to the technical field of robot visual detection, and in particular relates to a detection method for a visual target of a detection robot. Background technique [0002] Detection robot is a ground mobile platform used to perform special detection tasks in a specific area, and its autonomous reconnaissance technology has always been the focus of research at home and abroad. With the development of deep learning technology, the visual intelligent perception ability has also made unprecedented progress. Different from image processing, pattern recognition and other methods in traditional machine vision, today's robot vision target detection is mostly based on the convolutional neural network (Convolution Neural Network, CNN) architecture, using deep learning to extract image features, and realizing artificial feature paradigm driven by experience A Paradigm Shift to Data-Driven Representation Learning. [0003] Traditional CNN-bas...

AI Technical Summary

Problems solved by technology

If the one-stage method is used, the robot vision system has a higher probability of missing important targets or target recognition errors. If the two-stage method is used, the robot can better complete the detection task, but it cannot meet the real-time requirements.

Focusing on the real-time detection direction of the robot, the advantages of the two types of detection methods are combined, combined with the visual mixed attention mechanism, and the adaptive target scale perception method has not yet been seen.

Images