Driving Assistance Neural Network for Unified Detection and Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Advanced Driver Assistance Systems (ADAS) require separate deep-learning model architectures for object detection and image segmentation, leading to increased inference time and complexity due to the need for encoding, decoding, and complex post-processing steps.
Innovation Solution
A multi-task detection and recognition technology using a single deep learning model that integrates object detection and semantic segmentation, employing a block-based image detection approach to directly provide polynomial coefficients and category information, reducing the need for encoding and decoding and simplifying post-processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate deep-learning model architectures are used for object detection and image segmentation, then detection and recognition functions can be performed, but inference time is doubled and system complexity increases
Solution Approach 1:
The patent merges object detection and image segmentation into a single integrated deep-learning model architecture. The model simultaneously performs both tasks by sharing feature extraction layers and employing multiple output heads, eliminating the need for separate encoding and decoding processes. This integration reduces inference time while maintaining the reliability of both detection and segmentation functions.
Solution Approach 2:
The patent creates a universal deep-learning model that performs multiple functions: object detection, semantic segmentation, and lane line detection. The single model architecture handles diverse detection tasks through unified feature processing and multi-task loss functions, reducing overall system complexity and inference time compared to separate specialized models.
2Measurement precision
If image segmentation models with encoding and decoding processes are used, then pixel-by-pixel categorization can be achieved, but post-processing time increases due to grouping and polynomial fitting requirements
Solution Approach 1:
The patent extracts and eliminates the time-consuming encoding and decoding processes from traditional image segmentation models. By using a detection-oriented architecture that directly outputs object boundaries and lane line parameters, the model achieves pixel-level categorization accuracy without requiring subsequent grouping operations or polynomial fitting, significantly reducing post-processing time.
3Adaptability or versatility
If two different deep-learning model architectures are used for object detection and image segmentation, then comprehensive detection functions are achieved, but device complexity increases
Solution Approach 1:
The patent combines object detection and image segmentation architectures into a single unified model. The integrated structure shares backbone feature extraction layers and uses common processing pipelines, reducing the number of separate model components while maintaining comprehensive detection capabilities for both objects and semantic segments.
Solution Approach 2:
The patent develops a multi-functional deep-learning model that handles object detection, semantic segmentation, and lane line detection within a single architecture. The universal model uses shared feature processing and adaptive output mechanisms to perform diverse detection tasks, reducing device complexity compared to maintaining separate specialized models for each function.
Data Source
AI summary
The present invention provides a driving assistance system and a driving assistance computation method that utilize a deep neural network architecture to achieve object detection and semantic segmentation functionalities in a single inference of the same model.