USV Threat Evaluation Using CNN-Based Target Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing decision-making methods for unmanned surface vehicles are inefficient due to high computational requirements and long training times for target recognition and classification, limiting their ability to make rapid decisions in dynamic environments.
Innovation Solution
An intelligent decision-making system utilizing a convolutional neural network that processes image sequences from visible and infrared images, calculates threat factors, and evaluates threats in real-time, aided by external features like distance, speed, and trajectory, to rapidly adjust weight values and improve classification and recognition speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a full-connection neural network is used for target recognition and classification, then the system can achieve comprehensive target analysis, but the computing time and training time increase significantly
Solution Approach 1:
The patent divides the neural network into modular components: convolutional layers for feature extraction, pooling layers for dimensionality reduction, and fully connected layers for classification. This segmentation allows parallel processing and reduces overall training time while maintaining recognition accuracy.
Solution Approach 2:
The convolutional and pooling layers perform preliminary feature extraction and dimensionality reduction before data reaches the fully connected layers. This preliminary processing reduces the computational burden on subsequent classification stages, significantly decreasing training time.
2Measurement precision
If a full-connection neural network processes large-scale image data, then complete feature analysis is achieved, but computational resources are excessively consumed
Solution Approach 1:
The patent extracts essential features using convolutional and pooling layers, removing redundant information before it reaches the fully connected layers. This extraction process reduces the data volume requiring intensive computation, lowering energy consumption while preserving critical classification features.
Solution Approach 2:
The pooling layers perform dimensionality reduction by aggregating spatial information, transforming high-dimensional feature maps into lower-dimensional representations. This dimensional transformation maintains important patterns while reducing computational complexity for subsequent processing.
3Measurement precision
If traditional deep learning networks are used for real-time target recognition, then comprehensive analysis is performed, but the decision-making speed is too slow for dynamic environments
Solution Approach 1:
The convolutional and pooling layers perform preliminary feature extraction and dimensionality reduction in advance, preparing optimized data for rapid classification. This preliminary action enables the fully connected layers to process information faster, improving real-time decision-making speed.
Solution Approach 2:
The network is segmented into specialized layers that can process different aspects of target recognition in parallel. The convolutional layers handle local feature detection, pooling layers handle spatial aggregation, and fully connected layers handle classification, enabling efficient parallel computation for real-time applications.
Data Source
AI summary
An intelligent decision-making method and system for an unmanned surface vehicle are provided. The system comprises: an image acquisition module, configured to acquire an image sequence signal of an observation region; a target detection and recognition module, configured to receive the image sequence signal, recognize a preset interested target appeared in a scenario, determine and identify an image region occupied by the preset interested target; distinguish a type of the interested target, screen out the target that causes a threat to an unmanned platform itself according to the type of the target, continuously observe the target, and generate tracking trajectory information of the target; a target situation distinguishing module, configured to calculate a target threat factor; and a target threat evaluation and decision module, configured to evaluate a threat type, give a threat level, and lock the interested target with a high threat level for tracking in real time.