Surround View Image Recognition for Accurate Lane and Vehicle Positioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle front cameras with narrow angles of view struggle to accurately detect lanes and adjacent vehicles due to image distortion and loss of height information in top-view conversions, leading to obscured lanes and undetected lane changes.
Innovation Solution
The method employs an SVM camera to capture original images, semantically segments vehicle and lane pixels, and uses a matrix multiplication accelerator network (MMANet) for low-cost DSP processing to extract and convert object information into physical positions, minimizing distortion and noise through polynomial regression and Kalman filtering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If top-view conversion is used to extract object information, then position and physical distance recognition is improved, but image distortion increases and height information is lost
Solution Approach 1:
The patent uses SVM cameras mounted at multiple locations (front, rear, left, right) to capture images from different spatial dimensions. By processing these multi-dimensional original images rather than converting to a single top-view, the system achieves accurate position and distance recognition while preserving image geometry and avoiding distortion.
Solution Approach 2:
The patent applies semantic segmentation to divide the image into distinct regions (vehicle, lane, road, etc.) and processes each segment independently. This allows accurate extraction of object information from original images without requiring top-view conversion, thereby maintaining image shape while achieving precise measurement.
2Ease of operation
If top-view conversion is used to simplify coordinate conversion, then object position conversion is improved, but lane and vehicle detection accuracy deteriorates
Solution Approach 1:
The patent introduces a processing module that acts as an intermediary between the SVM cameras and the driving controller. This module performs semantic segmentation and coordinate transformation on original images, extracting object information while maintaining accuracy. The intermediary processing enables precise detection without requiring simplified top-view conversion.
3Shape
If SVM original images are used for recognition, then image distortion is reduced, but processing complexity increases
Solution Approach 1:
The patent replaces complex geometric transformations and manual processing methods with deep learning-based semantic segmentation networks. The neural network automatically performs feature extraction, object identification, and coordinate transformation on SVM original images, reducing processing complexity while maintaining image integrity and avoiding distortion.
Data Source
AI summary
A method of recognizing a surrounding driving environment based on a surround view mirror (SVM) original image. The method includes: acquiring an image captured by an SVM camera; semantically segmenting a pixel corresponding to a vehicle (hereinafter, a vehicle pixel) and a pixel corresponding to a lane (hereinafter, a lane pixel) from the image; post-processing the semantically segmented vehicle pixel and lane pixel to extract object information including adjacent lane information and adjacent vehicle information from the image; and converting the extracted object information into a physical position and transmitting the physical position to a driving controller.