In-Vehicle Perception Training Using Self-Supervised Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development of Automated Driving System (ADS) perception functionality is hindered by the need for large volumes of annotated training data, which is costly and labor-intensive, and the challenges of data privacy and bandwidth requirements in federated learning, making it difficult to share and annotate datasets effectively.
Innovation Solution
A self-supervised machine-learning algorithm is used to reconstruct ingested images, forming an evaluation dataset with masked permutations to determine reconstruction errors, which are used to update the perception network without manual annotations, enabling efficient and automated training of machine-learning algorithms for ADS-equipped vehicles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large volumes of annotated training data are used to improve perception function performance, then the performance and reliability of the ADS perception module is improved, but the cost and labor intensity of data annotation increases significantly
Solution Approach 1:
The system uses self-supervised learning where the model learns from unannotated data by creating synthetic annotations through masked image modeling. The perception module generates its own training signals by reconstructing masked regions, eliminating the need for expensive manual annotation while improving performance on rare scenarios and edge cases
Solution Approach 2:
The system performs preliminary unsupervised pre-training on large volumes of unannotated data before fine-tuning with少量 annotated data. This preliminary action allows the model to learn general features and patterns from raw data, reducing the subsequent need for extensive annotation while maintaining high performance
2Adaptability or versatility
If federated learning is used to share training data across multiple vehicles, then the diversity and volume of training data is improved, but data privacy concerns and bandwidth requirements increase
Solution Approach 1:
Each vehicle performs self-supervised learning locally on its own unannotated data without needing to share data with other vehicles or central servers. The model learns from local masked image modeling tasks, achieving data diversity through local adaptation while completely preserving data privacy and eliminating bandwidth requirements for data sharing
3Measurement precision
If manual annotation of training data is performed to improve detection accuracy, then the measurement precision of object detection is improved, but the time and resources required for annotation increase
Solution Approach 1:
The system generates synthetic annotations automatically through masked image modeling where the model predicts masked regions from visible portions. This self-annotation process creates training labels without human intervention, achieving high detection accuracy while eliminating annotation time and enabling continuous learning from new data
Solution Approach 2:
The system creates synthetic copies of annotations by generating predicted masks and bounding boxes from the model's reconstruction outputs. These copied annotations serve as training labels, replacing manual annotation processes while maintaining detection accuracy through the model's learned patterns
Data Source
AI summary
A method for updating a perception function of a vehicle having an Automated Driving System (ADS) is disclosed. The ADS has a self-supervised machine-learning (ML) algorithm for reconstructing an ingested image and a ML algorithm for an in-vehicle perception module for detecting one or more objects or free-space areas depicted in an ingested image. At first, an image of a scene in a surrounding environment of the vehicle is obtained. The obtained image is processed to obtain an output image with one or more detected objects or free-space areas. Then, an evaluation dataset is formed accordingly. The evaluation dataset and the obtained image is processed to obtain a reconstruction error value for each evaluation image and an evaluation image with highest reconstruction error value is selected among plurality of evaluation images. Using the obtained image and the selected evaluation image, the ML algorithm for the in-vehicle perception module is updated.