Autonomous Vehicle Perception With Multi-Frame Learning for Easier Maintenance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing environment perception systems in autonomous driving are difficult to maintain, have a low upper limit of algorithm performance, and are incapable of utilizing massive data, particularly in complex urban scenarios.
Innovation Solution
A method utilizing onboard sensors to obtain consecutive multi-frame perceptual object information, performing association matching and state prediction with a machine learning algorithm to enhance algorithm performance through iterative updates based on data-driven methods, reducing the need for manual parameter adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional numerical computation theory is used for multi-source fusion modules, then the system can perform association matching and state estimation, but the system becomes difficult to maintain and has low algorithm performance upper limit
Solution Approach 1:
The patent replaces conventional numerical computation theory (mechanical/systematic approach) with deep learning algorithms (intelligent approach). Specifically, it uses neural network models to perform association matching and state estimation that were previously done through Hungarian matching algorithms and Kalman filtering, thereby improving algorithm performance while reducing maintenance complexity through automated learning.
Solution Approach 2:
The patent changes the fundamental parameters of the computation system by transitioning from fixed mathematical models to adaptive neural network parameters. The system uses learnable parameters in deep learning models that automatically adjust to optimize performance, replacing the rigid parameter structures of conventional numerical methods.
2Extent of automation
If conventional multi-source fusion modules are used, then the system can process perception data, but it is incapable of automating the use of massive data
Solution Approach 1:
The patent implements self-service through deep learning models that automatically learn from massive data without manual intervention. The neural networks perform automatic feature extraction, association matching, and state estimation, eliminating the need for manual parameter tuning and rule configuration required by conventional systems.
Solution Approach 2:
The system incorporates feedback mechanisms where the deep learning models continuously learn from processed data, improving their performance over time. The automated processing pipeline uses feedback from prediction results to refine future predictions, enabling effective utilization of massive data for continuous improvement.
3Measurement precision
If conventional association matching algorithms are used, then the system can match perceptual objects, but it presents a low upper limit of algorithm performance
Solution Approach 1:
The patent replaces conventional association matching algorithms (Hungarian matching, greedy matching) with deep learning-based matching approaches. The neural network models learn optimal matching strategies from data, achieving higher accuracy and breaking the performance ceiling of traditional algorithms through adaptive feature representation and similarity measurement.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to the field of autonomous driving technologies, and specifically provides a method for obtaining environment perception information, a readable storage medium, and a smart device, which are intended to effectively solve the problems that existing environment perception systems are difficult to maintain, present a low upper limit of algorithmic performance, and are incapable of using massive data. In the invention, consecutive multi-frame perceptual object information is obtained based on data collected by onboard sensors, association matching and state prediction are performed on perceptual prediction objects in the consecutive multi-frame perceptual object information by using a machine learning algorithm, and an environment perception result of an environment in which a vehicle is located is obtained based on an obtained first association matching result and an obtained first state prediction result. The machine learning algorithm can implement iterative updates of the algorithm based on a data driven method, and allow for continuous improvement in algorithm performance by effectively utilizing massive data, so as to better support the application in complex scenarios, thereby effectively reducing difficulties and costs of maintenance.