Inspection Robot Visual Tracking Using Neural Network Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional visual tracking methods for inspection robots are inefficient and unreliable due to resource wastage, dependence on environmental conditions, and lack of robustness in detecting specific behaviors and human targets, especially under varying illumination and quick movements.
Innovation Solution
A neural network-based visual detection and tracking method that uses a camera to acquire dynamic images, preprocesses them with an embedded system, employs YOLOV4 for target detection, and an improved KCF for tracking, combined with a Densenet for depth feature extraction, allowing for real-time detection and tracking of human targets with specific behaviors without relying on external servers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If image information is acquired by robot body and uploaded to server for processing, then processing capability is improved, but resource waste and decrease in stability and instantaneity occur
Solution Approach 1:
The patent applies local quality by deploying a lightweight target detection model specifically on the robot's embedded processor rather than using a heavy server-based system. This allows the robot to perform detection locally with reduced resource consumption while maintaining adequate processing capability for tracking tasks.
Solution Approach 2:
The patent segments the processing tasks by separating target detection (performed locally on the robot using a lightweight model) from comprehensive image analysis (performed on the server). This segmentation enables the robot to handle real-time tracking independently while the server handles more complex analysis, reducing resource waste and improving responsiveness.
2Device complexity
If traditional people tracking algorithm is used, then simplicity is maintained, but stability dependence on environmental illumination and robustness on appearance deformation and quick movement deteriorate
Solution Approach 1:
The patent changes the parameters of the tracking algorithm by integrating deep learning-based target detection (YOLOV4) with traditional tracking methods. This hybrid approach maintains relative simplicity while significantly improving robustness to illumination changes, appearance deformation, and quick movements through the neural network's ability to learn invariant features.
Solution Approach 2:
The patent creates a composite tracking system by combining traditional tracking algorithms with deep learning-based detection. This composite approach leverages the simplicity of traditional methods while incorporating the robustness of neural networks, achieving both ease of implementation and high reliability under varying conditions.
3Loss of information
If server-based processing is used, then comprehensive analysis is achieved, but instantaneity and stability decrease
Solution Approach 1:
The patent segments processing tasks by performing time-critical target detection and tracking locally on the robot using a lightweight model, while sending only relevant data to the server for comprehensive analysis. This segmentation ensures instantaneity for tracking operations while maintaining analytical completeness through server processing.
Solution Approach 2:
The patent performs preliminary target detection and tracking locally on the robot before transmitting data to the server. This preliminary action filters out unnecessary data transmission and ensures that real-time tracking responds immediately to detected targets, while the server subsequently performs comprehensive analysis on the pre-processed information.
Data Source
AI summary
The present disclosure provides a neural network-based visual detection and tracking method of an inspection robot, which includes the following steps of: 1) acquiring environmental images of a dynamic background a movement process of the robot; 2) preprocessing the acquired images; 3) detecting human targets and specific behaviors in the images in the robot body, and saving the sizes, position information and features of the human targets with the specific behaviors; 4) controlling the orientation of a robot gimbal by using a target tracking algorithm to make sure that a specific target is always located at the central positions of the images; and 5) controlling the robot to move along with a tracked object. The neural network-based visual detection and tracking method of an inspection robot in the present disclosure has a quite high adaptive ability, achieves better detection and tracking effects on targets in a dynamic background scene.