Obstacle State Detection With Segmentation-Attention Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional panoptic driving perception systems face challenges in efficiently detecting obstacle states due to limited computational improvements in lightweight models with small input image sizes, leading to suboptimal performance in autonomous driving applications.
Innovation Solution
A system utilizing deep learning attention technology merges the output of a semantic segmentation sub-network with an object detection sub-network, incorporating a feature extraction module and distance table calibration to enhance accuracy without significantly increasing computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If lightweight models with small input image sizes are used, then computational complexity is reduced, but detection accuracy cannot be significantly improved
Solution Approach 1:
The patent merges the output of the semantic segmentation sub-network with the object detection sub-network by fusing feature maps at multiple levels. This integration allows the system to leverage semantic segmentation results to enhance object detection accuracy without requiring separate independent processing pipelines, thus avoiding significant increases in computational complexity while improving detection performance.
Solution Approach 2:
The system divides the detection task into two sub-networks: semantic segmentation sub-network and object detection sub-network. Each sub-network processes features at different levels of abstraction, with the semantic segmentation providing contextual information that enhances the object detection results. This segmented approach allows efficient processing at each stage while achieving improved overall accuracy.
2Measurement precision
If attention mechanism computation is applied to improve detection performance, then detection accuracy is improved, but computational load increases
Solution Approach 1:
The attention mechanism is applied selectively to specific feature maps and regions rather than uniformly across all data. The system focuses computational resources on processing feature maps that contain relevant obstacle information, applying attention weights locally where needed to enhance detection accuracy without performing unnecessary computations across the entire image data.
Solution Approach 2:
The semantic segmentation sub-network performs preliminary processing to identify and segment different road elements before the object detection sub-network processes the data. This preliminary action prepares the data in advance, allowing the attention mechanism to focus on already-segmented regions of interest, thereby reducing the overall computational load required for accurate detection.
3Speed
If feature maps are manually processed to extract useful information, then convergence speed is fast, but useful information may be lost
Solution Approach 1:
The system dynamically adjusts the processing of feature maps through learned attention mechanisms rather than using fixed manual processing rules. The attention weights are adaptively computed based on the input data, allowing the system to automatically identify and preserve useful information while maintaining fast convergence through efficient gradient-based optimization during training.
Solution Approach 2:
The system incorporates feedback loops where detection results and loss signals are used to adjust attention weights and feature processing in subsequent iterations. This feedback mechanism ensures that useful information is preserved and reinforced during training, while the system learns to prioritize important features automatically, maintaining both speed and information integrity.
Data Source
AI summary
The invention provides a system for detecting obstacle state and an operating method thereof, comprising an image capturing module, a semantic segmentation module, a feature extraction module, an object detection module, and a distance table calibration module. The invention is delivered a semantic segmentation information to a model for processing self-learning, and selected an output of an original image size, for a carrier of an attention mechanism.