Instance Segmentation via Dual-Thread Neural Network for Vehicle Perception
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for instance segmentation in automotive technology, such as deep neural networks, are computationally complex and inflexible, making them unsuitable for real-time object detection in vehicles due to high demands on computing power and error proneness, especially when dealing with complex contours and overlapping objects.
Innovation Solution
A method utilizing a single deep neural network with a decoder divided into two threads for semantic segmentation and edge detection, allowing for independent analyses that combine to achieve instance segmentation, enabling precise detection of objects with their exact contours and distinguishing adjacent objects of the same class, while being efficient and real-time capable.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple complex deep neural networks are combined for instance segmentation, then object detection precision is improved, but computing power demand and device complexity increase significantly
Solution Approach 1:
The patent segments the instance segmentation task into two independent analysis threads: semantic segmentation (first analysis) and edge detection (second analysis). Each thread processes the captured image independently through separate neural networks, and their results are combined to achieve precise object detection with contours. This segmentation reduces the complexity of any single network while maintaining high detection precision.
Solution Approach 2:
The patent merges the results of two independent analyses (semantic segmentation and edge detection) to achieve instance segmentation. By combining the class information from semantic segmentation with the contour information from edge detection, the system achieves precise object detection without requiring a single overly complex network.
2Measurement precision
If multiple complex deep neural networks are combined for instance segmentation, then object detection precision is improved, but real-time processing capability deteriorates
Solution Approach 1:
The patent segments the instance segmentation task into two independent analysis threads: semantic segmentation (first analysis) and edge detection (second analysis). Each thread processes the captured image independently through separate neural networks, and their results are combined to achieve precise object detection with contours. This segmentation reduces the complexity of any single network while maintaining high detection precision.
Solution Approach 2:
Instead of using a single comprehensive complex network, the patent employs two specialized networks that each perform partial functions (semantic segmentation and edge detection). This division allows each network to be optimized for its specific task, improving overall processing efficiency and enabling real-time performance.
3Measurement precision
If semantic segmentation method is used, then precise spatial division between object classes is achieved, but individual objects cannot be distinguished from each other
Solution Approach 1:
The patent segments the instance segmentation task into two independent analysis threads: semantic segmentation (first analysis) and edge detection (second analysis). Each thread processes the captured image independently through separate neural networks, and their results are combined to achieve precise object detection with contours. This segmentation reduces the complexity of any single network while maintaining high detection precision.
Solution Approach 2:
The patent uses edge detection as an intermediary to recover object boundaries that are lost in semantic segmentation. The edge detection thread identifies contours and boundaries, which when combined with semantic segmentation results, restores the ability to distinguish individual objects while maintaining precise spatial division.
4Loss of information
If object recognition method is used, then individual objects are detected at their positions, but precise spatial assignment cannot be made for complex contours
Solution Approach 1:
The patent segments the instance segmentation task into two independent analysis threads: semantic segmentation (first analysis) and edge detection (second analysis). Each thread processes the captured image independently through separate neural networks, and their results are combined to achieve precise object detection with contours. This segmentation reduces the complexity of any single network while maintaining high detection precision.
Solution Approach 2:
The patent merges the results of two independent analyses (semantic segmentation and edge detection) to achieve instance segmentation. By combining the class information from semantic segmentation with the contour information from edge detection, the system achieves precise object detection without requiring a single overly complex network.
Data Source
AI summary
The discloser relates to a method for evaluating an optical appearance in the surroundings of a vehicle and to a vehicle. The method has the steps of providing a captured image of the surroundings of a vehicle and extracting features from the captured image. Furthermore, the method comprises carrying out a first analysis of the captured image, wherein one or more objects are detected as surfaces and the result of the analysis is provided as a first analysis result. A second analysis of the captured image is also carried out, wherein edges of one or more objects are detected and the result of the analysis is provided as a second analysis result, the first analysis and the second analysis being carried out independently of each other. The method further comprises combining the first analysis result and the second analysis result to form an output image.