Autonomous Parking Model Training with Abnormal Trajectory Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous parking technologies face challenges in constructing accurate parking models due to the inclusion of inaccurate real-time data, which can lead to unsafe and unreliable control strategies, especially when dealing with abnormal road conditions such as avoiding pedestrians and other vehicles.
Innovation Solution
A method for generating a parking model that involves obtaining multiple sample sets of driving data from a preset spot to a target parking spot, constructing a virtual parking cruise space, identifying and deleting abnormal data points, and performing model training using the refined data to create a target parking model, thereby improving data quality and reducing noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If real-time data is directly used for model training, then data collection is simple and fast, but data accuracy and reliability deteriorate due to abnormal road conditions
Solution Approach 1:
The patent applies preliminary action by constructing a virtual parking cruise space and planning reference parking trajectories before model training. This preprocessing step allows abnormal data to be identified and removed in advance, ensuring that only reliable data is used for training while maintaining efficient data collection processes.
Solution Approach 2:
The patent introduces a reference parking trajectory as an intermediary standard to evaluate and filter real-time driving data. By comparing actual trajectories against the reference trajectory, the system can identify and remove abnormal data points, thus improving data reliability without sacrificing collection efficiency.
2Quantity of substance
If all real-time driving data is used for training, then training data volume is large, but model accuracy deteriorates due to noise from abnormal positions
Solution Approach 1:
The patent applies the taking out principle by extracting and removing data corresponding to abnormal positions from the training dataset. By identifying segments where the vehicle trajectory deviates significantly from the reference trajectory and removing only those portions, the system maintains large training data volume while eliminating noise that would degrade model accuracy.
Solution Approach 2:
The patent uses parameter changes by introducing a threshold parameter for trajectory deviation detection. By comparing the distance between actual and reference trajectories against this threshold, the system dynamically identifies abnormal positions, allowing flexible control over data filtering to maintain both data volume and accuracy.
3Adaptability or versatility
If abnormal data is included in training, then data collection is comprehensive, but control strategy safety deteriorates
Solution Approach 1:
The patent introduces a reference parking trajectory as an intermediary standard to evaluate data safety. By comparing real-time driving data against this reference, the system can identify abnormal positions that would compromise control strategy safety while preserving comprehensive data collection for normal driving conditions.
Solution Approach 2:
The patent implements feedback by using the reference trajectory to continuously evaluate and filter training data. This feedback mechanism ensures that only safe, normal driving patterns are used for training, improving control strategy safety while maintaining data comprehensiveness through systematic abnormality detection and removal.
Data Source
AI summary
The present disclosure discloses a method for generating a parking model, an electronic device and a storage medium, and relates a field of autonomous parking technologies. The detailed implementing solution includes: obtaining multiple sample sets; constructing a parking cruise space for the target vehicle based on each sample set, and extracting a first parking trajectory corresponding to each sample set from each parking cruise space; recognizing an abnormal position on each first parking trajectory, and deleting driving data corresponding to the abnormal position from a sample set corresponding to each first parking trajectory to obtain target sample data in each sample set; and performing model training based on the target sample data in each sample set to generate a target parking model.